text stringlengths 1.23k 293k | tokens float64 290 66.5k | created stringdate 1-01-01 00:00:00 2024-12-01 00:00:00 | fields listlengths 1 6 |
|---|---|---|---|
Future Prediction of Shuttlecock Trajectory in Badminton Using Player’s Information
Video analysis has become an essential aspect of net sports, such as badminton. Accurately predicting the future trajectory of balls and shuttlecocks can significantly benefit players by enhancing their performance and enabling them to devise effective game strategies. This paper aims to analyze data to provide players with an advantage in the fast-paced rallies of badminton matches. The paper delves into the innovative task of predicting future shuttlecock trajectories in badminton match videos and presents a method that takes into account both the shuttlecock position and the positions and postures of the players. In the experiments, players were extracted from the match video, their postures were analyzed, and a time-series model was trained. The results indicate that the proposed method improved accuracy by 13% compared to methods that solely used shuttlecock position information as input, and by 8.4% compared to methods that employed both shuttlecock and player position information as input.
Introduction
Recently, computer vision technologies have been employed to automate the analysis of video clips from net sports matches, such as tennis, volleyball, and badminton. These technologies enable player pose detection [1] and ball detection/tracking [2], allowing for the extraction of crucial information from match videos. This information can then be used to determine the high-level context, such as the players' actions, during the match.
Predicting the future movement of the ball, shuttlecock, and opponent is a critical aspect of sports video analysis. The ability to forecast these movements during a rally can give players a significant advantage over their competitors. Many players rely on their experience to make these predictions, and in fast-paced sports such as badminton, being able to anticipate the movement of the shuttlecock even a fraction of a second ahead can mean the difference between winning and losing the match. Performance analysis of players can also be considered so that players can understand strategies in matches. If a player's performance aligns with the prediction made by the proposed model, it implies that the shuttlecock's trajectory may be easily predicted by the competing player in the match. Players aim to play unpredictably to win the match.
Most research focused on predicting future movements in net sports has centered around predicting the landing point of the ball or shuttlecock [3][4][5][6][7][8]. However, in badminton, the shuttlecock must be hit without bouncing and from a higher, faster forward position, making the prediction of the landing point insufficient for gaining an advantage in the game. To achieve this, it is necessary to predict the shuttlecock's trajectory. Currently, trajectory prediction studies are limited to short-term events, such as the serve in table tennis or the toss in volleyball, and have not yet been applied to rallies in badminton.
This paper presents a method for predicting the future trajectory of the badminton shuttlecock during a match. One of the simple methods utilizes sequential models, such as recurrent neural networks (RNNs) [9], to model the motion of the shuttlecock by inputting previous shuttlecock trajectories and outputting future shuttlecock trajectories. To better reflect a player's decision-making process during a match, the proposed method also considers the player's position and posture information in addition to the shuttlecock's position information.
The proposed method was evaluated using the shuttlecock trajectory dataset [10] and demonstrated improved accuracy compared to methods that utilized only shuttlecock position information or methods that employed both shuttlecock and player position information. The results of the study can be summarized as follows: • This is a pioneering study on predicting the trajectory of the badminton shuttlecock during a match. • The proposed method predicts the shuttlecock's trajectory by considering the player's position and posture information, in addition to the shuttlecock's position information. • The results of the experiments show that the proposed method outperforms previous methods that use only shuttlecock position information as input or methods that use both shuttlecock and player position information as input.
Future Predictions in Net Sports
Recently, research on future prediction in various net sports, such as tennis, volleyball, table tennis, and badminton, has been growing in popularity. This section introduces previous research in the field and compares them with the current research. As shown in Table 1, previous research on future prediction in net sports has focused on predicting shot direction, landing point, stroke, and trajectory. A more detailed discussion of previous research in each sport is presented below. In tennis, Shimizu et al. [11] were the first to predict the future shot direction in three categories: right cross, left cross, and straight. They did so by using the player's continuous position and posture information up until the moment the ball was hit. They also developed a new dataset with shot directions for their study. However, in badminton, predicting only the direction is insufficient, as players have different movements for low and high trajectories.
In table tennis, Wu et al. [4,5] predicted the landing point of the service by using the player's motion information up until the moment just before hitting the ping-pong ball. This information was obtained through pose estimation. In volleyball, Sato et al. [6] predicted the landing point of the ball by using its velocity and 3D position, with an average error of about 0.3 m and about 1.5 s before the ball hit the floor. In badminton, Waghmare et al. [3] calculated the speed and direction of the shuttlecock to predict its landing point, using a two-dimensional laser scanner. While these methods help the player reach the landing point more quickly, they are not enough to give the player an advantage in the game by allowing them to hit the shuttlecock back faster and higher.
Stroke prediction involves predicting both the shot type and landing point. In tennis, Fernando et al. [7] predicted strokes by using a semi-supervised generative adversarial network (SGAN) [14], which combined a memory model with the automatic feature learning capabilities of a deep neural network. In badminton, Wang et al. [8] utilized a network called shuttleNet to predict the next stroke based on the current stroke. This was the first study to address stroke prediction in sports. While stroke prediction provides a better prediction of the shuttlecock's trajectory than landing point prediction, it still falls short compared to trajectory prediction.
In volleyball, Suda et al. [12] predicted the trajectory of the toss 0.3 s before it actually occurred, using the setter player's 3D joint positions. In table tennis, Lin et al. [13] predicted the trajectory of a subsequent serve by using a dual-network method. The method involved learning two separate parabolic trajectories: one from the service point to the landing point on the table (parabola 1) and one from the landing point to the hitting point (parabola 2). While trajectory prediction has been studied in both volleyball and table tennis, it has not yet been adequately studied in badminton.
All of the previous research has been conducted in recent years, and research on future predictions in net sports is still in its developmental stage. While shot direction prediction in tennis by Shimizu et al. [11] and stroke prediction in badminton by Wang et al. [8] have been explored, trajectory prediction has not been thoroughly studied, particularly in the case of badminton. Therefore, the aim of the current study is to predict the shuttlecock's trajectory in badminton.
Object Detection
Object detection is the task of identifying objects with specific attributes within an image or video and determining their location and extent by surrounding them with a bounding box.
There are two types of deep learning-based object detection methods: one-stage methods [15][16][17] that directly detect the target object from the input image, and two-stage methods [18-20] that first identify candidate regions in the input image and then perform detailed detection for each region. One-stage methods, such as YOLOv4 [17], prioritize processing speed and are suitable for real-time applications. Two-stage methods, such as Region-CNN (R-CNN) [18], Fast R-CNN [19], and Faster R-CNN [20], may have a slower processing speed compared to one-stage methods, but they offer higher detection accuracy. For this study, Faster R-CNN is used, as it is one of the highest-performing two-stage methods and provides more accurate information about the player's position and posture.
Pose Estimation
Pose estimation is the task of identifying information about a person's posture (eyes, nose, limbs, etc.) from an image or video showing a person. The joint points of a person are obtained as keypoints.
There are two types of deep learning-based pose estimation methods: the top-down method [1,[21][22][23][24][25][26][27], which performs pose estimation for each object identified after object detection in the image, and the bottom-up method [28][29][30][31][32][33], which first performs pose estimation for all objects in the image by connecting each keypoint to other objects of the same type. The former method tends to be more accurate because it estimates the posture of each object one by one, while the latter method is less accurate because it is challenging to learn to connect keypoints between the same objects. In this method, object detection and pose estimation are independent of each other in the top-down method, and HRNet [1] is applied to pose estimation after object detection.
Overview
This paper proposes a method for predicting the trajectory of a shuttlecock over the next (future) n f frames, based on the shuttlecock position, the positions, and postures of the two competing players detected during the previous (past) n p frames. The proposed method consists of two modules: a detection and pose estimation module for past frames, and a time series model module for future prediction, as illustrated in Figure 1. The first module takes a sequence of past frames from badminton match videos as input and performs shuttlecock detection, player detection, and player pose estimation for each image.
In the second module, the dimensions of the shuttlecock position, player position, and player posture information are each aligned in two dimensions. The six-dimensional information is then combined and used as input for trajectory prediction.
Detection and Pose Estimation from Past Frames
The shuttlecock position information is obtained from a shuttlecock detector, such as TrackNet [34]. These are two-dimensional coordinates x s and y s on the image.
For player detection and pose estimation, the MMPose framework [35] is utilized, which includes several pose estimation models and pre-trained models. An object detection model detects players and records their positions as bounding boxes, while a pose estimation model estimates the posture of the players and records their keypoints.
To detect humans, an object detector is utilized; we employ Faster R-CNN [20] trained on the Microsoft Common Objects in Context (MS COCO) dataset [36]. The detector provides bounding boxes and confidence scores as detection results. Bounding boxes are represented by four two-dimensional coordinate points on the image when a human is enclosed by a rectangle. The confidence score ranges from 0 to 1 and indicates the likelihood that the object within the detected bounding box is a human. By using the confidence score, only players, not referees or spectators, are detected. Referees and spectators that are detected alongside players have a lower confidence score, as they may be sitting, have only their faces in the image, be facing sideways, or appear small. Thus, the person with the highest confidence score is identified as the player, and the player's bounding box is obtained, as shown in Figure 2. The player position information is obtained by calculating the normalized center coordinates (x player , y player ) of the player with each coordinate in Figure 3 as follows: The players are distinguished by assigning a number to each of them in the image, starting with the player on the lower side. The player position information is represented as four-dimensional coordinates (x p1 , y p1 , x p2 , y p2 ), where x p1 represents the x coordinate of the player who is shown on the lower side of the image.
After the player detection, HRNet [1], which has been trained by MS COCO, is used as the pose estimator. As shown in Figure 4, 17 joints (eyes, ears, nose, shoulders, elbows, wrists, hips, knees, and ankles) are detected as keypoints in the bounding box obtained from player detection.
The coordinate values are expressed as absolute coordinates on the image. For example, if the coordinates of a keypoint are (x k , y k ) on the image, the normalized absolute coordinates (ax k , ay k ) of the keypoint are calculated using the values in Figure 3 as follows: where w, and h are the width and height of the input image, respectively. The posture information of the two players is represented as a 68-dimensional feature vector (x p1k1 , y p1k1 , . . . , x p2k17 , y p2k17 ), where x pikj represents the x coordinate of the first keypoint of the player who is shown on the lower side of the image.
Time Series Model for Future Prediction
The proposed method predicts the trajectory by inputting three types of information: the shuttlecock position information obtained by the shuttlecock detector, the player position information obtained by the object detector, and the player posture information obtained by the pose estimation.
At this point, the shuttlecock position information is 2-dimensional, the player position information is 4-dimensional, and the player posture information is 68-dimensional. In order to make three types of information into a feature vector of the same space, each of them is embedded into two dimensions by a fully connected layer as shown in Table 2. The shuttlecock position information and the player position information and the player posture information are combined to form a 6-dimensional feature vector. The combined information is fed into a long short-term memory (LSTM) network [37], which is the second module used for predicting the shuttlecock trajectory. The LSTM network uses multiple past inputs stacked together as its input, and its output is then forwarded to the fully connected layer for further processing.
Dataset
The shuttlecock trajectory dataset [10] was utilized, which was created for training and testing the TrackNet [34] and TrackNetV2 [38] models for badminton applications. This dataset comprises 26 match videos for training and 3 match videos for testing. The match videos have a resolution of 1280 × 720 and a frame rate of 30 fps, and are separated by rallies, which refer to recordings that start with a serve and end with a score.
Each frame in the dataset provides information about the shuttlecock's position and the moment it hits the racket. Twenty-three matches were used for training, excluding amateur matches, while three matches were used for testing. The professional matches used in this study were singles matches held in international tournaments between 2018 and 2021. The 23 match rally videos for training were randomly split into an 80% training set and a 20% validation set. The three-match rally videos for testing were used as the test set.
Data cleansing was carried out on the dataset to enhance prediction accuracy. The position coordinates of the shuttlecock in the dataset were set to (0, 0) whenever it was hidden by a person or not visible, which could negatively affect learning since the shuttlecock would appear to move unnaturally in the frames before and after it. Frames with such issues were removed from the dataset to ensure that consecutive frames with unnatural shuttlecock movements were not used for learning.
Data augmentation was also carried out to increase the diversity of the training data. Since flipping a badminton match video upside down would no longer be appropriate for the sport, only left-right flipping and translations were performed. The original image was flipped left and right with a probability of 50% and then translated to the right or bottom in a range of 0 to 50 pixels relative to the width and height of the image. Figure 5 shows the results of the data augmentation.
Evaluation Metrics
Two types of displacement errors, namely the average displacement error (ADE) and final displacement error (FDE), were used as evaluation metrics in this experiment. ADE is the average of errors across all output frames, while FDE is the error at the final point of the output trajectory. The Euclidean distance was calculated using the two-dimensional coordinates in a 1280 × 720 pixel image, and the unit of measurement is pixels. ADE is more important in this task since it predicts the trajectory for multiple frames rather than just the landing point.
Network Training
We implemented the proposed method in PyTorch [39] (1.12.1+cu102, with Python3.7.13) and ran it on the NVIDIA TITAN RTX processing unit using CUDA 11.4. For LSTM, the number of layers was set to 3, the hidden layer to 128 dimensions, and the network was optimized using Adam [40], with a weight decay of 10 −4 , the momentums β 1 = 0.5 and β 2 = 0.999, and a learning rate of 0.02. The model was trained for 400 epochs with four input frames and 12 output frames for all cases.
The mean squared error (MSE) was employed as the loss function: where F is the time-series model LSTM of the proposed method, x is its input data and gt is the ground-truth data of the output shuttlecock position. F(x) and gt are D-dimensional vectors and F(x) i is the value of F(x) on the i dimension. The output results were compared with the ground-truth data to calculate the error. Then, the parameters were updated to reduce the error using backpropagation.
Other Models Baseline Models
To verify the effectiveness of the proposed method, two baseline models were set up. One is to input only the shuttlecock position information without any player information, and the other is to input the shuttlecock and player position information without any player posture information.
Other Time-Series Models LSTM was used as the model for trajectory prediction, but four other time-series models were also examined: an RNN [9], gated recurrent unit (GRU) [41], Transformer [42] and sequence-to-sequence learning with neural networks (Seq2Seq) [43]. The network parameters were set the same as in the proposed method model for all time-series models, except for the learning rate. Only the learning rate was set as shown in Table 3.
Other Representations of Posture Information
In addition to the absolute coordinate values on the image used in this method, we also examined other ways of representing the posture of the players: the relative joint positions in the bounding box and the heat map generated for the pose estimation.
If the coordinates of a keypoint are (x k , y k ) on the image, the normalized relative coordinates (rx k , ry k ) to the bounding box are calculated using the values in Figure 3 as follows: The accuracy of three inputs to the time-series model was compared: absolute coordinate values on the image, relative coordinate values to the bounding box, and heatmap generated by pose estimation, respectively.
When using absolute coordinates on the image, the spatial coordinates are the same as the position information of the shuttlecock and players. When using relative coordinates to the bounding box, the spatial coordinates are different from the position information of the shuttlecock and players. Relative coordinates have a more significant influence on the posture information than absolute coordinates because the change in posture is greater with relative coordinates. Figure 6a shows the overview in the case of using the joint positions to represent the postures of the players.
When using the heatmap, a heatmap was first generated for each of the two players in the pose estimation part and superimposed, as shown in Figure 7. Then, the heatmap including the posture information for the two players was sent to ResNet-18 [44] to extract a 512-dimensional feature vector, which was input to the time-series model together with the shuttlecock position and the players' positions. Figure 6b shows the overview in the case of using the heatmap to represent the postures of the players.
Data Augmentation Values
Experiments were conducted for data augmentation with different probabilities of left-right flipping and ranges of parallel shifts. The probability of left-right reversal was set to 0%, 25%, 50%, or 75%, where 0% means that the image is not flipped. The range of translation was set to 0 pixels, 50 pixels, or 100 pixels, where 0 means no translation.
The Number of Frames of Past/Future
Experiments were conducted with the number of past and future frames fixed at 4 and 12, respectively. For the results shown in Section 3.3, five different combinations of the number of past and future frames were tested by increasing the number of past frames from 4 to 12, while keeping the total number of past and future frames at 16.
Comparison with Other Models
As for the quantitative evaluation, the results of comparing this method with the baseline methods and the methods using other time-series models are shown in Tables 4 and 5. As a qualitative evaluation, the results of comparing this method with the two baseline methods are shown in Figure 8, and the results of comparing this method with methods using other time-series models are shown in Figure 9.
The proposed method shows the best results for ADE and FDE. It also qualitatively predicts the trajectories closest to the ground truth.
Compared to the method with the baseline methods, the proposed method improves ADE by about 13% compared to the method using only the shuttlecock position information as input, and by about 8.4% compared to the method using the shuttlecock and player position information as input. This shows that the proposed method effectively uses all the information on the shuttlecock position, player position, and player posture. Compared to the method with the highest accuracy for each time-series model, the proposed method using LSTM improves ADE by about 9.8%, 5.0%, 20%, and 12% compared to the methods using RNN, GRU, Transformer, and Seq2Seq, respectively. Therefore, it is shown that the proposed method using LSTM is the best model among the time-series models considered.
Comparison by Data Augmentation Values
The results of the experiments with different probabilities of left-right flipping and different ranges of translation are shown in Tables 6 and 7. The proposed method shows the best results for ADE and FDE. This indicates that a 50% probability of left-right flipping and a 50-pixel range of translation are appropriate among those considered in this study, and that the data augmentation is effective.
Comparison of the Number of Frames of Past/Future
The accuracy improves as the number of input frames increases and the number of frames to be predicted decreases as shown in Figure 10, Tables 8 and 9.
Limitation
This study proposed a method for predicting future trajectories in badminton match videos. However, several limitations exist in the framework.
The first limitation is the position of the cameras. The dataset used in this study includes high-level tournaments, such as the World Championships, all of which were recorded from a distance so that the entire court could be seen. This makes it easy to obtain information on shuttlecocks and players. However, recording from a distance requires equipment, such as a tripod and a photographer, making it difficult to set up a camera when attending a match alone or when there is no access to a suitable recording location. In addition, since the camera is recording from a distance, it is likely that people may pass in front of the camera. If there are many frames where the shuttlecock is not in the angle of view and cannot be tracked, it would be difficult to predict.
Second is the adaptation to sudden changes in trajectory. As shown in Figure 11, if the trajectory changes at the moment of hitting back or the moment the shuttlecock touches the floor, it is quite difficult to predict. Even if it could be predicted that the trajectory would change after bouncing back, the direction of the change could be wrong. In this method, the position and posture information of the player was input in addition to the shuttlecock's position information, which enabled some frames to adapt to changes in trajectory, but there were still many frames that were difficult to predict.
Future Work
There are two things that need to be worked on in the future to make this research work with in-the-wild data.
The first is to improve the generalization performance of the model. For the practical scenario, it is necessary to confirm that the model is general enough by testing the proposed model on multiple data sets in badminton (e.g., those with different camera locations and those containing matches of amateur players) and other net sports, such as tennis and table tennis. For this purpose, it would also be useful to be able to automate the rally segmentation and shuttlecock position detection in match videos.
The second is to perform trajectory prediction in 3D space. Since this research is based on the novel task of performing trajectory prediction in badminton, we performed the prediction in 2D space as a first step. If we can further develop this task to predict in 3D space and project the predicted trajectory into real space, it can be used for training to predict the trajectory of a shuttlecock hit by an opponent or to prevent an opponent from predicting the trajectory of a shuttlecock hit by oneself. This can further help players improve their skills.
Conclusions
This paper approached the novel task of predicting the trajectory of the shuttlecock in a badminton match video and proposed a trajectory prediction method that uses information about the shuttlecock's position and the players' positions and the players' postures. Experiments comparing the proposed method with the baseline method were conducted to confirm the effectiveness of the proposed method. Furthermore, experiments with different time-series models show that the LSTM used in this method achieves the highest accuracy.
Conflicts of Interest:
The authors declare no conflict of interest. | 6,035.4 | 2023-05-01T00:00:00.000 | [
"Computer Science"
] |
Effect of thermal annealing on the optical stability of amorphous Ge–Se–Te films
Ge–Se–Te chalcogenide films with Ge content from 10% to 27% were prepared using thermal evaporation. The films were annealed with different times, and the changes of the linear refractive index and optical band gap of the films with different chemical compositions were investigated. It was found that, after 30 h of annealing, the Ge20Se8.5Te71.5 film exhibits the smallest change ratio in terms of linear refractive index (<0.5%), optical band gap (<1.5%), and thickness (<2.5%). Therefore, this component has the best optical stability in the Ge–Se–Te system studied in this paper. The optical band gap of Ge20Se8.5Te71.5 is about 0.8 eV, and the refractive index exceeds 3.4, which is beneficial to the applications in Te-based optical waveguide devices.
Introduction
Chalcogenide glass is an amorphous compound consisting of one or more chalcogen elements (S, Se, Te) that are covalently bonded with other elements like Ge, As, Sb. The chalcogenide glasses are widely used in optical and information storage devices because of their excellent optical properties, like broad infrared transmission range, high linear and nonlinear refractive index, and low phonon energy [1][2][3][4]. High nonlinear refractive index can make them play an important role in applications such as supercontinuum, all-optical switching, and wavelength conversion [5][6][7][8]. In the fabrication process of many optical devices based on planar waveguide, a very important step is to prepare high quality optical films with stable physical properties. However, the internal defects of the films and their amorphous nature lead to that the structural relaxation usually occurs under external energy input (such as light illumination, ion-irradiation, thermal annealing, etc). This in turn would cause the change of the optical parameters, like the transmission, refractive index, optical band gap, etc, leading to the deterioration of the performance in chalcogenide-based optical devices. Many studies have concentrated on the effect of external energy input on the properties of the films, for example, photo-bleaching and photodarkening effects in the films [9][10][11][12][13], thermal stability was also investigated in As-S, Ge-As-Se, Ge-Sb-Se, Ge-Ga-S thin films [14][15][16][17]. However, these compounds containing toxic arsenic may volatilize during deposition or annealing, which are environmentally unfriendly.
In contrast, Te-based chalcogenide glasses are becoming attractive due to the emerging applications of the materials in a waveguide range beyond 10 μm. Te element has a larger relative atomic mass and lower phonon energy than the S and Se elements, resulting in a higher nonlinear coefficient and a far infrared cutoff wavelength [18,19]. At the same time, the material does not contain arsenic, which seems to be more environmentally compatible. However, the development of tellurium-based materials is limited by poor stability, and they tend to crystallize, resulting in a small glass-forming area of the material [20]. In order to screen the best film for the applications in Te-based planar waveguide devices, it is very important to find the relatively stable composition in the formation of the films [21]. This paper mainly studies the effect of annealing on the optical stability of amorphous Ge-Se-Te films. We deposited a series of the films with different compositions and annealed them with different times, and investigated the changes of the optical parameters such as refractive index, optical band gap, and transmittance of the films with an aim to find out the film with stable optical performance.
Experimental
Ge x Se 10 Te 90−x glasses with different compositions (x=15, 17.5, 20, 22.5, 25, and 30 at.%) were prepared by a melt-quenching technique. Appropriate amount of high purity Ge, Se and Te elements were weighted and placed into a quartz tube and then evacuated to 10 -5 Torr before being sealed using hydrogen torch. The tube then was placed into a rocking furnace with a maximum temperature of 950°C. After 24 h homogeneous rocking of the mixture, the tube was rapidly cooled in an ice-water mixture, and then annealed at a temperature of 20∼30 ℃ below its T g for 2 h. The as-prepared glasses were then used as starting materials for thermal evaporation.
The films were deposited on both silicon and quartz substrates using thermal evaporation. Before deposition, noteworthily, the substrates of silicon were stripped with a solution of 20% HF in order to eliminate native oxide layer. All the substrates were ultrasonically cleaned using acetone, alcohol, and pure water, step by step, and then dried with nitrogen gas flow. After that, the substrates were placed onto the rotatable substrate holder. The distance from the evaporation boat to the substrate is 300 mm. In each run of the experiment, the chamber was evacuated to 10 -7 Torr, and then Ar gas was introduced. Before deposition, Ar + was used to clean the substrate for 10 min and this procedure has been demonstrated to be able to improve the adhesion between the film and substrate. The beam current was 25 mA. The deposition rate can be changed by tuning the value of the current applied onto the evaporation source. The thickness of the thin film can be in situ read via a quartz thickness monitor placed near the substrates, and ex situ measured using a DEKTAK 3 Veeco mechanical profiler. The compositions of the as-deposited films were examined by using an Energy Dispersive x-ray Spectrometer (EDS) installed in a Tescan VEGA3 SB-Easyprobe Scanning Electron Microscope. X-ray diffractometer (Bruker D2) was employed to examine possible crystallization of the film, and check the amorphous or crystalline nature of the films. The transmission spectra of the films were measured in a range from 500 to 2500 nm using an ultraviolet visible near infrared spectrophotometer (PerkinElmer Lambda 950), and the energy band gap spectrum was estimated from the Tauc plot of the absorption edge. The refractive index of the films was measured using an infrared ellipsometer (J. A. Woollam IR-Vase II). Differential scanning calorimetry (DSC) was used to determine the glass transition temperature (T g ) of the Ge-Se-Te glasses. Noteworthily, in the previous literature [9], the researcher measured T g both of the bulk glasses and the films by using a flash DSC which only tens of nanograms of materials are needed, and found the difference of T g between the bulk and film is less than 5°C. It indicates that, there is only negligible change of T g between the film and bulk with the same composition. Thus, we here obtained the T g of Ge-Se-Te film by measured T g values of the bulk glasses with the same compositions of the films using the conventional DSC. The T g values are shown in table 1. Thermal annealing of the films was performed in a vacuum annealing furnace, and the annealing temperature was around 20 ℃ below T g with different duration time of 5, 10, 15, 20, 25, and 30 h.
Results and discussion
The compositions of the films were shown in table 1. It is well known that thermal evaporation generally leads to the deviation of the composition from the starting materials. In this case, we tuned the compositions via different evaporation rates, and found that, Ge content in the film generally is lower than that in the corresponding starting material. Nevertheless, we chose a group of the films with similar Se content around 10 at.%. The annealing temperature (T a ) is about 20°C lower than T g , which is obtained from the bulk glasses that has the similar composition as the films. See the details in table 1. XRD pattern of the as-deposited and 30 h-annealed films were shown in figures 1(a) and (b), respectively. No sharp diffraction peaks can be found, and the broad XRD features indicate that thermal annealing below T g for 30 h cannot induce any crystallization in the films. Therefore, any changes of the optical parameters in the rest part of the paper is not due to the phase change behavior that has been frequently reported in Te-based films since the annealing temperature is too low to induce phase change.
The transmission spectra of the as-deposited thin films were shown in figure 2(a). We are particularly interested in the transmission edges that move towards the short wavelength with the decrease of Te content. In order to obtain the exactly value for the short wavelength cutoff, an enlarge transmittance spectra in visible-near infrared is provided in figure 2(b). The value of short wavelength cutoff is ∼1040, ∼980, ∼960, ∼900, ∼850, and ∼780 nm (the corresponding transmittance is ∼5%), for Ge10, Ge12, Ge17, Ge20, Ge23, and Ge27, respectively. This is in good agreement with that Te element has greater atomic mass and polarizability, leading to smaller optical bandgap with increasing Te content in the films.
The absorption coefficient (α) of the films can be calculated from the relationship [22], where d is the thickness of films and T is the transmittance. The optical band gap E g opt can be calculated according to the Tauc formula [23], where B is the tail parameter, and the value of n is selected as 2 for indirect transition amorphous materials like the Ge-Se-Te films [24]. The main panel of figure 3(a) shows the relationship between ahv 1 2 ( ) / and hv of the asdeposited films. The value of E g opt was estimated by extending the linear region to the intersection of the abscissa. The optical band gap of the as-deposited film was further shown in the inset of figure 3(a). It is evident that, as the Te content increases, the optical band gap decreases gradually. As we know that, the size of the atom can play a very important role in changing the energy band-gap of materials [25]. In the amorphous films, the insufficient number of atoms will produce some unsaturated bonds and saturated bonds, and the unsaturated bonds will form some defects in the amorphous structure. Te element has a larger atomic radius, which will increase the number of unsaturated bonds and the local density of states, and ultimately reduce the optical band gap of the amorphous film [26]. Figure 3(b) is the refractive index of the films measured by the infrared ellipsometer, where the Cauchy model was used to fit the optical parameters of the films with a mean square error (MSE) less than 2.5. With the increase of Te content, the refractive index of the films is gradually increased. The refractive index of the sample is generally related to the polarizability of the constituent elements. The Te element has a greater relative atomic mass and ion polarizability, which leads to an increase in the refractive index in the Te-rich films.
To study the refractive index dispersion below the energy band gap for the as-deposited Ge-Se-Te thin films, we here performed the Wemple-DiDomenico Single oscillator model [27]. Such model consists of inter band transition of electron as a single oscillator. The dependance of refractive index on photon energy below interband absorption edge (hv< E g ) is given as [28], where E 0 is the single oscillator energy, E d is the dispersion energy, E=hv=photon energy. Figure 4 shows plots of (n 2 − 1) −1 versus (hv) 2 of the as-deposited Ge-Se-Te thin films. By linearly fitting the above curve, the single oscillator energy (E 0 ) and dispersion energy (E d ) can be derived. The parameter E 0 corresponds to the distance between centers of gravity of the valance and conduction band, while the other parameter E d corresponds to the average strength of the inter-band optical transitions [29]. We list the specific values of E 0 and E d of these films in table 2. The introduction of Te element can improve the nonlinear parameters of the material. Similarly, the smaller single oscillator energy (E 0 ) indicates that the material has superior nonlinear optical quality. We investigated the effect of annealing on the films with different durations below T g . Figure 5 shows the variation of the thickness change ratio of Ge-Se-Te thin films with different annealing time, where d is the thickness of the as-deposited films and Δd is the thickness difference between the as-deposited and annealed film. We found that, with increasing annealing time, the thickness of Ge10, Ge12, and Ge17 gradually decreases; while that of Ge20, Ge23 and Ge27 increases. Figures 6(a)-(c) show the dispersion of the refractive indexes of Ge10, Ge20 and Ge27 films with different annealing time, respectively. It can be seen that, the refractive index at a certain wavelength in figure 6(a) increases with the increase of the annealing time, but that in figures 6(b) and (c) decreases with the increase of annealing time. However, the variation range in Ge20 is smaller compared with other two films. In order to clearly elucidate, figure 6(d) shows the annealing time dependent refractive index at a wavelength of 3 μm for all the Ge-Se-Te films. With the increase of annealing time, the refractive index of Ge10, Ge12, and Ge17 films gradually increases, while the refractive index of Ge20, Ge23, and Ge27 films decreases. Previous studies have confirmed that the refractive index is closely related to the density of the material [30,31]. Because thermal annealing induces the change of the thickness in the films as shown in figure 5, the decrease of the film thickness would increase the density, leading to the increase of the refractive index. Otherwise, the decrease of the thickness would induce the decrease of the refractive index.
We have annealed each group of films for different times, and tested the transmission spectra of the films after each annealing. The transmission spectra of the three typical groups of films are shown in figure 7. We can see that, after the Ge10 film is annealed for different times, the cutoff edge of the transmission spectrum has a red shift. However, Ge20 and Ge27 show the opposite trend to Ge10. The transmission spectra of Ge12 and Ge17 also have a red shift, while the transmission spectra of Ge23 have a blue shift after annealing (details are not shown in here). The change rule of the transmission spectrum corresponds to the refractive index. All films conform to this pattern, and the red shift of the cut-off edge corresponds to a larger refractive index. Figures 8(a)-(c) shows the typical changes of the optical band gap of Ge10, Ge20, Ge27 films with the annealing time, respectively. Apparently, as the annealing time increases, the optical band gap of Ge10 tends to decrease, but the optical band gap of Ge27 tends to increase. Interestingly, the optical band gap of Ge20 changes insignificantly with the annealing time. Figure 8(d) shows the change of the optical bandgap for all the films. We found that, the optical band gaps of Ge10, Ge12, and Ge17 films gradually decrease with the increase of annealing time, while the tendency of the changes in Ge20, Ge23, and Ge27 films is opposite to those films with lower Ge contents. The change of the optical band gap is usually correlated with the atomic polarizability [32]. Previously investigations on chalcogenide materials indicated that a material with smaller E g opt was more likely to show a higher refractive index n, especially n increased with decreasing E g opt when a chalcogenide glass was subjected to photodarkening [33]. This phenomenon can also be explained by the model proposed by Mott and Davies [34], suggesting that thermal annealing should improve network topology and reduce the number of the defects. Since the films are created under non-equilibrium conditions, they contained a large amount of the disorders, which occupied the band tails or at positions within the bandgap. The change of the numbers of the disorders leads to the change of the bandgaps upon thermal annealing.
In order to show the changes of optical parameters with annealing time more intuitively, we list all the film parameters in table 3. The fitting error of refractive index (n) and optical band gap (E g opt ) is less than 0.3% and 0.5%, respectively.
Conclusion
In this paper, Ge-Se-Te chalcogenide thin films were prepared by thermal evaporation method. We found that, the actual compositions of the films are closer to the bulk glass when the deposition rate is 5-10 Å s −1 . The optical properties of the deposited films were characterized, and then the effects of different annealing times on the optical properties of the films were investigated. The results show that the optical band gap of the films increases with the increase of annealing time, and the change rate of the band gap is the lowest when the content of Ge is 20%. Simultaneously, as the annealing time increases, we found that the refractive index of films with a Ge content of less than 20% gradually increases, while the refractive index of films with a Ge content of more than 20% gradually decreases. The change in refractive index shows the opposite trend to the change in the thickness of the films. Comprehensive experimental results show that, the refractive index, optical band gap, and thickness change index of the film with 20% Ge content are the lowest in the whole annealing process. Among the thin film components studied here, Ge 20 Se 8.5 Te 71.5 showed the best thermal stability, and this component may be a potential material for photonic device applications. | 4,005.4 | 2020-11-10T00:00:00.000 | [
"Materials Science"
] |
Novel Titanium Nanospike Structure Using Low-Energy Helium Ion Bombardment for the Transgingival Part of a Dental Implant
Aim(s): The aim of the study was to fabricate a nanospike surface on a titanium alloy surface using a newly established method of low-energy helium ion bombardment. Various methods to achieve nanospike formation on titanium have been introduced recently, and their antibacterial properties have been mainly investigated with respect to Escherichia coli and Staphylococcus aureus. Oral pathogens such as Porphyromonas gingivalis play an important role in the development of peri-implantitis. For that reason, the antibacterial properties of the novel, nanostructured titanium surface against P. gingivalis were assessed, and a possible effect on the viability of gingival fibroblasts was evaluated. Materials and Methods: Helium sputtering was employed for developing titanium surfaces with nanospikes of 500 nm (ND) in height; commercially available smooth-machined (MD) and sandblasted and acid-etched titanium disks (SLA) were used as controls. Surface structure characterization was performed through scanning electron microscopy (SEM) and atomic force microscopy (AFM). Following incubation with P. gingivalis, antibacterial properties were determined via conventional culturing and SEM. Additionally, the viability of human gingival fibroblasts (HGFs) was tested through MTT assay, and cell morphology was assessed through SEM. Results: SEM images confirmed the successful establishment of a nanospike surface with required heights, albeit with heterogeneity. AFM images of the 500 nm nanospike surface revealed that the roughness is dominated by large-scale hills and valleys. For frame sizes of 5 × 5 μm and smaller, the average roughness is dominated by the height of the titanium spikes. ND successfully induces dysmorphisms within P. gingivalis cultures following the incubation period, while conventional culturing reveals a 17% and 20% reduction for ND compared to MD and SLA, respectively. Moreover, the nanospike surfaces did not affect the viability of human growth fibroblasts despite their sharp surface. Conclusion(s): This study successfully developed a novel titanium-nanospike-based structuration technique for titanium surfaces. In addition, the nanospikes did not hinder gingival fibroblast viability. Enhanced antimicrobial effects for such a novel nanospike-based resurfacing technique can be achieved through further optimizations for nanospike spacing and height parameters.
Introduction
The development and utilization of osseointegrated dental implants is one of the main routine dental surgery procedures applied in dental patient care on a global scale, providing major benefits in patients with tooth loss caused by a spectrum of conditions. However, as with any other surgical procedure, such implants do carry with them a risk of periand post-procedural complications. One of these complications is peri-implantitis, which typically manifests itself within the patient via the development of a microbial biofilm over the dental implant surface-especially on the transgingival part of the implant [1]. Peri-implantitis can lead to acute infections and induce severe inflammatory outcomes that ultimately damage peri-implant tissue, in a similar manner to permanent periodontal disease. However, unlike periodontal disease, the pathogenesis of peri-implantitis can be more widespread, leading to more rapid tissue devastation in such cases [2].
Typically, peri-implantitis, due to microbial biofilm development, is affected by multiple factors, including the biofilm-colonizing causative microbial agent(s), as well as the dental implant material with its unique surface topography. In addition, the fact that the oropharynx provides a level growth field for a myriad of micro-organisms due to the variety of food consumed by the individual, coupled with the fact that oral mucosa is continuously shed, leads to the prerequisite that biofilm-developing microbial colonies form an ecosystem within the oral cavity [3]. However, biofilms that develop on the hardened surfaces of native teeth or dental implants, which are not able to shed outer cellular layers, allow for the consequent accumulation of biofilms, inducing dental conditions that include periodontitis and peri-implantitis [3]. Furthermore, microbial biofilm development is facilitated on rough surfaces, consequently proving that material selection is pivotal for dental implant survival [3,4].
Although peri-implantitis-inducing microbial biofilms can consist of diverse microbial species, retrospective studies have revealed that the vast majority of dental implant periimplantitis cases involve overwhelming colonization by Porphyromonas gingivalis, followed by Prevotella intermedia and Prevotella nigrescens [5,6]. P. gingivalis is a Gram-negative anaerobic bacterium with previously established notoriety as a pathobiont-being present in the oral biome with the ability to proliferate rapidly within periodontal regions of the oral cavity [7]. In addition, P. gingivalis is also capable of thwarting the innate immune system of the host, together with exacerbating inflammation, due to a spectrum of virulenceinducing components such as gingipains and fimbriae, together with the presence of lipopolysaccharides and capsule formation [8]. Previous in vivo studies on peri-implantitis using murine models highlighted the issue that P. gingivalis-driven peri-implantitis leads to exacerbations in bone loss within implant regions, rather than native tooth regions [9].
With the knowledge of the etiology and pathogenesis of P. gingivalis-driven periimplantitis, the design and development of the appropriate materials for use in dental implants is crucial-especially since the unique topography of such materials can directly affect the risk of P. gingivalis-colonized biofilm accumulation [10]. Ultimately, the long-term success rate for osseointegrated dental implant surgical procedures depends on the ability of the dental implant surfaces to fend off-as well as possible-any attempted P. gingivalisbased biofilm development on the transgingival part of the dental implant or abutment surfaces.
Regarding the development of antibacterial dental implant surfaces, lessons can be learned through the study of defense mechanisms adopted by other organisms [11,12]. All such naturally occurring surfaces have in common the ability to counteract microbial threats and, consequently, are a source of inspiration for mimicking such topographies onto dental implants (among other potential bio-utilized prosthetics) for the purpose of conferring antibacterial effects and preventing consequent biofilm accumulation [12].
The study conducted by Bhadra et al. in 2015 adopted the model employed by dragonfly wings, with nanopatterned surfaces, for the development of a bespoke titanium nanowire topography on dental implants. Apart from having effective antibacterial capacity, such bespoke nanowire surfaces on the dental implant enabled enhanced proliferative growth and development of primary human fibroblasts over the course of 10 days [13].
Another highly important study was carried out in 2018 by Hazell et al., wherein black silicon structures with nanoneedle-covered surfaces were analyzed for antibacterial properties. The study concluded that black silicon and black diamond surfaces (the latter being developed artificially from black silicon) offer a less accommodating environment for bacterial culture attachment and growth [15]. The investigation performed by Saraeva et al. in 2020 focused on the development of nanospikes through laser ablation of a 500 nm thick gold film, followed by analyzing the antibacterial properties-utilizing live/dead staining techniques-of this nanostructured film pre-and post-ablation [18].This study highlighted how nanospike-bearing gold film can successfully exhibit antibacterial properties. This is due to the increased red propidium iodide staining of dead bacteria, which is a result of punctured bacterial cell membranes [18]. The recent study carried out earlier this year by Elliott et al. adopted alkaline hydrothermal treatment for engineering a titanium alloy surface, for the generation of nanospikes ranging between 250-350 nm (following a 4 h treatment) and 100-1250 nm (following an 8 h treatment). This investigation additionally demonstrated that the 8 h alkaline, hydrothermally treated nanospike titanium alloy surface managed to kill close to 40% of static bacterial cultures following a one-hour-long incubation period [21].
The antibacterial effects provided by nanopillar/spike resurfacing are mediated through mechanical piercing of bacterial cell walls and/or cell membranes, together with impeding appropriate adhesive functionalities typically adopted by bacterial pathogens [22]. The nanopillar/spike characteristics should have the capacity to penetrate both Gramnegative and Gram-positive bacterial species, as the discrepancy in bacterial cell wall thickness can have a major influence on the nanopillar/spike design's antibacterial efficacy. The study carried out by Ivanova et al. utilized silicon nanopillar arrays (35 nm diameter/90 nm periodicity) with three heights (220, 360, and 420 nm), which were developed using deep UV immersion lithography, with the 360 nm nanopillar found to have antibacterial properties against both Gram-negative Pseudomonas aeruginosa and Grampositive Staphylococcus aureus bacterial colonization of the nanopillar surface [23]. The study carried out by Jenkins et al. adopted a thermal oxidation methodology for developing nanopillars over grade 5 titanium alloy (Ti-6A1-4V), followed by a five-minute high-temperature treatment (850 • C), leading to nanopillar formation [24]. Such nanopillars were found to be effective in thwarting bacterial colonization by Escherichia coli and Klebsiella pneumoniae (both Gram-negative bacterial species), as well as by-although to a reduced effect-Gram-positive Staphylococcus aureus bacterial populations [24].
In order to be able to use such a technology adequately for the transgingival part of dental implants, the surface must support soft tissue integration [25]. However, no consensus has been reached on the ideal nanostructural design for conferring antibacterial properties to the transgingival part of dental implants, thus creating a vacuum for such a research niche.
In essence, the aim of this study was to fabricate nanospikes with 500 nm spike dimensions on titanium alloy surfaces using a newly established method of low-energy helium ion bombardment. Various methods to achieve nanospike formation on titanium have been introduced recently, and their antibacterial properties have been mainly investigated towards Escherichia coli and Staphylococcus aureus [26]. Oral pathogens such as P. gingivalis play an important role in the development of peri-implantitis, which can lead to implant loss. As a result, the antibacterial properties of the novel, nanostructured titanium surface against P. gingivalis were assessed, and a possible effect on the viability of gingival fibroblasts was evaluated. It is also important to highlight that, to the best of the authors' knowledge, no previous studies have been performed to develop nanopillar/nanospike surfaces for dental implants that can exert antibacterial properties against the major causative agents for peri-implantitis, such as P. gingivalis-although such titanium-based nanospikes have been found to confer antibacterial properties against other bacterial populations, as described above. This renders such research efforts highly attractive for possibly developing novel nanospike-coated dental implants that could have long-lasting efficacy against the risk of peri-implantitis within recipient patients in future clinical settings.
Preparation of Specimen and Material Characterization
Heated titanium samples were exposed to helium (He) plasma using an unbalanced magnetron sputtering source; by controlling the temperature (350 • C), the negative voltage on the sample (−120 V), and the time of exposure, the surface developed a nanostructured pattern (data in submission process). The spikes' formation was reported for the first time in 2014 by Kajita et al., who claimed that the process is related to physical mechanisms that cause the various morphological changes, including the following: the penetration and diffusion of irradiating helium atoms, helium bubble formation near the surface, physical sputtering, and redeposition of radicals. In particular, it seems that physical sputtering brings about significant difference from the morphological changes in tungsten, where the effect of the physical sputtering is negligible when the incident ion energy is less than 100 eV [27].
Three bespoke 15 mm diameter titanium discs with a 1.5 mm thickness were employed in this study, with the custom-made study group consisting of nanostructured discs with a 500 nm long nanospiked surface (ND). In the present study, the control groups consisted of commercially available smooth-machined titanium discs (MD) with a polished surface obtained through a grinding process, which also corresponded to the clinically standardized polished and smooth regions typically used for tissue-level dental implants and SLA ® titanium discs (SLA), which are sandblasted and acid-etched (Institute Straumann AG, Basel, Switzerland). The SLA surface, which corresponds to the implant's osseointegrated part, was used as an additional control group. All tested control groups were manufactured by Institute Straumann AG, Basel, Switzerland.
Scanning Electron Microscopy (SEM)
The specimens' surfaces were gold-sputtered and visualized with a scanning electron microscope (SEM, field emission at 5 kV) (Hitachi, Tokyo, Japan).
Atomic Force Microscopy (AFM
AFM experiments were performed in a dry environment, under a nitrogen gas atmosphere and at room temperature. Three samples (ND, MD, and SLA) were examined by means of atomic force microscopy (AFM) in intermediate-contact and full-contact modes. The topographic images and roughness were obtained in intermediate-contact mode AFM with a Nanosensors PPP-NC cantilever (Nanosurf, Liestal, Switzerland). The sensor stiffness and frequency were equal to k = 28 N/m and f = 160 kHz, respectively. Six different AFM images of every surface were taken, successively reducing the image size from 20 µm 2 down to 0.3 µm 2 . The adhesion force and adhesion energy measurements were performed in contact mode AFM with a PPP-CONT cantilever, with sensor stiffness and frequency equal to k = 0.1 N/m and f = 11 kHz, respectively. To obtain adhesion values, 30 force-distance curves on every surface were acquired.
Adhesion of P. gingivalis on Specimens
A 10 µL sample of Porphyromonas gingivalis (ATCC 33277) stock solution was inoculated to 10 mL of thioglycolate (BBL TM , Becton Dickinson, Allschwil, Switzerland) enriched with 0.5 mg/L menadione (Merck, Buchs, Switzerland) and 5 mg/L hemin Merck, Buchs, Switzerland), and the culture was incubated for 96 h in anaerobic conditions at 37 • C. Thereafter, the bacteria were harvested in a stationary growth phase by centrifugation (8500 rpm, 5 min, RT), resuspended in 10 mL of simulated body fluid [28] enriched with 0.2% glucose, and allowed to adhere to the three different material groups-ND, MD, and SLA-for 6 h, at 37 • C, in static anaerobic conditions. Thereafter, the discs were gently dipped in 0.9% NaCl (Merck, Buchs, Switzerland), with bacterial cultures being either harvested and cultivated by conventional culturing on Columbia blood agar (BBL ® , BD Becton Dickinson™, Allschwil, Switzerland), or fixed in 2% glutaraldehyde (Merck, Buchs, Switzerland), dehydrated in stepwise increasing concentrations of ethanol, critical-point dried, and coated with 10 nm of gold for scanning electron microscopy (SEM) [25]. Consequently, inoculated plates were incubated for seven days at 37 • C in anaerobic conditions, prior to assessment of colony-forming units for each individual titanium disc, as previously described [29].
The reduction in adhesion was calculated using the following formula: reduction% = 100 × ((CFU M or SLA − CFU ND )/CFU M or SLA ).
MTT Assay
To evaluate the effects of the various surfaces (ND, MD, and SLA) on gingival fibroblasts, an MTT cell viability assay was performed. In total, 30,000 HGF cells were cultured on various discs within 24-well plates for 72 h, followed by the addition of thiazolyl blue tetrazolium bromide (MTT) at a concentration 0.1 mg/mL. HGF cells were consequently incubated for a further 4 h, and the reaction was finally stopped by adding 125 µL of dimethyl sulfoxide (DMSO). MTT and DMSO were purchased from Merck™ (Buchs, Switzerland). All supernatants were harvested, and the optical density was measured at 590 nm, as previously described [29].
Cell Morphology by Scanning Electron Microscopy (SEM)
First, 4 × 10 4 HGF-1 cells were seeded on the specimen surface (ND, MD, and SLA). After 24 h of culture, cells were washed twice with phosphate buffered saline (PBS) and fixed with 2% glutaraldehyde in PBS for 30 min. Glutaraldehyde was removed and free aldehyde groups were quenched by adding 1 mL of 0.1 M glycine in PBS. Cells were washed twice with PBS and subsequently fixed with 2% osmium tetroxide in 0.1 M cacodylate buffer and incubated for 30 min. Cells were washed twice with cacodylate buffer. Dehydration was performed with graded ethanol (twice each with 50, 70, 90, and 100% ethanol for 2 min. Samples were critical-point dried with CO 2 (Critical Point Dryer, CPD 030, BAL-TEC) and sputtered (SCD 050, Sputter Coater, BAL-TEC AG, Balzers, Liechtenstein) with approximately 50 nm Au-Pd to make the cells electroconductive. Cell morphology was visualized with SEM (ESEM XL30, Philips, Eindhoven, the Netherlands), which was performed once, with duplicate probes (n = 2).
Unless otherwise stated, all reagents, chemicals, culture media, serum and PBS were purchased from Sigma-Aldrich Inc. (St. Louis, MO, USA).
Statistical Analysis
Data were collected on an Excel sheet (Microsoft Corporation, Richmond, CA, USA) for descriptive analysis. The normality of the results was tested using the Shapiro-Wilk test, and Student's t-test was applied (IBM ® , SPSS ® Statistics software Version 26.0 (IBM Corp., Armonk, NY, USA)) to assess statistically significant differences between the adhesion and viability of P. gingivalis and the viability of gingival fibroblasts on ND, SLA, and MD discs in the experiments conducted. The level of significance was set to p < 0.05. Figure 1A,B show views at a 52 • tilt angle, while Figure 1E,H show views at an 85 • tilt angle. The height of the spikes was directly proportional to the exposure time, although the base size (~200 nm) was not influenced. As seen in Figure 1 (Figure 1A,B, and more precisely in Figure 1C), the spike density was not modified for 500 nm nanospikes, and was influenced by the exposure time. The height of 500 nm was achieved after 24 h exposure time. As seen in Figure 1D-F, MD had a polished and smooth surface with microgrooves, which were a result of the polishing process. In contrast, SLA discs ( Figure 1G-I), which were produced by coarse grit-blasting with 0.25-0.5 mm corundum grit at 5 bar, followed by acid etching, had a high number of peaks/valleys across the surface.
AFM
The topography of the spikes was also investigated using atomic force microscopy (AFM) in intermittent-contact mode (Figure2). These characterizations were carried out on 500 nm tall spike samples. For scan frame sizes larger than 5 × 5 μm, the roughness was
AFM
The topography of the spikes was also investigated using atomic force microscopy (AFM) in intermittent-contact mode (Figure 2). These characterizations were carried out on 500 nm tall spike samples. For scan frame sizes larger than 5 × 5 µm, the roughness was dominated by large-scale hills and valleys. For frame sizes of 5 × 5 µm and smaller, the average roughness was dominated by the height of the titanium spikes (Figure 2A,B). When the scanned area was reduced further, the average roughness leveled off at a value of~200 nm, which is consistent with the SEM images (not shown here). Figure 2D,F show topographic images and profiles ( Figure 2E,G) of MD and SLA titanium surfaces. The MD surface showed a trench-like characteristic for mechanical polishing, whereas chemical etching of the SLA surface led to the formation of a rough surface. The average roughness measured by AFM for ND, MD, and SLA surfaces is summarized in Table 1 for three different surface areas. The highest roughness values were reported for the SLA titanium surface, whereas the machined titanium showed the lowest average roughness. A significant difference in roughness was found between the SLA and MD/ND surfaces. The change in roughness was also correlated with adhesion. The average roughness measured by AFM for ND, MD, and SLA surfaces is summarized in Table 1 for three different surface areas. The highest roughness values were reported for the SLA titanium surface, whereas the machined titanium showed the lowest average roughness. A significant difference in roughness was found between the SLA and MD/ND surfaces. The change in roughness was also correlated with adhesion. Table 1. The comparison of average roughness of ND, MD, and SLA surfaces measured via intermittent-contact AFM. The roughness depends on the scan area. For ND titanium surfaces and for large scan frame sizes, the roughness was determined by large-scale hills and valleys, whereas roughness for smaller scan areas was mainly due to the height of the titanium spikes.
Oral Pathogen Morphology
SEM images revealed similar images between the control, MD, SLA, and ND surfaces. The heterogeneous surface observed here was confirmed by conventional culturing, showing high variance in bacterial adhesion and survival rates (Figure 4).
Conventional Culturing Assays
The reduction in bacterial adhesion on ND in comparison to MD and SLA was 17 and 20%, respectively. No statistical differences in bacterial logarithmic count were found between groups (p > 0.05) following 6 h of bacterial adherence (Table 2).
Conventional Culturing Assays
The reduction in bacterial adhesion on ND in comparison to MD and SLA was 17 and 20%, respectively. No statistical differences in bacterial logarithmic count were found between groups (p > 0.05) following 6 h of bacterial adherence ( Table 2).
Comparison of Specimens in MTT Assay
The incubation of HGF cells on ND, MD, and SLA surfaces did not influence their viability. Moreover, all discs showed very similar results, as absorbance peaks were almost identical between the three surfaces tested. Bar graphs representing the MTT assay results ( Figure 5A) highlighted minimal discrepancies between the study and control groups, with such discrepancies not having any statistical significance. Table 2. Table representing mean CFU/disc following P. gingivalis exposure for possible bacterial adhesion on ND, MD, and SLA, for five replicates. ND showed a reduction of 17% and 20% compared to MD and SLA, respectively.
Specimen Interaction with Human Gingival Fibroblasts (HGFs) 4.3.1. Comparison of Specimens in MTT Assay
The incubation of HGF cells on ND, MD, and SLA surfaces did not influence their viability. Moreover, all discs showed very similar results, as absorbance peaks were almost identical between the three surfaces tested. Bar graphs representing the MTT assay results ( Figure 5A) highlighted minimal discrepancies between the study and control groups, with such discrepancies not having any statistical significance.
Cell Morphology
Cell morphology was qualitatively assessed via SEM after 24 h of culture. SEM images of ND, MD, and SLA showed that flattened and elongated HGF-1 cells adhered to the surfaces. Filopodia attachments were found on all samples; however, these attachments were more abundant on ND. MD showed spindle-shaped fibroblasts, whereas ND showed reticular-shaped HGFs. SEM images revealed that fibroblasts spread very flat and attach tightly to the smoother surfaces of MD and ND compared to SLA. However, HGF cells do not fully extend into the sandblasted and acid-etched morphology of SLA; moreover, filopodia attachments stretch over longer distances ( Figure 5B).
Cell Morphology
Cell morphology was qualitatively assessed via SEM after 24 h of culture. SEM images of ND, MD, and SLA showed that flattened and elongated HGF-1 cells adhered to the surfaces. Filopodia attachments were found on all samples; however, these attachments were more abundant on ND. MD showed spindle-shaped fibroblasts, whereas ND showed reticular-shaped HGFs. SEM images revealed that fibroblasts spread very flat and attach tightly to the smoother surfaces of MD and ND compared to SLA. However, HGF cells do not fully extend into the sandblasted and acid-etched morphology of SLA; moreover, filopodia attachments stretch over longer distances ( Figure 5B).
Discussion
A new approach for creating a nanostructured surface with nanospike heights of 500 nm (ND) was successfully introduced in this study. Nanospikes can serve as a physicomechanical-based antibacterial measure against P. gingivalis, which plays a leading role in the development of peri-implantitis. Antimicrobial titanium surfaces can be developed via a spectrum of methodologies, e.g., glancing angle sputter deposition, nanoimprint lithography, hydrothermal manipulation, or helium ion bombardment, which was introduced in the present study. The ladder technique used to create such a nanospike surface Through its sharp shape, the nanospike structure appears to perforate the fibroblast at some spots.
Discussion
A new approach for creating a nanostructured surface with nanospike heights of 500 nm (ND) was successfully introduced in this study. Nanospikes can serve as a physicomechanical-based antibacterial measure against P. gingivalis, which plays a leading role in the development of peri-implantitis. Antimicrobial titanium surfaces can be developed via a spectrum of methodologies, e.g., glancing angle sputter deposition, nanoimprint lithography, hydrothermal manipulation, or helium ion bombardment, which was introduced in the present study. The ladder technique used to create such a nanospike surface proved to be successful, especially in terms of the ability to modify nanospike height as desired. The control group in this study consisted of smooth-machined titanium discs (MD) with a polished surface-corresponding to the machined transgingival part of an implant, typically utilized for tissue-level dental implants-and SLA titanium discs (SLA), which were sandblasted and acid-etched (Institute Straumann AG, Basel, Switzerland), corresponding to the osseointegrated part of a dental implant.
When compared to an unstructured machined titanium surface, the nanospike surface had no detrimental or cytotoxic effects on fibroblast cell viability. This characteristic is essential in order to allow proper implant establishment with the surrounding tissues, consequently increasing the chances of successful soft tissue integration [30]. Additionally, HGF-1 cells exhibited a spindle-shaped morphology, and were oriented along the grooves on MD, according to our findings. On the other hand, ND and SLA mostly showed cells with a reticular shape and the formation of filopodia, which is consistent with other recent studies in the field [29,31]. This might be explained by the applied helium ion bombardment, which may have partially removed the machining features on ND (as shown in Figures 1 and 5B). These findings suggest that nanogrooves and machining features with the relevant morphological variations influence changes in cell alignment. Fibroblasts can respond to the micro-and nanotopography of the substratum surface, which is also known as 'contact guidance', referring to the orientation, changes in cell shape, polarity, and alignment of the cell as a result of the micro-or nanostructure of distinct surfaces [32]. Cells on smoother surfaces expand and establish a strong actin cytoskeleton to mechanically anchor themselves onto the topography, whereas topographic characteristics on rough surfaces are utilized to stabilize the cells [33,34].
Helium ion bombardment created a slightly higher surface roughness compared to the polished surface of MD, whereas SLA was 4.8× rougher. Several studies have found that the viability and proliferation rate of human gingival fibroblasts are significantly increased on polished surfaces when compared to alternative groups [35,36]. The cell response on various surfaces may not be ascribed to nanoscale roughness of the surface alone; rather, it appears that the variation in microprofile and chemical characteristics can alternate cell responses. It has been reported that an implant with a rough surface may exhibit a more prominent development of peri-implantitis than an implant with a smooth surface, although rough surfaces allow enhanced osseointegration [37,38]. In this regard, ND surfaces might be more favorable for peri-implant tissue health than implants with rough surfaces. Despite the rougher surface of ND in comparison to MD, no difference in the viability of HGFs could be detected, which is a promising sign for future application on the transgingival part of dental implants, as well as for possible soft tissue integration.
AFM adhesion force and work measurements were performed between ND, MD, SLA surfaces and the silicon AFM tip. The flat MD surface showed the greatest adhesion, whereas on the rough SLA surface the adhesion force and adhesion energy were reduced. This appears to be directly related to the surface roughness. Due to the flat shape of MD, the AFM tip seems to have a larger surface area to adhere to, despite the increased surface created by sandblasting and acid etching for SLA, or by helium ion bombardment for ND. Our findings suggest that adhesion force and energy are affected by both micro-and nanostructure. A recent study by Lagonegro et al. suggested that osteoblasts preferentially adhere to the peaks of the microstructure of SLA [39]. These findings are highly comparable to the data provided in this study, as SEM images revealed that the fibroblasts spread very flat and adhere tightly to the smoother surface of MD and ND compared to SLA. However, HGF cells do not fully extend into the sandblasted and acid-etched morphology of SLA, and filopodia attachments stretch across larger distances.
ND successfully induces dysmorphisms within P. gingivalis cultures following an incubation period of six hours. Bacteria attached to the ND surfaces appeared to be deflated or stretched when compared to bacteria incubated on the machined control surfaces, with the tips of the nanopillars protruding through the top sides of the collapsed cells; this outcome is consistent with earlier published studies [40,41]. Conventional culturing showed a bacterial reduction of 17 and 20% compared to MD and SLA, respectively. Our findings reveal a percentage-wise antibacterial tendency, but not a sufficient logarithmic reduction based on colony-forming units to claim sufficient antibacterial efficacy. Previous research investigated the effects of nanospike surfaces on S. aureus and E. coli, whereas the present work is centered on P. gingivalis. One probable cause is the difference in size and diameter between E. coli (2 m long, 0.25-1 m diameter) and P. gingivalis (1.5 m long, 1 m diameter), which may require a distinct nanostructured design of the nano-micro-hierarchical surface to boost antibacterial effectiveness. These distinct orders of magnitude might have a significant influence on the nanostructure's properties against various pathogens [42]. Other bacterial traits, such as structural and morphological properties, membrane thickness, or elasticity, may either hinder bactericidal action or encourage initial attachment. The murein thickness of bacteria is critical to the efficacy of antibacterial nanostructures. Ivanova et al. (2012) demonstrated that cicadae wings exhibit bactericidal effects on P. aeruginosa, with a death rate of approximately 2.05 × 10 5 CFU/(min × cm 2 ), while the present study showed a death rate for P. gingivalis of approximately 1.3 × 10 8 CFU/(min × cm 2 ), which is three magnitudes higher [40]. This surface is made up of nanopillars with diameters ranging from 50 to 250 nm, with varying heights, spacing, and densities. In comparison to the nanospike surface in the present work, the nanostructures were nearly half the length, implying that they were more efficient in rupturing the murein wall of the mentioned pathogen (P. aeruginosa). This suggests that when designing a nanostructure to combat multiple pathogens, a one-size-fits-all approach is not feasible.
A recent study by Bright et al. concluded that bactericidal efficacy is reduced in a stepwise manner as bacterial concentrations are increased, with declines in bacterial viability observed for S. aureus at 10 5 CFU/disc and P. aeruginosa above 10 6 CFU/disc. Surprisingly, biofilm depth analysis revealed a decrease in bacterial viability. P. gingivalis showed a 1.3 × 10 8 CFU adherence after 6 h in the present study, which is two orders of magnitude greater than in the work by Bright et al. The large number of colony-forming units at the start of the experiment could potentially be a reason for the different outcome observed when compared to other study findings, as suggested by Bright et al. [43].
The SEM scans, on the other hand, revealed that the ND study group had heterogeneity in nanospike growth across all examined regions for each disc. Such heterogeneity in nanospike surfaces could possibly be the main reason for the lack of the tangible antibacterial effect desired from such nanostructured biomaterials. Rather than providing efficient antibacterial properties, ND did lead to morphological deformities in P. gingivalis. It has been shown that stretching from the nanopatterned surface leads to deformation and mechanical rupture of P. aeruginosa [40,44]. During the adhesion process, if the degree of stretching is sufficient [45], it will lead to cell rupture and death; nevertheless, this was not the case for the nanostructures used in the present study, which could be explained by the different morphology of the pathogen used in this study. Despite such limitations, the results obtained from this study are highly comparable, and are equivalent to similar parallel studies on the development of antibacterial nanospike surfaces for dental implants. This study achieved nanospikes of up to 500 nm in length via a newly introduced technique, while in the recent study carried out by Elliott et al. such nanospike lengths peaked at 1250 nm [21]. In addition, although the investigation carried out by Elliott et al. did achieve success, with 11-40% of bacterial cultures being killed when attempting attachment on the nanospike surface [21], this reduction is highly comparable with the results (17-20%) presented in this study-although a percentage reduction is not enough to claim antibacterial properties; rather, a logarithmic bacterial reduction is a defining marker.
Following such additional knowledge, future research endeavors should focus on optimizing the nanospike deposition methodology in order to enhance nanospike surface homogeneity across the target regions. Other optimizations could include modification of the inter-nanospike spacing as a means for further potentiating antibacterial effects against P. gingivalis and other pathogens. In addition, future assays focusing on biofilm exposure to nanostructured surfaces should include various incubation time periods to increase bacterial load in order to simulate the accumulation of biofilm over time. This measure ensures possible acclimatization and establishment/attachment of such bacterial cultures on the investigated biomaterials.
Conclusions
In essence, the nature-inspired technology of nanostructured surfaces for the purpose of providing antibacterial properties against microbial threats is still in its infancy; 500 nm nanospikes have a tendency to hinder P. gingivalis adhesion, whilst also not interfering with the viability of gingival fibroblasts. However, nanospike length and spacing must be addressed in future studies in order to further influence the level of P. gingivalis adhesion reduction and the strength of HGF-1 adhesion.
The results demonstrated in this study confirm steady progress in the beneficial potential and clinical value of implementing such novel biomaterials for use in dental implants and other medical aids that can provide an increased quality of life within recipient patients in the near future.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/nano12071065/s1, Figure S1: Heated titanium samples were exposed to helium (He) plasma using an unbalanced magnetron sputtering source by controlling the temperature (350 • C), the negative voltage on the sample (−120 V) and the time of exposure, the surface developed a nanostructured pattern (data in submission process). Top view images for 24 h exposure (A-C) and for 48 h exposure (D-F). Images A, B, D and E are views at 52 • tilt angle. The height of the spikes was directly proportional to the exposure time, though the base size (~200 nm) was not influenced. As seen in Figure 1 (A and B and more precisely in images C and F), spikes density was not modified for both heights and is influenced by the ex-posure time. Roughly, twice the number of spikes is visible for the 500 nm height in comparison to 1000 nm spikes. | 7,880.6 | 2022-03-24T00:00:00.000 | [
"Medicine",
"Materials Science",
"Engineering"
] |
EMERGENCE AND SELF-ORGANIZATION OF COMPLEX SYSTEMS : THE ROLE OF ENERGY FLOWS AND INFORMATION A PHILOSOPHICAL APPROACH
How order emerges from noise? How higher complexity arises from lower complexity? For what reason a certain number of open systems start interacting in a coherent way, producing new structures, building up cohesion and new structural boundaries? To answer these questions we need to precise the concepts we use to describe open and complex systems and the basic driving forces of self-organization. We assume that self-organization processes are related to the flow and throughput of Energy and Matter and the production of system-specific Information. These two processes are intimately linked together: Energy and Material flows are the fundamental carriers of signs, which are processed by the internal structure of the system to produce system-specific structural Information (Is). So far, the present theoretical reflections are focused on the emergence of open systems and on the role of Energy Flows and Information in a self-organizing process. Based on the assumption that Energy, Mass and Information are intrinsically linked together and are fundamental aspects of the Universe, we discuss how they might be related to each other and how they are able to produce the emergence of new structures and systems.
INTRODUCTION
To respond the basic question of the following paper we consider necessary to give a more precise definition of the concepts we use to describe the process of emergence and selforganization.On the other hand we need to say that we talk about open and complex systems in general abstract way, without dealing with specific physical, biological or social systems.
We try to describe a very basic process, which assumes different qualities at different levels of complexity.So far, concepts like information, signal, signs, and others are used independently of concepts like meaning or understanding in the sense of human communication.
BASIC PHYSICAL CONCEPTS : MASS, MATTER AND ENERGY
Matter is historically one of the oldest concepts closely related to the form of physical objects.On the other hand the term energy comes up only in the 19 th century as a counter-concept to matter and, using the words of C.F.Weizsäcker, as a "substantialisation" of the 17 th century's concept of force. 1 Energy, on the other hand, is often defined as mass in movement.This general concept of energy has two basic and antagonic aspects: a) Energy as Heat, playing the role of a universal "random generator" producing chaos, and Emergence and self-organization of complex systems:the role of energy flows and information a philosophical approach Norbert Fenzl ARTIGO b) Energy as Work, as a kind of order "generator", producing organized, system-specific structures. 2t this point we would like to recall the important statement of Tom Stonier (1992).
He said: what mass is to matter, heat is to energy, organization is to information.In agreement to this, we say that matter is organized mass, or in more precise terms, organized movement of mass.So far, to organize mass the therefore needed energy appears as system specific (useful) work.
We remember also that heat and mass are supposed to appear nearly simultaneously (protons emerge approximately 10 -11 seconds after the supposed Big Bang) with the beginning of the Universe.So we can say that heat, as unorganized movement of mass, is the mother of all forms of energy we know.All the other types of energy are ultimately expressed as work, or as ability to organize mass to matter.
Without going into thermo-dynamical details, for our further understanding we consider that input energy of open systems is basically used in two basic ways: a) to perform (system specific) work and b) to overcome structural inertia, the dissipated part of energy, called entropy.
Finally in our understanding, the concepts of work and entropy are always system specific in the sense that they only make sense "inside" the system and cannot be exported.
For example, entropy cannot be "exported" as entropy, but only as energy, or heat in a wider sense.
OPEN SELF ORGANIZED SYSTEMS
In the present analysis we concentrate our attention specifically on, open, selforganized systems (Fig. 1).
These systems are composed of three basic dimensions of space-time: -The microscopic dimension on the level of the individual elements, -The mesoscopic dimension of the whole structure, limited by the structural boundary.
-The macroscopic dimension of the field of interaction or relevant environment, limited by the system boundary.
These three dimensions of space-time are inseparable parts of open and complex systems.The dialectic relationship between these three dimensions is a basic characteristic of complex systems.
Element
The concept element is related to the smallest unity of the structure, which is still relevant for the mesoscopic characteristics of a system.The elements (also called parts) constitute the microscopic dimension of the system.For example if we talk about some complex organic molecule, the atoms (C, H, or others) are classified as elements.This is not the case of an organism, such as a plant for example, where the concept of element makes sense only if it designates at least a cell, or a society, where element designs the individual member.
Structure
The concept structure is related to the body aspect of a system.Elements are linkes by forces, which guarantee the coherence and cohesion of the structure.The sum of these binding forces (from physical up to even psychological or cultural forces in the case of human society) characterizes the identity of a system and consequently makes the difference to the rest of the Universe by creating a structural boundary, which defines the mesoscopic dimension of the system. 3Structural boundaries are interfaces, assuming a very important role as mediators between the inner and the outer space of open systems.The basic characteristic of the structure is its structural inertia, the resistance of organized matter against movement (changes).
During the vital flow and throughput of energy, part of the energy input is required to overcome structural inertia, producing entropy (heat), which is lost by dissipation.The remaining amount of energy, used to organize (or reorganize) the systemic structures is what we call system specific work.The antagonism between these two fundamental forms of energy seems to be the basic driving force of self-organization.
Field of Interaction
The concept of field of interaction (Wirkungsfeld), or system-relevant environment, 3 Meso -to describe its intermediate position between micro-and macro-dimension of the whole system.
Emergence and self-organization of complex systems:the role of energy flows and information a philosophical approach Norbert Fenzl
ARTIGO
permanently with its relevant environment through the exchange of energy and matter, characterized by energy-input of higher quality (E1) and energy-output of lower quality (E2).
The input-energy is used by the system to: a) Weaken or break up the bonds (the cohesion) between the elements of the systemto overcome the structural inertia -dissipating energy (Entropy).
b) Reorganize the elements with the aim to (re) stabilize the mesoscopic structure by realization of Work (Stonier, 1990).This is the basic process, which gives adaptability to the system with respect to environmental changes.
Energetic-Material Metabolism (EMM)
The complete process of input-transformation (throughput)-output of energy, called energetic-material metabolism (EMM) of the system, imposes specific changes to its inner (microscopic) and outer (macroscopic) space.During their "life time" open systems transform part of their survival-relevant environment, creating a specific macroscopic dimension of space-time, the field of interaction, or also called system-relevant environment, which turns out to be a characteristic and inseparable part of all open complex systems.
The fact, that we include the field of interaction into the concept of system, means that we distinguish between the structuralboundary and the system-boundary.So, open systems cannot be reduced to their structural dimensions and what we call system is necessarily greater than the physical dimensions of its structure.
On one hand the process of EMM produces its corresponding interaction field.On the other hand, the same process obliges the system to react to all changes of its relevant environment.So far, the necessary structural changes are related to the external changes, which are partly caused by the system itself.
Maybe we can recall at this point the analogy between the relationships like: particle/wave, body/mind and mass/field.
A system only can interact with its relevant environment according to the dynamics of its own structural organization, or, according to the dynamics and needs of its own EMM, which functions according to the structural organization of the system and needs to adapt the environment to its specific requirements.The units of these so caused changes in the field of interaction we call, in a very general way, signs, without distinction if these signs are intentional or not.A specific type of structural organization produces signs in its relevant environment and so far, they naturally reflect essential characteristics of the structure, which produces them.
INFORMATION
To define the concept of Information and analyse its role in the process of emergence of new systems, it is necessary to focus briefly on some underlying concepts, like sign, signal, and data.
Signs
Basically, a sign is something that stands for something else.But here we can find different approaches.For example, Korzybski defines sign as a map, which means a territory.
As an example we can mention the different worldwide used signs to design bathrooms, airports, danger, etc.For C. S. Peirce, "a sign is something that stands for something else to someone in some respect".This definition is more complex and Peirce includes the subject in his concept.
Generally we can say that people working with linguistics would say that words are signs: So, a certain word (lets say car) has a potential meaning (what is written in the dictionaries) and a specific meaning (for a specific person) when used in a specific context.
We use the terminology in the following way: Signs stand for something else "in some respect" and do not represent the entire thing or experience to which they refer.For example a footprint, or all kinds of changes in the environment, which can be computed or recognized by the structure of an open and complex system, is considered a sign in Complex System Theory.
Signals
The main two definitions for signal which are satisfying the purpose of our present analysis are: In a most general context of complex systems, changes (differences) of the energy/matter input mean something for the system if they cause perceptible changes in its structural organization.So far, when signs are carried by the EMM input and cause a systemic reaction or structural changes in the system, they are codified as signals.
Data
People working with computer science define generally data as information that has been translated into a convenient form to store, move or process.Relative to today's computers and transmission technology, data is information converted into binary digital form.
In telecommunications, data sometimes means digital-encoded information to distinguish it from analogue-encoded information such as conventional telephone voice calls.
Data can often be sent in packets that arrives separately in pieces.
In the present discussion data is used as the unit of changes (delta dIs) structural Information is going through during its EMM.The word Information itself is composed by in and form, something is put "in-to a form" and seems to be a kind of synthesis between "self formation" and alien induced transformation.So far, we agree in general terms, with all the authors who define Information as a measure of quantity of form, or as a measure of structural organization.But it seems that even on inorganic levels of evolution this is only one aspect of the information concept. 4 On the other hand the concept of information is, at least since the works of Shannon and Weaver, closely related to the idea of transformation, emergence of the new or novelty. 5
The Concept of Information
There is also a large consensus that the concept of information is related to the idea of emergence of difference, which leads us to the concept of bit, as the unit of difference, and so far as unit of information. 6 Now, if we resume the different concepts of Information actually used, we can find at least the following useful definitions: a) Transmission of Information (I) is related to the transmission of Energy (E) and Entropy (S).But (I) is not equal to (E) nor to (S) (Ebeling 1993, Wiener, 1973, Stonier, 1990); b) The emergence of Information is only possible in self organized systems (Fenzl, Hofkirchner, Stockinger) 7 ; c) Information reduces the uncertainty of a system.Information as a measure for difference (Shannon, Weaver, 1949); d) Pragmatic Information requires novelty (in the sense of Shannon) and receptivity (in the sense of Weizsäcker, 1979) ; e) Ayres, (1994) distinguishes between two basic forms of Information: -D-Information (D -doubt) -SR-Information (SR -survival relevant) f) Information is neither matter nor energy (Wiener 1973); g) Structural Information measures the complexity of a system (Stonier, 1990).Basically we consider the definition of Stonier (1990) the most adequate.But in a very general way, all these concepts can be grouped basically into 3 different types of Information.
1. Structural Information (Is): Information, which represents the structural organization and the functionality of a system.
Pragmatic Information (Ipr):
Information, which appears at a mesoscopic level and represents the way system structures act on its relevant environment.
Potential Information (Ipt):
Information that exists only in potential form, such as a set of signs, not yet received and organized to structural Information by the system (or co-systems).
It is important to point out, that the emergence of irreversible differences in evolution shows a consistent internal logic: we need to admit some logical relation between the past, the present and the future to be able to talk about information.This step leads us directly to irreversibility and probability as basic parts of evolution and of the concept of Information. 8 Since open systems need to maintain its metabolism, they require permanent interaction with its environment.So far, the cycle of inputprocessingoutput of energy flows is essential and nothing can reach the structural space of such a system without energy flows.
Following the previously discussed definitions, we resume that: -Signs are produced (as changes, differences) in the relevant environment by metabolic activities of a system, -If signs, carried and transmitted by EMM energy flows are strong enough to cause structural changes, they are codified as signals and produce the actualisation of structural information.
Emergence and self-organization of complex systems:the role of energy flows and information a philosophical If we consider, that the same process occurs in (B) with respect to (A), and (B) reacts on (A), the first step of a coherent interaction between both systems is accomplished.Now, each system reacts to the other in its specific way, but they have at least two main possibilities in common: attraction or repulsion with all possible intermediate reactions.
In the case of attraction, systems with compatible reactions start to "behave" in a coherent way and establish cohesive links.
Finally we can say that open systems don't need to be in direct structural contact to interact; self-organization is intermediated by processes beyond their respective structural boundaries and by corresponding internal changes due to: a) The exchange of energy and matter with their relevant environment; Emergence and self-organization of complex systems:the role of energy flows and information a philosophical approach Norbert Fenzl
ARTIGO
b) The permanent actualisation of structural information (Is) and production of pragmatic information (Ipr).
CONCLUSION
To maintain the necessary flexibility to survive external changes, open complex systems must be able to respond internally by reorganizing its micro-state and externally by organizing their environment (macro-state) according to their own patterns of structural organization by setting signs, the smallest possible changes that structural transformation is able to imprint to its relevant environment.
So, structural Information is transmitted in "small energetic units" (signs), and must be "re-assembled", or in other terms, codified by the structure of the receiver system.The exchange of signals between systems requires some overlapping of their respective relevant environment (field of interaction) to create channels, able to transmit the signs through the activity of its EMM.
The incoming signals are compared to the already "embodied" structural information and classified in useful or harmful to the (Ayres, 1994).We also can say in a more "physical" way, that the incoming signals (p.ex.waves) are modifying -and being modified by -the system-specific organization of matter.
the macroscopic dimension of a system.As we know, all open systems are submitted to the same basic cycle: emergence, development, decay and death.During this cycle the structure of open systems suffers characteristic transformations interacting
. 2 -
Fig.2 -Interaction of 2 Complex Systems Of course signals also are used in the scope of human communication.In this case signals generally design a type of message which can consist of one or more letters, words, characters, signal flags, visual displays, or special sounds, with prearranged meaning and which is conveyed or transmitted by visual, acoustical, or electrical means.
ARTIGO Complexitas -Rev.Fil.Tem., Belém, v. 2, n. 1, p. 31-45, jan./jun.2017 -ISSN: 2525-4154 But as we stated initially we use the concept in a very general way, considering human communication only a quality of specific open systems, such as human beings or human society, for example.So far we do not link the concept to some kind of meaning, in the sense of human understanding. | 3,978.8 | 2018-03-13T00:00:00.000 | [
"Physics"
] |
Integration of hemicellulose recovery and cold caustic extraction in upgrading a paper-grade bleached kraft pulp to a dissolving grade
In this study, a hemicellulose recovery process was integrated with a cold caustic extraction (CCE) process in upgrading paper-grade bleached kraft pulp to dissolving grade. Under the conditions of 15% NaOH, 10% pulp consistency, 30 °C and 1 h, a paper-grade softwood bleached kraft pulp was purified to a dissolving-grade pulp with 97.57% α-cellulose and 1.67% pentosan contents. The spent liquor from the cold caustic extraction process was sequentially extracted with ethanol to precipitate and recover the dissolved hemicelluloses, followed by evaporation to recover the ethanol. After the recovery of hemicelluloses and ethanol, the spent liquor can be reused as the caustic solution for the CCE process without compromising the resulting pulp properties. The results demonstrated that it is feasible to integrate hemicellulose production with the cold caustic extraction process of dissolving pulp production, based on the concept of biorefinery.
INTRODUCTION
Dissolving-grade pulps are purified bleached pulps with high cellulose content of 90-99% for the production of regenerated cellulose fibers and cellulose derivatives, which has great market growth over the last ten years. For high quality of dissolving pulp, the hemicelluloses, lignin and extractives contents have to be low. 1 Cold caustic extraction (CCE) can be a good approach to upgrade prehydrolysis kraft pulp to acetate grade dissolving pulp under the 20~50 ℃ temperature in 10~20% alkaline solution, aiming to selectively dissolve short chain carbohydrates, mainly hemicelluloses. [2][3][4][5] One main challenge of CCE application is the recovery of the relatively high concentration of alkaline solution, due to the presence of hemicelluloses in the CCE filtrate. It has been proposed that a sequential extraction process may be performed to isolate the hemicelluloses from the CCE filtrate for value-added products, 6 and then the caustic liquor can be reused for the CCE process. So the two processes can be integrated by the concept of biorefinery. However, few results have been reported in the literature.
In this study, we systematically investigated the effect of CCE process parameters including temperature, alkaline concentration and pulp consistency on resulting pulp properties and the recovery of hemicelluloses. The CCE filtrate was sequentially extracted with various concentrations of ethanol (25%-50%-75%) to recover the hemicelluloses dissolved in the CCE process, and the recovered hemicelluloses were characterized by HPAEC. After hemicelluloses were recovered, the spent liquor was recycled back to the CCE process.
Materials
A commercial bleached softwood kraft pulp was obtained from a paper mill in China. The pulp sheets were torn into small pieces manually and stored in a sealed plastic bag at room temperature.
Cold Caustic Extraction (CCE)
The CCE experiments were carried out according to a 3-factor Box-Behnken design, as shown in Table 1. The three factors were: pulp consistency (PC), alkali concentration (AC) and temperature (T). The CCE treatment of the pulp was carried out in polyethylene bags heated in a water bath for 60 min. The bags were taken out every 15 minutes for mixing by kneading manually for 10~15 s. After the CCE treatment, the pulp samples were filtered and the filtrate was collected for hemicellulose recovery. The pulp was then washed to neutral pH with distilled water and air-dried for further analyses.
Hemicellulose Recovery
A 3-step sequential extraction process was applied to recover the hemicelluloses from the spent CCE liquor. In the first step, 35 ml of 95% ethanol was added slowly to 100 ml of the CCE filtrate at room temperature under constant stirring to deposit the first fraction of the dissolved hemicelluloses. The first hemicellulose fraction (labeled as H 1 ) precipitated from above process was separated from the liquid phase by centrifugation, washed with 95% ethanol, and vacuum dried at 60 ℃. The clear liquid phase was labeled as AL 1 . In the second step, another 50 ml of 95% ethanol was added to 100 mL of AL 1 to obtain the second hemicellulose fraction (labeled as H 2 ), in the same way as in the first step. The clear liquid phase was labeled as AL 2 . Similarly, in the 3 rd step, another 125 ml of 95% ethanol was added to 100 ml of AL 2 to obtain the third hemicellulose fraction which was labeled as H 3 . The ORIGINAL PAPER www.Bioresources-Bioproducts.com 21 clear liquid from the 3 rd step was marked as AL 3 which was saved for ethanol recovery and caustic recycling.
Reutilization of the CCE Filtrate
The AL 3 liquid was evaporated under vacuum to recover the ethanol, and the remaining caustic liquor was reused for the CCE process as mentioned in the section 2.2. The alkali concentration of the recovered caustic solution was determined by titration with double indicators.
Carbohydrates Analysis
The constituent neutral sugar and uronic acids in the recovered hemicelluloses were determined by high-performance anion exchange chromatography (HPAEC). 10 The neutral sugars and uronic acids in the recovered hemicelluloses were liberated by hydrolysis with 72% H 2 SO 4 for 45 min at 25 ℃ followed by a high temperature hydrolysis at 105 ℃ for 2.5 h by dilution to 3% H 2 SO 4 . After hydrolysis, the samples were diluted and injected into the HPAEC system (Dionex ISC 3000, USA) with an amperometric detector, a CarbopacTMPA-20 column (4 mm × 250 mm, Dionex), and a guard PA-20 column (3 mm × 30 mm, Dionex). Neutral sugars and uronic acids were separated in isocratic 5 mM NaOH (carbonate free and purged with nitrogen) for 20 min, followed by a 0.75 mM NaAc gradient in 5 mM NaOH for 15 min with a flow rate of 0.4 mL/min. Calibration was performed with standard solutions of L-arabinose, D-glucose, D-xylose, D-mannose, D-galactose, glucuronic acid, and galacturonic acid.
Effect of CCE Conditions on Treated Pulp Properties
The pulp properties after the CCE treatments under various conditions were listed in Table 1. Before the CCE treatment, the pulp's viscosity was 742 mL/g, and its α-cellulose and pentosan contents were 86.32% and 5.44%, respectively. In all cases, the CCE treatment increased the α-cellulose content and decreased the pentosan content markedly. The α-cellulose content represents the level of un-degraded and higher-molecular-weight cellulose, while the pentosan content indicates the amount of hemicelluloses as impurity, which should be kept as low as possible for dissolving pulp. 8,9 The results in Table 1 show that after the CCE treatment, the α-cellulose content of the pulp increased from 86.72% up to 97.6%, and the pentosan content decreased from 5.44% to as low as 1.67%, a reduction of almost 70%, while the viscosity change was relatively small. Table 2 shows the statistical analyses of a single CCE parameter's effect on the resulting pulp properties (α-cellulose, pentosan and viscosity). It can be seen that increasing the temperature from 20 to 30 ℃ benefited the viscosity and α-cellulose content, probably due to increased removal of short chain carbohydrates. 1,3 However, further increasing the temperature had negative effect on the viscosity and α-cellulose due to alkaline degradation. These results are in agreement with those reported earlier. [11][12][13] For the effect of pulp consistency on the CCE performance, the analyses in the Table 2 indicate that the CCE performed best at the lowest pulp consistency. This is understandable if we consider the fact that at a lower pulp consistency more caustic solution was available for treating a given amount of fibers.
The caustic concentration had a positive effect on the CCE performance: the higher the caustic concentration, the higher the viscosity and α-cellulose, and the lower the pentosan content of the resulting pulp. Similar effect of enhanced alkaline concentration was reported earlier on the extraction of hemicelluloses. 13 In general, mild reaction temperature, low pulp consistency and high alkalinity favored the CCE process in purifying the pulp, which is in accordance with the results in Table 1 (30 ℃, 15% alkalinity with 10% pulp consistency). Under these conditions, a high grade of dissolving pulp can be produced with high viscosity, high α-cellulose content and low pentosan content.
Recovery of Hemicelluloses from CCE Filtrate
To recover the dissolved hemicelluloses, the filtrate from the CCE process (30 ℃, 15% alkalinity with 10% pulp consistency) was treated by a sequential extraction process with ethanol of different concentration. Hemicelluloses are essentially insoluble in ethanol/water solution, depending on the molecular weight of hemicelluloses and the ethanol concentration.
www.Bioresources-Bioproducts.com 22
The high-molecular-weight fraction of the hemicelluloses can be firstly separated from the CCE filtrate by deposition at a low ethanol concentration.
And then lower Mw fractions of hemicelluloses can be separated sequentially by raising the ethanol concentration in the system by stepwise. Table 3, about 3.87% and 6.88% of hemicelluloses were separated from the CCE spend liquor in 1 st and 2 nd extraction stages, respectively. No carbohydrates were found in the H 3 fraction, suggesting that two stages of ethanol extraction were sufficient to isolate most of the hemicelluloses dissolved in the CCE spent liquor. This finding was in agreement with those reported by Peng et al. 14 The recovered hemicelluloses consisted mainly of xylose and mannose, which is in accordance with the findings by Tian et al. 15 The recovered hemicelluloses may be further treated by an enzymatic saccharification and hydrolysis process to isolate the main monosaccharide for high-value application. 16,17 It is interesting to note that the GluA and GlaA were only identified in the H 1 fraction, but not in other fractions. These results indicated that the recovered hemicelluloses from the H 1 fraction consisted more branched xylans (i.e. arabinoglucuronoxylan). Meanwhile, the Ara/Xyl ratio of the H 1 fraction was higher than that for the H 2 fraction, also confirming that more branched xylans were recovered in the 1 st stage at a lower ethanol concentration. After the hemicelluloses have been recovered by precipitation in ethanol, the liquid mixture was evaporated under vacuum to recover the ethanol. The remainder liquid was essentially sodium hydroxide water solution. The concentration of NaOH in the recovered caustic solution was about 14.6%, slightly lower than the fresh CCE liquor. The decreased NaOH concentration may be explained by i) the conversion of NaOH to Na 2 CO 3 during the treatment; ii) reaction of NaOH with ethanol to form CH 3 COONa. When the recovered CCE caustic solution was applied to the CCE process to replace the fresh caustic solution, the resulting pulp had similar properties as the pulp treated with fresh alkali liquor, in terms of viscosity, α-cellulose and pentosan contents.
These results demonstrate the feasibility of applying the CCE process to an existing paper-grade softwood kraft pulp mill to produce dissolving grades of pulp to meet market demands. As shown in Fig. 1, with the integrated CCE technologies, the pulp mill can produce dissolving grades of pulp, as well as high value products from hemicelluloses.
CONCLUSION
Hemicellulose recovery was successfully integrated with the CCE process for dissolving pulp production. The optimal CCE process conditions were found to be 30 ℃, 10% pulp consistency and 15% caustic concentration. The dissolved hemicelluloses in the CCE process can be recovered and fractionated by sequential precipitation at different levels of ethanol concentrations. It was found that the recovered hemicelluloses consisted mainly of xylose and mannose which can be potentially converted to high value products. More branched xylans were separated in the first isolation stage at a lower ethanol concentration. The CCE spent liquor can be reused after the recovery of the hemicelluloses and ethanol, without affecting the resulting pulp properties in terms of viscosity, α-cellulose and pentosan contents. | 2,681 | 2017-01-25T00:00:00.000 | [
"Materials Science",
"Environmental Science"
] |
The Design of Miniature Frequency Reconfigurable Antenna Based on Inductive Loading Technology
. A circularly polarized (CP) and frequency recon fi gurable microstrip antenna with loading inductive is presented in this paper. The designed antenna is comprised of a radiating patch, four short-circuited grounded metal posts, and four coupling branches. Each coupling branch has an end that is coupled to the shorted ground post and is also connected to the parasitic branches by means of a group of PIN diodes. By controlling the state of the PIN diodes connected to each parasitic branch, the working resonant frequency of the antenna can be changed. In order to further understand the mechanisms of operation of the antenna, the equivalent circuit model was built, and the circuit model of the antenna was analyzed, and this analysis was used for the development of the frequency recon fi gurable microstrip patch antenna. Furthermore, the parameters of speci fi c equivalent circuits can be solved by the three lengths of branch. Meanwhile, the calculated results derived from the given resonant frequency formula for the antenna are in good agreement with the simulation results of the antenna. Simulated results for the input impedance of the antenna are also in good agreement with the calculated values for the equivalent circuit. Finally, the antenna is fabricated and measured, and the measured results show that the antenna can not only achieve frequency recon fi guration at 1.14GHz, 1.21GHz, and 1.39GHz but also accord well with the simulation value, while maintaining a compact size.
Introduction
With the rapid development of wireless communication technologies and applications in recent years, and to satisfy the requirements of many applications, such as mobile and satellite communications, base stations, cognitive radios, navigation, and remote sensing, the reconfigurable antennas and relevant techniques are researched for low cost and space-saving [1][2][3][4][5].Especially, the frequency reconfigurable technology will meet the antenna operating in numbers of multisystems [5][6][7][8], which becomes an alternate antenna in a wide band wireless system.
The inherent disadvantage of microstrip antennas is their narrow bandwidth, which makes it difficult to realize multiband or broadband systems [9].However, the frequency reconfigurable microstrip antenna is a suitable option for antennas working in a wide range of multiband and other special applications, owing to its limited size and low cost [10].
The transmission line model of microstrip patch antenna is proposed by Munson [10].It is a common method for the analysis and optimization of microstrip patch antennas.Sun et al. have given the equivalent circuit of a microstrip antenna with a shorting probe [9,11,12].The analysis method is using simulation software to calculate input impedance and resonance frequency of equivalent circuit and antenna.Further research and further documents always use this analysis method.
Many studies have reported the use of software simulation to analyze parasitic inductance or capacitance in antenna models in order to construct equivalent circuits.Similarly, in [9], the distributed parameters are utilized to increase the compactness of the antenna.Coupled between the patch and the branch is capacitance, and the larger the coupling size, the larger the coupling capacitance.In [13], an antenna is designed that has a frequency reconfigurable feature loaded by patch-shorted circuits and which can switch between three band operating frequencies of BeiDou navigation.As described in [14], two switches were inserted into the slot to serve as three reconfigurable frequency bands to achieve the triple-frequency switched antenna.A reconfigurable PIN diode is introduced in [15] in order to adjust the radiation path of current, which enables frequency reconfiguration.These frequency reconfigurable ways either change the current path of the antenna radiator or control the feed structure or feed mode, thus increasing the complexity and difficulty of the reconfiguration and reducing the radiation efficiency of the antenna.Whereas in our previous work [16], a novel method was proposed by changing the equivalent capacitance of the patch itself by changing the length of the variable branch, thereby obtaining the frequency reconstruction.Unfortunately, this work only gives the preliminary simulation results and does not involve the specific parameter values of the equivalent circuit of the antenna, the in-depth discussion of the equivalent circuit, and the verification of the fabricated antenna.
The analysis shows that changing the length of the variable branch will change its corresponding equivalent capacitance, thus changing its resonant frequency.In this paper, three branch lengths corresponding to three capacitance values are used to determine or solve lumped parameters of equivalent circuits.The resonant frequency of the antenna is derived from the resonant frequency calculation formula of the circuit.Therefore, this research can not only determine the consistency of the input impedance of the antenna and the equivalent circuit but also determine the accuracy of the lumped parameters of each component in the circuit and further determine the accuracy of the resonant frequency.At present, there are few reports about this method.
A frequency reconfigurable antenna is presented in this paper, which is a further in-depth study of our previous work [16].By using PIN diodes and four variable branches, the operation frequency could be controlled at about 1.14 GHz, 1.21 GHz, and 1.39 GHz.The |S 11 | is below -15.7 dB at all operation frequencies, and the measured gain is about 1 dB at 1.14 GHz, 1.6 dB at 1.21 GHz, and 2.39 dB at 1.39 GHz.The dimension is only 20 × 20 × 20 mm 3 , while a bigger test board is used to fix the RF connector.
Basic Mechanism of the Inductive Loading
Frequency Reconfigurable Antenna The illustration of the proposed antenna is shown in Figure 1.At the top of the antenna, there is a radiator board, which is printed by radiator and branches on both sides.Four copper poles serving as inductors and supporting structures are connected between the radiator and the ground board.Meanwhile, two feeding probes could be used to support the radiator board either.The ground board is at the bottom of the antenna (see Figure 1(a)).The coaxial connector is fixed under the ground board, and a 3 dB 90 °hybrid coupler is used to radiate CP wave from the antenna.Figures 1(b) and 1(c) are the two sides of the radiator board.The upper side is called as top layer, and the underside is called as 2nd layer.To facilitate control of the diodes inserted in the variable branches, variable branches are printed on the top layer.The radiator is designed as a rectangular microstrip patch and printed in the 2nd layer, with four slots etched at four corners.Each slot is arranged with a strip, which could bring the coupling capacitance and called as coupling branch (Figure 1(c)).Four coupling strips are connected to the inductance poles.The variable branches, which are connected to the inductance pole either, and L-shaped feeding patches are printed at the top layer.Each variable branch is consisted by three pieces of strips which are connected by 2 PIN diodes.And the states of In this study, the length of variable branches is chosen as 1 mm, 9 mm, and 12 mm which corresponds to the resonant frequency of 1.39 GHz, 1.21 GHz, and 1.14 GHz separately.
The coupling branches and variable branches bring coupling capacitance between them and the radiating patch, while copper poles will generate parasitic inductance.Considering that the coupling capacitance is grounded by the parasitic inductance of the poles, this technique is known as inductive loading.When parasitic parameters change, the resonant frequency also changes, thus forming a frequency reconfigurable antenna.
Equivalent Circuit of Antenna.
The popularity of the transmission line model may be gauged by the number of extensions to this model which have been developed in [17,18].The transmission line model of the rectangle microstrip antenna is shown in Figure 2. In this model, the rectangular microstrip antenna consists of a microstrip transmission line with a pair of loads at either end [19].The input impedance or admittance is calculated by this model.The input impedance or admittance including the patch is considered the transmission line and the loads (Figure 2(a)).So, it must be an impedance generally.It should have a real part and an imaginary part (Figure 2(b)).
In Figure 2(b), the R, R', C3, and C3' only refer to the input impedance or the input admittance of the loads.The resistance R represents the radiation loss at both ends of the transmission line.C3 and C3' represent the whole reactance of the input impedance, which includes the distributed capacitance and inductance.The capacitance and inductance are related to the resonance frequency.In a circuit, see Figure 3, the resonant frequency could be calculated either if the C3 is a capacitance or the parallel capacitance and inductance, whose susceptance value is equal to jωC3.
The equivalent circuit of the proposed antenna is shown in Figure 3.
Because the CP is realized by the same but orthogonality linear polarization feed method, Figure 3 is an equivalent circuit of one of the linear polarizations.The phase difference between the two outputs of the 3 dB coupler is 90 °, thus achieving CP.Where the L1, C1, and C2 are caused by the L-type feeding patch and the radiator and the ground, T1 and T2 are the length of the radiator patch which is divided by the feeding point, and R and R' are the radiation resistance of the antenna.C3 and C3' are the coupling capacitance between radiator and ground board.C4 and C4' are the coupling capacitance between radiator and coupling branches in the slot of the radiator.C5 and C5' are the coupling capacitance between radiator and variable branches.L2 and L2' are the inductive of the inductive loading pole [9].Because C4 and C4' and C5 and C5' are connected to the ground through L2 and L2', the parallel relationship in Figure 3 is formed.
Equivalent Circuit of Antenna.
The resonant frequency of the antenna is determined by the above parameters, which is not mentioned in our literature [16].C5 and C5' are controlled by the length of the variable branches.The states of the diodes are used to control the length of these branches.Therefore, the change of parameters leads to the reconstruction of antenna resonant frequency.
The two diodes of each branch exhibit transmission behavior when they are both on, and the resonant frequency decreases as the branch is extended.Nonetheless, when two diodes of each branch are turned off, the coupling capacitance between the branch and the radiator patch is eliminated.Therefore, C5 and C5' are nearer to zero.
So, the resonance frequency could be calculated by the half-side circuit of Figure 3.It is given in Calculate the parameter values of the equivalent circuit in Figure 3 through simulation and solving Formula (1), as shown in Table 1.
C5 and C5' are the capacitance of variable branches.The length is corresponding to the coupling capacitance between branches and radiator.In this study, the branches are divided into three segments while connected by two PIN diodes.When the length is 12 mm, 9 mm, and 1 mm, the coupling capacitance is 3.05 pF, 2.26 pF, and 0.85 pF.Using Formula (1), the resonance frequencies of the antenna are 1.144 GHz, 1.214 GHz, and 1.393 GHz.
Simulation of the Antenna
3.1.Resonance Frequency of Antenna.As calculated previously by Formula (1), the resonance frequency is 1.144 GHz, 1.214 GHz, and 1.393 GHz.Using antenna simulation software, the resonance frequency could be gained either.
Figure 4 shows the simulation results of the antenna reflection coefficient in several different states, in which the lengths of the variable branches are 12 mm, 9 mm, and 1 mm, and the resonant frequencies of the corresponding antennas are 1.140 GHz, 1.216 GHz, and 1.392 GHz, respectively.The |S 11 | of three resonance frequencies is -18.7 dB, -15.7 dB, and -21.7 dB, respectively.The simulated results agree well with the calculated results.
Input Impedance Simulation of the Antenna and
Equivalent Circuit.The input impedance of the antenna can be calculated by the antenna simulation software and the equivalent circuit model of the antenna.A simulation of the results is depicted in Figures 5-7.It is the impedance simulation results of three frequencies related to three lengths (L) of variable branches.
Figure 5(a) shows the input impedance curve with the branch length of 12 mm, while the C5 = C5 ' = 3 05 pF.Compared with the simulation result of equivalent circuit in International Journal of RF and Microwave Computer-Aided Engineering Figure 5(b), the difference of the real part is only 3 Ω while the imaginary part is 7 Ω. Figure 6 is the input impedance curve with the branch length of 9 mm, while the C5 = C5 ' = 2 26 pF.Compared with the simulation result of the antenna, the difference of the real part is 7 Ω while the imaginary part is 6 Ω. Figure 7 is the input impedance curve with the branch length of 1 mm, while the C5 = C5 ' = 0 85 pF.Compared with the simulation result of the antenna, the difference of the real part is 5 Ω while the imaginary part is 9 Ω.The results show that the simulation of the antenna software is in good agreement with the calculation curve of the equivalent circuit.
In the previous research, the input impedance of the microstrip antenna was calculated by software simulation, and then, the equivalent circuit model was built with this as the target, and the simulation software was used to approach the target to obtain the parameter values.It is based on comparing the input impedance curves of the antenna and the equivalent circuit, and the results of the two are consistent because of the same input impedance.In this paper, according to the equivalent circuit model, the resonant frequency calculation formula of the circuit is given.With fixed parameters (L2, C3, and C4), one of the three reconfigurable resonant frequencies can be obtained by Formula (1), and the input impedance and resonant frequency of L2, C3, and C4 are adjusted to be consistent with the antenna simulation results.Then, Formula (1) is used to calculate C5 at the other two reconfigurable frequencies, and
Radiation Pattern of the Frequency Reconfigurable
Antenna.An analysis of the performance of the proposed antenna at different reconstruction frequencies is presented.The frequency reconstruction can only be achieved by adjusting the state of the PIN diode embedded in each variable branch, as well as obtaining a compact structure for the antenna.
Figure 8 depicts the radiation pattern of the proposed antenna at three frequency points in two orthogonal planes.As a result, the simulated gains at the three reconstructed frequency points are 1 dB, 1.6 dB, and 2.4 dB, and the corresponding 3 dB beamwidths are also 135 °, 125 °, and 102 showing a wide beam characteristic.Moreover, the PIN diode embedded for reconfiguration has a slight effect on the radiation pattern.The radiation characteristics of the wide beam indicate that the antenna has the potential to be applied in satellite navigation systems.
Measurement and Discussion
Figure 9 shows the fabricated and test scene of the antenna.The variable stripes with diodes are etched on the top layer, which is convenient for controlling the reconfiguration.The radiator has printed the 2nd layer, which did not influence the radiation.
The state of the diode used for reconstruction is controlled by the DC bias circuit.The schematic diagram of the control circuit is shown in Figure 10, two capacitors could prevent the DC shorted to each other, and the two inductors could prevent the RF current leak through the DC conductor line.The states of antenna change with the states of diodes are shown in Table 2.
The measured |S 11 | is shown in Figure 11.Compared with Figure 4, the resonant frequency of measurement is fit to the simulation results.The radiation pattern measured results are shown in Figure 12.Compared with simulation results, the pattern is matched to each other.The simulated and measured gains are slightly different.The simulation and measured gains are shown in Table 3.
Measurement results are a bit lower.It is caused by the loss of the feeding line and the substrate material.In the simulation, the loss tangent is given.But it is not very accurate, International Journal of RF and Microwave Computer-Aided Engineering the simulated and measured three frequency points, respectively.It can be seen from the figure that the efficiency of the antenna at these three reconfigurable frequency points is over 80%.According to the measured axial ratio, the designed reconfigurable antenna also shows good circular polarization characteristics.
Conclusions
This paper presents a frequency reconfigurable antenna and designs and analyzes the measurements of the antenna.By in-depth investigation, the equivalent circuit model of the frequency reconfigurable antenna is given, and the formula for calculating the resonant frequency of the antenna is derived according to the circuit model.Through the calculation of the formula, the results show that the calculation results are in good agreement with the simulation and measurement results.The realization of frequency reconfiguration is that the PIN diode is used to change the length of the variable branches which are printed on the top layer.The resonant frequency of the antenna is controlled by the state of diodes.Changing the length of the branch does not change the pattern.By adding more diodes, the proposed design can obtain more frequency substitution values.This method can be used for patch antennas operating in other frequency bands.The area of the radiator is smaller than the microstrip antenna without the loading probe.
Figure 5 :
Figure 5: Input impedance comparison of antenna and circuit (L = 12 mm): (a) antenna simulation result and (b) equivalent circuit simulation result.
Figure 6 :
Figure 6: Input impedance comparison of antenna and circuit (L = 9 mm): (a) antenna simulation result and (b) equivalent circuit simulation result.
Figure 7 :Figure 8 :
Figure 7: Input impedance comparison of antenna and circuit (L = 1 mm): (a) antenna simulation result and (b) equivalent circuit simulation result.
Figure 9 :
Figure 9: Fabricated photographs of the frequency reconfigurable antenna.
Figure 10 :
Figure 10: Schematic diagram of the control circuit.
Figure 12 shows the radiation patterns of three frequency points, respectively.It can be seen from the figure that the measured results show good circular polarization and wide beamwidth performances.The measured 3 dB beamwidths of the three frequency points are 140 °, 111 °, and 102 °, respectively.At the same time, Figure13also shows the efficiency and axial ratio of
Figure 12 :
Figure 12: Measured and simulated results of the radiation pattern of the antenna.
Figure 13 :
Figure 13: Measured and simulated results of efficiencies and axial ratio.(a) Efficiencies and (b) axial ratio.
Table 1 :
Parameter of the equivalent circuit.
Table 2 :
State control of the diodes.
6International Journal of RF and Microwave Computer-Aided Engineering as given by the manufacturer.And the measurement error has always existed.This can cause the difference between simulations and measurements.Measurement results are very near to the simulation results.
Table 3 :
Comparison of gain between simulation and measurement. | 4,333 | 2023-11-23T00:00:00.000 | [
"Engineering",
"Physics"
] |
Anomalous Weak Values Without Post-Selection
A weak measurement performed on a pre- and post-selected quantum system can result in an average value that lies outside of the observable's spectrum. This effect, usually referred to as an"anomalous weak value", is generally believed to be possible only when a non-trivial post-selection is performed, i.e., when only a particular subset of the data is considered. Here we show, however, that this is not the case in general: in scenarios in which several weak measurements are sequentially performed, an anomalous weak value can be obtained without post-selection, i.e., without discarding any data. We discuss several questions that this raises about the subtle relation between weak values and pointer positions for sequential weak measurements.Finally, we consider some implications of our results for the problem of distinguishing different causal structures.
INTRODUCTION
All quantum measurements are subjected to a fundamental trade-off between information gain and disturbance of the measured system. In particular, one can perform weak measurements that provide little information but only weakly perturb the system. A particularly interesting situation arises when weak measurements are combined with post-selection [1]. This can be conveniently described within the von Neumann model of quantum measurements, where the quantum system to be measured is coupled via a joint unitary operation to another quantum system, the pointer, which represents the measurement device. The measurement is then completed by performing a strong measurement of the pointer.
More formally, consider a system initially prepared (or pre-selected) in a pure state |ψ , and an observable to be weakly measured on it. The system-pointer interaction is generated via a Hamiltonian of the form H = γ ⊗p, wherep denotes the momentum operator acting on the pointer. The latter is initially in a state |ϕ(0) , which we shall take here to be a Gaussian wave packet centred at a position x = 0 with spread σ. Assuming that we are in the weak measurement regime, with the coupling constant γ and interaction time ∆t such that g := γ∆t is small enough compared to the spread of the pointer, the global state after the coupling is given by e −iĤ∆t |ψ |ϕ(0) ≈ (1 1 − igÂp) |ψ |ϕ(0) (1) (where tensor products are implicit, and taking = 1).
For simplicity we will henceforth choose units so that g = 1; the strength of the measurement will then be controlled solely by the pointer spread σ, and the validity of the weak regime will depend only on this being sufficiently large (see Appendix for details). Next, the system is post-selected onto the state |φ (e.g. via a strong projective measurement). The final state of the pointer is then (up to normalisation) where is the so-called weak value of the observable given the pre-selection in the state |ψ and post-selection in the state |φ [1]. The mean position of the pointer is thus displaced (via the displacement operator e −iA φ ψp ; see Appendix) to x ≈ ϕ(0)| e iA φ ψpx e −iA φ ψp |ϕ(0) = Re(A φ ψ ).
Note that the definition (3) of a weak value can be generalised to post-selections on a given result for any general quantum measurement [2, 3]. In particular, a trivial, deterministic measurement of the identity operator 1 1 amounts to performing no post-selection. This allows one to also consider a weak value with no post-selection, defined (see Appendix) as With this definition, Eq. (5) gives x = A 1 1 ψ = Re(A 1 1 ψ ): we recover the same relation as in Eq. (4), although now A 1 1 ψ is restricted to lie in [λ min (Â), λ max (Â)] since here it is simply equal to the expectation value ofÂ.
The phenomenon of a weak value outside the spectrum of is referred to as an "anomalous weak value" [1, 4, 5]. This has been observed in many experiments [6][7][8], and appears to be directly linked to various (a priori unrelated) areas such as tunnelling times [9] and fast light propagation [10,11]. In practice, anomalous weak values allow for the detection and precise estimation of very small physical effects [12][13][14][15], via a form of signal amplification. While astonishing at first sight, anomalous weak values can in fact be intuitively understood in terms of destructive interference of the pointer state, which occurs as a result of post-selection. With this in mind and given the rudimentary analysis above, it is rather natural to attribute the origin of anomalous weak values to the presence of post-selection; this opinion indeed seems to be widely shared in the community.
Here we show, however, that this is not the case in general, and that anomalous weak values can in fact be obtained deterministically, without any post-selection at all. Specifically, we consider a situation in which two successive weak measurements are performed on a quantum system. The experiment thus involves two pointers, one associated to each weak measurement. Considering observables that are simply given by projectors, one expects to find the mean position of each pointer between 0 (the system's state being orthogonal to the projector) and 1 (the system's state aligning with the projector). Yet, we will see that the average of the product of the pointer positions can become negative. This may be understood in terms of the second measurement acting as an effective post-selection of the system, thus creating the desired interference. Importantly however, no data is discarded. Below, after discussing in detail a simple example of this effect, we provide more general insight and results on anomalous weak values and pointer positions obtained with no post-selection.
ILLUSTRATIVE EXAMPLE
To start with, let us consider a qubit system initially prepared in the state |0 , undergoing a sequence of two weak von Neumann measurements of the projection observables |ψ j ψ j | (j = 1, 2), where the states |ψ j and their orthogonal states |ψ ⊥ j are defined as To each measurement is associated a pointer in the state |ϕ j (x j ) , where x j is the mean position of the pointer wavefunction. The two pointers are initially independent, and both centred at x j = 0. The initial state of the system and pointers is therefore Following the von Neumann measurement procedure described earlier with interaction HamiltoniansĤ j = γ j |ψ j ψ j |p j , the average post-measurement position of the corresponding pointer is (with appropriate units so that γ j ∆t j = 1 as before) x j = 1 if the state of the system is |ψ j ; if the state is |ψ ⊥ j then the pointer does not move. The state of the system and pointers after the interaction with the first pointer is thus After interacting with the second pointer, it evolves to |ψ 2 |ϕ 1 (1) |ϕ 2 (1) + 3 4 |ψ 2 |ϕ 1 (0) |ϕ 2 (1) Tracing out the system, one finds that the joint pointer state is a mixture of the following two states, with probabilities p (1) = 5/8 and p (0) = 3/8 respectively. Finally the positions of the pointers are measured. The quantity of interest is the average of the product of the pointer positions, i.e., the expectation value x 1 ⊗x 2 . Since both states in Eq. (11) are product states of the two pointers it follows that, for each state in the mixture, x 1 ⊗x 2 = x 1 x 2 . As x 2 = 1 for |Φ (1) and x 2 = 0 for |Φ (0) , the expectation value of x 1 ⊗x 2 is thus simply p (1) Φ (1) |x 1 |Φ (1) .
Note that we have not yet specified the strength of either measurement. Considering Gaussian pointers with widths σ j for each measurement, we find (see Appendix) Notice that this quantity depends on σ 1 but not on σ 2 : the strength of the second measurement has no effect here.
Since both observables being measured are projectors with spectra {0, 1}, one would naturally expect an average value within the range [0, 1]. Independently of the strength of either measurement, each pointer, taken individually, indeed has an average position in [0, 1]: specifically, x 1 = 1/4 and x 2 = 5/8 (see Appendix). In the regime where the first measurement is strong (i.e. σ 1 → 0), Eq. (12) gives x 1 ⊗x 2 ≈ 1/16, which is consistent with the above argument. However, if the first measurement is sufficiently weak (i.e. σ 1 is large enough), the average value can become negative. In the limit σ 1 → ∞ we get This pointer reading is anomalous in that it gives an average value outside of the natural range. As we will discuss in more detail below, this result can be linked to an anomalous weak value without post-selection, (|ψ 2 ψ 2 | · |ψ 1 ψ 1 |) 1 1 0 := 0|ψ 2 ψ 2 |ψ 1 ψ 1 |0 (see Eq. (18) below); specifically, we have here We emphasise that this anomalous value is obtained despite the absence of post-selection. This effect can nevertheless be understood intuitively by considering that the second measurement acts as an effective post-selection on |ψ 2 , as the corresponding pointer moves only in this case. This becomes apparent upon rewriting the above weak value as which differs from the standard weak value (|ψ 1 ψ 1 |) ψ2 0 for a post-selection on |ψ 2 only by the factor | ψ 2 |0 | 2 , which is the probability that the projection of |0 onto |ψ 2 is successful. As it turns out, this factor ensures in particular that the anomalous weak value without postselection cannot be arbitrary large, a fact that we prove further below. For a sequence of two projection observ-ables andB (with eigenvalues 0 and 1), the above value of −1/8 for the real part is indeed the most anomalous value obtainable (see Appendix).
ANALYSIS FOR ARBITRARY OBSERVABLES
In order to place the previous example in a more general framework, let us recall some facts about sequential weak measurements of noncommuting observables [16][17][18]. To this end, consider a system prepared in the pure state |ψ , which is subjected to a sequential weak measurement of the observables thenB, before being postselected onto the state |φ . The system-pointer interaction Hamiltonians areĤ 1 = γ 1Âp1 andĤ 2 = γ 2Bp2 . We will choose again, for simplicity, the coupling constants and interaction times such that γ j ∆t j = 1, and take Gaussian pointers initially in the states |ϕ 1 (0) and |ϕ 2 (0) with widths σ 1 and σ 2 , which dictate the measurement strengths.
However, while the notion of an anomalous weak value for single (non-sequential) weak measurements is intimately linked to the pointer displacement (and even justified) by the relation x = Re(A φ ψ ), the relationship between the mean pointer positions and (BA) φ ψ is more subtle for sequential weak measurements. In the presence of postselection, it has instead been shown [16,19] that within the weak regime (with large enough widths σ 1 and σ 2 ). This cautions that some care must be taken when linking (possibly anomalous) pointer positions to weak values.
Let us now generalise the sequential weak value of Eq. (6) to the case without post-selection, by defining, in a similar way to before, the sequential weak value with no post-selection as Connecting this to the pointer positions, we prove in the Appendix that, contrary to Eq. (17) (which was obtained with post-selection), we recover here the direct relation as anticipated already in Eq. (14), which holds as long as the first measurement is sufficiently weak. This justifies that our earlier illustrative example could indeed be interpreted as yielding an anomalous weak value without post-selection. Crucially, although for a single measurement without post-selection A 1 1 ψ is simply the expectation value ofÂ, no such interpretation can be given to (BA) 1 1 ψ sinceB is only Hermitian -and thus defines an observable -if andB commute. In particular, this implies that (BA) 1 1 ψ need not be contained within the interval where Λ min(max) (Â,B) = min(max) k,ℓ λ k (Â)λ ℓ (B), as one one would naturally expect for the product of outcomes for a measurement of thenB.
Nevertheless, as we noted after Eq. (15), the value of (BA) 1 1 ψ cannot be amplified arbitrarily. It is possible to place a more quantitive bound on the values that it can in fact take. Using the Cauchy-Schwartz inequality, we indeed have (where · is the spectral norm). Thus, although one can obtain anomalous weak values without post-selection, their magnitude cannot be pushed outside what one can obtain using strong measurements. The bound above implies in particular that for observables with symmetric spectra, the real part of the weak value -and therefore the mean product of pointer positions, see Eq. (19) -cannot be anomalous; anomalous pointer positions are only obtained for observables with asymmetric spectra, such as projection observables. Nevertheless, one can also obtain complex weak values for observables with symmetric spectra, which can similarly be considered anomalous. Take, for example, a system initially prepared in the (+1)-eigenstate |0 of the Pauli matrixσ z , on which a sequential weak measurement of the Pauli observablesσ y andσ x is performed. One thus obtains (σ x σ y ) 1 1 0 = i. The imaginary part of the weak value here can be detected by measuring the pointer momenta [15,19] (see Appendix). We note again that such a complex anomalous weak value cannot be obtained without post-selection with only a single weak measurement.
MORE MEASUREMENTS
Eq. (20) might bound how anomalous a weak value can be without post-selection, but it is not generally tight. For two projection observables andB (with eigenvalues ±1), for example, it only implies a bound Re[(BA) 1 1 ψ ] ≥ −1; nevertheless, as we prove in the Appendix, the value of −1/8 obtained earlier for the real part of the weak value is the most negative value that one can obtain. Can one do better by considering longer sequences of successive weak measurements? Here we will see that this question has a subtle answer: the weak value itself can approach −1, but this will not mean the average product of the pointer positions does so as well.
For a sequence of n observables 1 , . . . , n to be measured weakly on the state |ψ before a post-selection on |φ , the sequential weak value is defined (following, e.g., Ref. [16]) as When no post-selection is performed, this can be generalised to in analogy to the cases discussed earlier. As we show in the Appendix, a similar bound to Eq. (20) can be derived, namely For n projection observables, this implies the bound Re[(A n · · · A 1 ) 1 1 ψ ] ≥ −1. As it turns out, it is possible to obtain an anomalous sequential weak value without post-selection approaching −1 and thus saturating this bound in the limit n → ∞. To see this, take the initial state of the system to be |ψ = |0 and consider the sequence of n qubit projectors j = |a j a j | with |a j = cos( jπ n+1 ) |0 +sin( jπ n+1 ) |1 for j = 1, . . . , n. This sequence of weak measurements gives Note that for n = 2 this coincides precisely with the explicit two-measurement example we began with. As discussed above, for two sequential weak measurements in the absence of post-selection, the mean product of the pointer positions gives precisely the real part of the sequential weak value; see Eq. (19). However, for n > 2 measurements this direct relationship is broken and the mean product of the pointer positions corresponds instead to a mixture of sequential weak values for 2 n−2 different permutations of the observables (see the Appendix for an explicit expression). For example, for n = 3 we have, in the weak regime, The real part of (A 3 A 2 A 1 ) 1 1 ψ is thus not directly observed. However, as we show in the Appendix, its value (as well as the imaginary part) can nonetheless be deduced experimentally by measuring several different expectation values of the products of pointer positions and momenta [16].
Interestingly, by numerically minimising the mean product of the pointer positions for sequences of up to 5 projection observables, we were unable to obtain a value smaller than −1/8, and we conjecture that this is in fact the case for all n. Thus, although the weak value itself can be brought arbitrarily close to −1, it seems that additional sequential weak measurements may not lead to "more anomalous" pointer positions. This behaviour highlights oft-overlooked subtleties in the connection between anomalous weak values and pointer positions: for individual weak measurements, there is a direct correspondence between the pointer position and (the real part of) the weak value, and an anomalous weak value has an immediate physical relevance. For sequential weak measurements, a distinction must be made between anomalous weak values and anomalous pointer positions (with post-selection, this is already the case for two measurements; see Eq. (17) or Ref. [16]).
This divergence between weak values and pointer positions for sequential weak measurements means that, in general, it is more difficult to give a clear physical interpretation to sequential weak values, anomalous or not. Indeed, while some authors have argued that weak values for single weak measurements should be considered real properties of quantum states with direct physical meaning [20,21], it is unclear whether such arguments are justified for sequential weak values [16] given the lack of examples of physical scenarios where they play a direct, crucial role.
FURTHER DISCUSSION
Compare the situation of a sequential weak measurement of two observables andB with the alternative in which a bipartite system |ψ ab ∈ H a ⊗ H b is prepared and andB are weakly measured on the two different substituent systems. One can view this either as a measurement of the joint observableÂ⊗B (with two different pointers, one coupled to each observable) or a sequential measurement of the commuting observables ⊗ 1 1 and 1 1 ⊗B. In the absence of any post-selection, one has (A ⊗ B) 1 1 ψ ab = ψ ab | ⊗B|ψ ab which, being just an expectation value, cannot lie outside the spectrum of the product observable ⊗B. For tensor product measurements, an anomalous weak value is thus unobtainable without post-selection.
This observation raises some interesting implications. Consider for example a scenario in which two parties, Alice and Bob, each operate in a closed laboratory. Each receives a system, performs a weak measurement, and sends the resulting system out; they then come together to jointly measure their pointers. By repeating this many times (or on a large number of systems), they thus determine x 1 ⊗x 2 . If Alice and Bob have no knowledge of their causal relationship, they could unknowingly be weakly measuring the same system at different times (either Alice then Bob, or vice versa), or measuring different parts of a (potentially entangled) bipartite system. By observing an anomalous weak value they can differentiate between these two scenarios.
The problem of distinguishing these two causal structures for quantum systems -the former is known as a direct cause relationship, while the later a common cause relationship, since any correlations must be due to a (possibly quantum) common cause -has been the subject of recent interest; see, e.g., Refs. [22][23][24][25][26]. An anomalous weak value thus provides a novel way to witness a direct causal relationship and distinguish between these cases. We leave it as an open question whether such a witness can be found whenever Alice and Bob are capable of signalling to each other; i.e., if whenever they are connected by a quantum channel of nonzero capacity they can always find a pair of observables to measure that would generate an anomalous weak value without post-selection.
Recently, there has also been substantial interest in quantum processes that are not consistent with any definite (possibly stochastic) causal ordering [27], and practical approaches to witness such "indefinite" causal orders have been developed [28,29] and experimentally tested [30,31]. It would be interesting to see whether indefinite causal orders may also be witnessed by, for example, producing larger anomalous weak values than possible in a well-defined causal structure. [ The weak measurement regime can in general be defined for any type of von Neumann measurement interaction scheme, as introduced in the main text, by comparing the various parameters that describe it: the strength of the measurement interaction (γ), the time of the interaction (∆t), the width of the measurement pointer (σ), the eigenspectrum of the observable being measured as well as the weak values to be considered. The relation between weak values and pointer positions can then be obtained by taking the appropriate limits.
As the point of our paper is to analyse specific cases of anomalous weak values and anomalous pointer positions, for simplicity we choose a specific form for the pointer states, namely, Gaussian states. By a Gaussian pointer, we mean a pointer whose state |ϕ(a) is described by a Gaussian wavefunction as follows: where {|x } x is a continuous eigenbasis of the pointer positionx. For a ∈ R, |ϕ(a) is properly normalised; for a complex value of a, its norm is e Im(a) 2 4σ 2 . The mean position of the pointer in the state |ϕ(a) (possibly after renormalisation) is x = Re(a) and its variance is x 2 − x 2 = σ 2 , while the mean value of the momentum operator Prior to the measurement, we always start with a = 0. Note that an operator of the form e −iαp (for any α ∈ C) acts as a displacement operator, such that e −iαp |ϕ(a) = |ϕ(a + α) .
Let us clarify here the conditions that define the standard weak measurement regime, under which the approximations of Eqs. (1)-(2) of the main text are valid. As everywhere in the paper, we choose units such that g = γ∆t = 1. By considering the spectral decomposition = k a k |a k a k | of the observable under consideration and the completeness relation 1 1 = k |a k a k | (with {|a k } k an orthonormal basis of the system Hilbert space), one can write e −iÂ⊗p = k |a k a k | ⊗ e −ia kp and 1 1 − i ⊗p = k |a k a k | ⊗ (1 1 − ia kp ), so that the difference between the left and right hand sides of Eq. (1) is The approximation of Eq. (1) is valid if for each k (for which | a k |ψ | is non-negligible), the norm of |δ k is small enough (compared e.g. to that of the lhs of Eq. (1), which is 1). Using Eq. (A1), one finds Similarly, the difference between the two lines of Eq.
(2) is (ignoring the common prefactor φ|ψ ) which is also small if Thus the weak regime is valid whenever the two conditions (A4) and (A6) are fulfilled.
Relating weak values to pointer positions and momenta
The case of a single weak measurement It will be useful to relate here the formula for a weak value to the pointer position and momentum in a more general setting than that considered in the main text, where the initial state is not considered a priori to be pure, and where post-selection is conditioned on a given result of an arbitrary Positive-Operator Valued Measure (POVM) measurement on the system, rather than a projective measurement. In such a setting, Eq. (3) can be generalised to which now defines the weak value of the observableÂ, given the pre-selection in the state ρ and post-selection by the POVM element E. This definition was first proposed explicitly in Ref.
[2] (although earlier alluded to in Ref. [32]), and shown to be indeed the natural generalisation of the standard definition (3) of a weak value. Note already that this definition indeed reduces to Eq. (3) if the preparation is a pure state |ψ (i.e. for ρ = |ψ ψ|) and the post-selection is a projection onto another pure state |φ (for E = |φ φ|). It also allows one to define a weak value with no post-selection by taking a trivial POVM element E = 1 1, which indeed reduces to the definition of Eq. (6) in the case of a pure state ρ = |ψ ψ| -and which simply coincides here, in the case of a single observable, with the expectation value ofÂ; note in particular that, contrary to a general weak value, the weak value with no post-selection is linear in the pre-selected state.
In the von Neumann measurement scenario that we consider here, one thus prepares the density matrix ρ, weakly measures (= k a k |a k a k |) (with a pointer in a Gaussian state as described above), and finally post-selects an outcome corresponding to the POVM element E. The initial density matrix of the system and pointer is given by Under the interaction between the system and the pointer, the joint state evolves to Due to the post-selection upon E, the state of the pointer is then projected onto the (unnormalised) state Tr E |a k a k | ρ |a ℓ a ℓ | |ϕ(a k ) ϕ(a ℓ )| (A11) (where Tr S is the partial trace over the state of the system).
The expectation value of the position of the pointer, given that the post-selection was successful, is mn Tr E |a m a m | ρ |a n a n | ϕ(a n )|ϕ(a m ) .
(A12)
Evaluating the expressions in the fraction above for the Gaussian pointer of Eq. (A1) (with a k , a ℓ , a m , a n ∈ R), and taking the weak limit approximation in which σ ≫ |a k − a ℓ |, σ ≫ |a m − a n |, one finds, to the lowest order, so that we are left with x ≈ kℓ Tr E |a k a k | ρ |a ℓ a ℓ | (a k + a ℓ )/2 mn Tr E |a m a m | ρ |a n a n | = Tr(EÂρ) + Tr(EρÂ) /2 Tr(Eρ) = Re Tr(EÂρ) where we used the spectral decomposition of and the cyclic property of the trace, together with the fact that E, and ρ are Hermitian and thus Tr(EPÂ) = Tr((EPÂ) † ) * . Recalling the generalised definition (A7) of a weak value, we thus find that the mean position of the pointer (when the post-selection is successful) is, as in Eq. (4), One may also consider measuring the expectation value of the momentum of the pointer instead, conditioned again on a successful post-selection: mn Tr E |a m a m | ρ |a n a n | ϕ(a n )|ϕ(a m ) .
(A16)
The relevant quantity for the pointer states is, in the weak regime approximation, from which (together with Eq. (A13)) we find that the expectation of the pointer's momentum is mn Tr E |a m a m | ρ |a n a n | = 1 2σ 2 Tr(EÂρ) − Tr(EρÂ) /2i Tr(Eρ) = 1 2σ 2 Im that is, Note that, unlike the expression for x , this depends explicitly on the width of the pointer. The expectation value of the momentum is thus directly linked here to the imaginary part of the weak value. Using Eqs. (A15) and (A19), one may therefore recover both the real and imaginary parts of A E ρ from the expectation values of the pointer's position and momentum (in a regime where σ is large enough to ignore higher order terms, but not so small as to render the term that remains above, with the pre-factor 1 2σ 2 , unmeasurable).
Two sequential weak measurements
Let us now turn to the sequential measurement of two observables andB. The sequential weak value (BA) φ ψ for a system prepared in the pure state |ψ and post-selected in |φ , defined in Eq. (16), was introduced in Ref. [16]. As in the previous section, one may consider a more general setting where the initial state ρ is not necessarily assumed to be pure, and the post-selection is conditioned on a general POVM element E. In such a case, the definition of Eq. (16) can, similarly to Eq. (A7), naturally be generalised to (BA) E ρ := Tr(EBÂρ) Tr(Eρ) . (A20) One can indeed verify that one recovers Eq. (16) for ρ = |ψ ψ| and E = |φ φ|. As before, this definition also allows one to define a sequential weak value with no post-selection by taking the trivial POVM element E = 1 1, as in Eq. (18) for the case of a pure state ρ = |ψ ψ|. Again, and contrary to a general sequential weak value, the sequential weak value with no post-selection is linear in the pre-selected state; note, however, that it no longer coincides with an expectation value, as in general the productB is not Hermitian, and thus does not define a valid observable.
We consider here a sequential von Neumann measurement scenario where two separate Gaussian pointers (labelled by the subscripts j = 1, 2) are used to measure (= k a k |a k a k |) andB (= m b m |b m b m |) on a system prepared in the state ρ and post-selected on a POVM element E. Similarly to the analysis in the previous section, the final (unnormalised) state of the two pointers after the post-selection is given by (with implicit identity operators) Using the weak regime approximations of Eq. (A13), for both the weak measurement of and ofB, we find that the expectation value of the product of the pointer positions, given that the post-selection was successful, is The real and imaginary parts of the weak value (BA) E ρ = Tr(EBÂρ) Tr(Eρ) are thus not directly given by the mean values of the pointer positions and momenta (conditioned on a successful post-selection), as observed previously in Ref. [16,19], but can still easily be recovered by combining the mean values above as follows: Nevertheless, if no post-selection is made (E = 1 1), then the two summands in Eq. (A22) and in Eq. (A23) are equal, and one again directly obtains (as in Eq. (19) for x 1 ⊗x 2 ). Moreover, these expressions hold as long as the weak regime is applicable for the first measurement, irrespective of the strength of the second. To see this, note that, in the absence of post-selection, Eq. (A21) reduces to Eq. (A13) thus holds with exact equalities for the terms in Tr(x 1 ⊗x 2 η) and Tr(p 1 ⊗x 2 η) corresponding to the second pointer, and the weak regime approximation is thus only required for the first measurement. This also justifies further the claim that the second measurement can be seen as performing an effective post-selection, since it can be taken to be arbitrarily strong.
The example of an imaginary anomalous weak value without post-selection using Pauli observables described in the main text, for which (σ xσy ) 1 1 0 = i, can, from Eq. (A27), thus be observed by measuring p 1 ⊗x 2 .
Generalisation to n sequential weak measurements Considering now n sequential weak measurements, the previous definition of a weak value, for a system prepared in the (possibly mixed) state ρ and post-selected on a POVM element E can be generalised to (A n · · · A 1 ) E ρ := Tr(EÂ n · · ·Â 1 ρ) Tr(Eρ) .
which concludes the proof.
Note, furthermore, that by the linearity of the weak value with no post-selection with respect to the pre-selected state, (BA) 1 1 ρ = i q i (BA) 1 1 ψi for a mixed state ρ = i q i |ψ i ψ i |, which implies that the bound above also holds for (BA) 1 1 ρ ; that is, Re[(BA) 1 1 ρ ] ≥ −1/8 for any two projection observables andB and any mixed state ρ.
We further note that by using the linearity of the weak value with no post-selection and the triangle inequality, it is easy to see that Eq. (23) also holds for a preparation in any mixed state ρ = i q i |ψ i ψ i |: | 7,666.4 | 2018-05-23T00:00:00.000 | [
"Physics"
] |
Calixarenes: Generalities and Their Role in Improving the Solubility, Biocompatibility, Stability, Bioavailability, Detection, and Transport of Biomolecules
The properties and characteristics of calix[n]arenes are described, as well as their capacity to form amphiphilic assemblies by means of the design of synthetic macrocycles with a hydrophilic head and a hydrophobic tail. Their interaction with various substances of interest in pharmacy, engineering, and medicine is also described. In particular, the role of the calix[n]arenes in the detection of dopamine, the design of vesicles and liposomes employed in the manufacture of systems of controlled release drugs used in the treatment of cancer, and their role in improving the solubility of testosterone and anthelmintic drugs and the biocompatibility of biomaterials useful for the manufacture of synthetic organs is emphasized. The versatility of these macrocycles, able to vary in size, shape, functional groups, and hydrophobicity and to recognize various biomolecules and molecules with biological activity without causing cytotoxicity is highlighted.
In this article, the characteristics and properties that enable the calixarenes to self-assemble in order to form amphiphilic assemblies, micelles of cylindrical or ellipsoidal geometry, vesicles, bilayers, liposomes, and nanoparticles such as nanocapsules and nanospheres that allow the encapsulation, stability, bioavailability, solubility, or even the biological activity of the drugs are reviewed. Also reviewed are the types of molecules that can be recognized by calixarenes, their interactions, and the techniques that allow their determination and the evaluation of the stoichiometry of the complex formed in solution and solid state. In particular, calixarene toxicity, their characteristics, the variation in the size of the cavity and in their functional groups, and their applications, as well as the advantages of these macrocycles, are described.
Summary Description of Its Structure
Polyhydroxylated platforms such as calixarenes, resorcinarenes, or pyrogalloarenes ( Figure 1) are a very interesting class of compounds. They have a remarkable ability to act as receptors for a variety of guest species, depending on their structural properties [23][24][25][26][27], which can be modified by changing the size of the substituents or by adding functional groups as a part of the scaffold [28][29][30][31][32]. Among the described macrocyclic compounds, calixarenes are probably the most promising for application in the area of host-guest recognition of toxicological molecules. Their synthetic availability, low toxicity, and presence of reactive sides are characteristics that make them relevant within supramolecular chemistry. The calixarenes are a family of macrocyclic compounds with a variable number of phenol units linked by methylene bridges in ortho position [33]. The number of units of aromatics can be between 4 and 20, although the calixarenes of 4, 5, 6, 7, and 8 are the most common [34,35]. As shown in Figure 2, the cyclic structure of calix[n]arenes is similar to other polyhydroxylated macrocyles [36,37], and the macrocycle cavity will depend on the number of aromatic units in the system. These compounds have a three-dimensional cavity that can accommodate host molecules during a process called hostguest complexation. These systems have an advantage as a synthetic receptor, owing to different conformational isomeric forms, which allow different uses and applications. For instance, the calix [4]arenes can adopt several different conformers, including the cone, partial cone, 1,2-alternate, and 1,3-alternate. In the rigid cone conformation, all the phenolic -OH groups form strong hydrogen bonds that stabilize the structure (Figure 2). The calixarenes are a family of macrocyclic compounds with a variable number of phenol units linked by methylene bridges in ortho position [33]. The number of units of aromatics can be between 4 and 20, although the calixarenes of 4, 5, 6, 7, and 8 are the most common [34,35]. As shown in Figure 2, the cyclic structure of calix[n]arenes is similar to other polyhydroxylated macrocyles [36,37], and the macrocycle cavity will depend on the number of aromatic units in the system. These compounds have a three-dimensional cavity that can accommodate host molecules during a process called host-guest complexation. These systems have an advantage as a synthetic receptor, owing to different conformational isomeric forms, which allow different uses and applications. For instance, the calix [4]arenes can adopt several different conformers, including the cone, partial cone, 1,2-alternate, and 1,3-alternate. In the rigid cone conformation, all the phenolic -OH groups form strong hydrogen bonds that stabilize the structure (Figure 2 In the cavity of calix [4]arene in the cone conformation, it is possible to distinguish two edges: the lower rim, where the methylene bridges are, and the upper rim on the opposite side. In the center is the annular system, where the aromatic rings are; in addition, the upper rim is larger than the lower rim ( Figure 3) [38]. The molecular dimensions of the cavities vary depending on the number of units of aromatics; thus, the diameter of the upper rim has been estimated to be 3.8 Å for calix [4]arenes and 5.0 Å for calix [6]arenes [39]. In the cavity of calix [4]arene in the cone conformation, it is possible to distinguish two edges: the lower rim, where the methylene bridges are, and the upper rim on the opposite side. In the center is the annular system, where the aromatic rings are; in addition, the upper rim is larger than the lower rim ( Figure 3) [38]. The molecular dimensions of the cavities vary depending on the number of units of aromatics; thus, the diameter of the upper rim has been estimated to be 3.8 Å for calix [4]arenes and 5.0 Å for calix [6]arenes [39]. The calix [4]arenes can be functionalized on the upper rim or on the lower rim with several functional groups such as amides, imines, sulfur, azo, semicarbazone, and alkyl groups, among others, producing a wide variety of macrocycle compounds with different properties of recognition, selectivity, solubility, and degree of hydrophobicity [40]. This last aspect is very important, because with the introduction of polar groups to the calixarenes it is possible to design amphiphilic The calix [4]arenes can be functionalized on the upper rim or on the lower rim with several functional groups such as amides, imines, sulfur, azo, semicarbazone, and alkyl groups, among others, producing a wide variety of macrocycle compounds with different properties of recognition, selectivity, solubility, and degree of hydrophobicity [40]. This last aspect is very important, because with the introduction of polar groups to the calixarenes it is possible to design amphiphilic macrocycles with a hydrophilic head and a hydrophobic tail, which may self-assemble into micelles, vesicles, liposomes, and other aggregates useful in the transport of drugs [19]. Functionalization of calixarenes can be done from the starting materials; varying the nature of the substituent group on the phenol facilitates the modification on the upper rim of the macrocyclic system. On the lower rim, the obvious site for chemical modification is that of the hydroxyl groups. The functionalization of calixarenes with polar groups can lead to various structures, as shown in Figure 4. The reactivity of the calixarenes is mainly located at two points: on the hydroxyl groups (lower rim) or on the position of the hydroxyl group (upper rim). Functioning substances such as carboxilates [41], phosphates [42], ammonium groups [43] or sulfonates [44] can be introduced into the hydroxylated platform by means of easily accessible reactions and selectively, with good yields. The importance of the calix[n]arenes lies in the fact that they are able to recognize and accommodate into their cavity guest molecules via non-covalent interactions. Some of these interactions are H-bonding, cation-π, π-π stacking, van der Waals interactions, and the so-called anion-π, in cases where the aromatic system is electron-deficient [45,46]. The magnitude of the interaction also depends on the conformation of the macrocycle [47].
In conclusion, the advantages of the calix [4]arenes are that they are easy to synthetize and that they can be modified to obtain compounds according to the guest that one wants to complex. This The importance of the calix[n]arenes lies in the fact that they are able to recognize and accommodate into their cavity guest molecules via non-covalent interactions. Some of these interactions are H-bonding, cation-π, π-π stacking, van der Waals interactions, and the so-called anion-π, in cases where the aromatic system is electron-deficient [45,46]. The magnitude of the interaction also depends on the conformation of the macrocycle [47].
In conclusion, the advantages of the calix [4]arenes are that they are easy to synthetize and that they can be modified to obtain compounds according to the guest that one wants to complex. This synthesis has good yields, and the reagents are inexpensive [48]. In addition, the simple derivatives of calixarenes have not exhibited toxicity or immunogenic properties [19]. All of this makes calixarenes highly valued in supramolecular chemistry as complex agents for the transport of drugs and their controlled release, among other applications.
Complexing Properties of Calixarenes
As indicated, the cavity of the calixarenes is of variable size and is sufficiently large to accommodate anions, cations, or neutral molecules, including biologically important molecules. This, together with their ability to trap guest molecules by means of noncovalent interactions, has resulted in widespread interest in the calix[n]arenes within supramolecular chemistry, particularly with respect to the host-guest phenomena ( Figure 5) [49,50]. Furthermore, as mentioned above, the calix[n]arenes can be functionalized at both rims, R and -OH ( Figure 3) [51][52][53], producing a wide variety of compounds that can have different complexing properties. In particular, calix[n]arenes functionalized with a sulfonic group in the para position have been found to have several pharmaceutical applications, due to their potential for encapsulating drugs, increasing their solubility, bioavailability, oral absorption, and stability under heat, light, and acidic conditions [54][55][56]. In addition, they possess good biocompatibility and innocuousness [56,57]. Other applications of the calix[n]arenes are as phase-transfer agents, sensors, antibacterials, ionselective electrodes, and use as catalysts and in separations science [58,59].
Other advantages of the calix[n]arenes over other macrocyclic systems include a preorganized cavity of variable size and an electron-rich option of modification and selective binding with the guest [58,59]. This increases their potential use as agents of specific recognition, for example of a guest of toxicological interest.
Amphiphilic Assemblies Based on Calixarenes
As was discussed above, calixarenes are substances of easy access and of easy chemical modification, thus allowing one to obtain systems with the introduction of polar groups to the calixarenes, making it possible to design amphiphilic macrocycles with a hydrophilic head and a hydrophobic tail, which may self-assemble into micelles, vesicles, liposomes, nanocapsules, and other aggregates. The development of self-organizing synthetic amphiphilic calixarenes with properties of inclusion and encapsulation allows complex hydrophobic molecules to be transported in a hydrophilic environment; this is especially useful in the transport of drugs ( Figure 6). Figure 5. Three-dimensional representation of the complexation process of calix [4]arenes.
+
Furthermore, as mentioned above, the calix[n]arenes can be functionalized at both rims, R and -OH ( Figure 3) [51][52][53], producing a wide variety of compounds that can have different complexing properties. In particular, calix[n]arenes functionalized with a sulfonic group in the para position have been found to have several pharmaceutical applications, due to their potential for encapsulating drugs, increasing their solubility, bioavailability, oral absorption, and stability under heat, light, and acidic conditions [54][55][56]. In addition, they possess good biocompatibility and innocuousness [56,57]. Other applications of the calix[n]arenes are as phase-transfer agents, sensors, antibacterials, ion-selective electrodes, and use as catalysts and in separations science [58,59].
Other advantages of the calix[n]arenes over other macrocyclic systems include a preorganized cavity of variable size and an electron-rich option of modification and selective binding with the guest [58,59]. This increases their potential use as agents of specific recognition, for example of a guest of toxicological interest.
Amphiphilic Assemblies Based on Calixarenes
As was discussed above, calixarenes are substances of easy access and of easy chemical modification, thus allowing one to obtain systems with the introduction of polar groups to the calixarenes, making it possible to design amphiphilic macrocycles with a hydrophilic head and a hydrophobic tail, which may self-assemble into micelles, vesicles, liposomes, nanocapsules, and other aggregates. The development of self-organizing synthetic amphiphilic calixarenes with properties of inclusion and encapsulation allows complex hydrophobic molecules to be transported in a hydrophilic environment; this is especially useful in the transport of drugs ( Figure 6). Figure 6. Possibilities of amphiphilic calixarenes in self-assembly.
Types of Molecules that Can Be Recognized by Calixarenes and Their Effect
As mentioned above, the calixarenes have potential applications in diverse fields of medicine and in the pharmaceutical and biological contexts. They are very versatile molecules, because they have a hydrophobic character, but when aqueous environments are needed, for example in biological media, the calixarenes can be modified by the introduction of hydrophilic groups at the upper and lower rim of the macrocycle, leading to water-soluble structures. In this section, we give a brief description of both hydrophilic and hydrophobic molecules that can be recognized or transported by calixarenes.
Recognition of Dopamine by Calixarenes
Among the many substances that can be recognized by calixarenes is dopamine (Figure 7), which is a neurotransmitter of vital importance for the normal function of the central nervous system. The lack of this substance is one of the causes of Parkinson's disease, an illness that affects more than 10 million people worldwide [60]. In addition, alterations in the normal levels of dopamine are associated with attention-deficit hyperactivity disorder, which affects between 2 and 7% of children, adolescents, and adults in the world [61].
Types of Molecules that Can Be Recognized by Calixarenes and Their Effect
As mentioned above, the calixarenes have potential applications in diverse fields of medicine and in the pharmaceutical and biological contexts. They are very versatile molecules, because they have a hydrophobic character, but when aqueous environments are needed, for example in biological media, the calixarenes can be modified by the introduction of hydrophilic groups at the upper and lower rim of the macrocycle, leading to water-soluble structures. In this section, we give a brief description of both hydrophilic and hydrophobic molecules that can be recognized or transported by calixarenes.
Recognition of Dopamine by Calixarenes
Among the many substances that can be recognized by calixarenes is dopamine (Figure 7), which is a neurotransmitter of vital importance for the normal function of the central nervous system. The lack of this substance is one of the causes of Parkinson's disease, an illness that affects more than 10 million people worldwide [60]. In addition, alterations in the normal levels of dopamine are associated with attention-deficit hyperactivity disorder, which affects between 2 and 7% of children, adolescents, and adults in the world [61]. Dopamine can be recognized by 4-tert-butylcalix [6]arene when it is co-spread with cellulose acetate to form a Langmuir film and transferred to a gold electrode. The cellulose acetate maintaining the calixarene remains in the vertical cone conformation in the air-water interface, while the presence of calixarene in the Langmuir film provides selectivity to a sensor. The lineal range where the dopamine can be detected is between 5 and 100, and 100 and 7500 nm, and the limit of detection is 2.54 nM. The ratio of calixarene/cellulose acetate has been optimized to show that when the percent of calixarene is 30 wt %, the reuptake of the dopamine is greatest [62].
Improved Solubility of Testosterone with Calixarene
The interaction of calixarenes with steroids has been studied by several authors. In particular, a study of the complexation of testosterone with a water-soluble calixarene has been published by Millership [63]. Testosterone ( Figure 8) is a steroid hormone involved in the male sexual response. It is recognized that erectile dysfunction can result from low levels of testosterone, and therefore the assessment of serum testosterone levels before establishing a possible treatment is important (Figure 2) [64]. Because testosterone is not soluble in water, a method that allows its solubilization is desirable. In this way, Millership improved the solubility of this steroid in water by means of the complexation of the testosterone with 4-sulphonic calix[n]arenes. The author measured the solubility in the presence and absence of several calixarenes, such as 4-sulphonic calix [4]arenes, 4-sulphonic calix [6]arenes, and 4-sulphonic calix [8]arenes below the critical micelle concentration, and in all cases found the formation of soluble complexes 1:1, with 4-sulphonic calix [8]arene being the most soluble. This was explained by the fact that the steroid does not enter into the cavity of 4-sulphonic calix [4]arenes, due to their small size, while it can enter comfortably into the cavity of 4-sulphonic calix [6]arenes and 4-sulphonic calix [8]arene. He also found that the pH of the solution influences the shape of the calixarene because of the effect of the hydrogen bonding of the hydroxyl groups. Experiments conducted in phosphate buffers at pH 5.8, 7.3, and 10 demonstrate that the greatest solubility is reached at pH 7.3, which suggests that the conformations adopted by calixarene at this pH favor the process of complexation. The complex constants at pH 7.3 for complexes testosterone-4-sulphonic calix [4]arenes, testosterone-4-sulphonic calix [6]arenes, and testosterone-4-sulphonic calix [8]arenes were 26± 22, 346 ± 39, and 156 ± 9, respectively [65].
Biomaterial Modification with Calixarenes to Avoid Allergy or Infection
There are several implantable medical devices useful for the repair of soft and hard tissue, but many of these produce acute inflammation. Charnley et al. [65] proposed coating these medical Dopamine can be recognized by 4-tert-butylcalix [6]arene when it is co-spread with cellulose acetate to form a Langmuir film and transferred to a gold electrode. The cellulose acetate maintaining the calixarene remains in the vertical cone conformation in the air-water interface, while the presence of calixarene in the Langmuir film provides selectivity to a sensor. The lineal range where the dopamine can be detected is between 5 and 100, and 100 and 7500 nm, and the limit of detection is 2.54 nM. The ratio of calixarene/cellulose acetate has been optimized to show that when the percent of calixarene is 30 wt %, the reuptake of the dopamine is greatest [62].
Improved Solubility of Testosterone with Calixarene
The interaction of calixarenes with steroids has been studied by several authors. In particular, a study of the complexation of testosterone with a water-soluble calixarene has been published by Millership [63]. Testosterone ( Figure 8) is a steroid hormone involved in the male sexual response. It is recognized that erectile dysfunction can result from low levels of testosterone, and therefore the assessment of serum testosterone levels before establishing a possible treatment is important (Figure 2) [64]. Dopamine can be recognized by 4-tert-butylcalix [6]arene when it is co-spread with cellulose acetate to form a Langmuir film and transferred to a gold electrode. The cellulose acetate maintaining the calixarene remains in the vertical cone conformation in the air-water interface, while the presence of calixarene in the Langmuir film provides selectivity to a sensor. The lineal range where the dopamine can be detected is between 5 and 100, and 100 and 7500 nm, and the limit of detection is 2.54 nM. The ratio of calixarene/cellulose acetate has been optimized to show that when the percent of calixarene is 30 wt %, the reuptake of the dopamine is greatest [62].
Improved Solubility of Testosterone with Calixarene
The interaction of calixarenes with steroids has been studied by several authors. In particular, a study of the complexation of testosterone with a water-soluble calixarene has been published by Millership [63]. Testosterone ( Figure 8) is a steroid hormone involved in the male sexual response. It is recognized that erectile dysfunction can result from low levels of testosterone, and therefore the assessment of serum testosterone levels before establishing a possible treatment is important (Figure 2) [64]. Because testosterone is not soluble in water, a method that allows its solubilization is desirable. In this way, Millership improved the solubility of this steroid in water by means of the complexation of the testosterone with 4-sulphonic calix[n]arenes. The author measured the solubility in the presence and absence of several calixarenes, such as 4-sulphonic calix [4]arenes, 4-sulphonic calix [6]arenes, and 4-sulphonic calix [8]arenes below the critical micelle concentration, and in all cases found the formation of soluble complexes 1:1, with 4-sulphonic calix [8]arene being the most soluble. This was explained by the fact that the steroid does not enter into the cavity of 4-sulphonic calix [4]arenes, due to their small size, while it can enter comfortably into the cavity of 4-sulphonic calix [6]arenes and 4-sulphonic calix [8]arene. He also found that the pH of the solution influences the shape of the calixarene because of the effect of the hydrogen bonding of the hydroxyl groups. Experiments conducted in phosphate buffers at pH 5.8, 7.3, and 10 demonstrate that the greatest solubility is reached at pH 7.3, which suggests that the conformations adopted by calixarene at this pH favor the process of complexation. The complex constants at pH 7.3 for complexes testosterone-4-sulphonic calix [4]arenes, testosterone-4-sulphonic calix [6]arenes, and testosterone-4-sulphonic calix [8]arenes were 26± 22, 346 ± 39, and 156 ± 9, respectively [65].
Biomaterial Modification with Calixarenes to Avoid Allergy or Infection
There are several implantable medical devices useful for the repair of soft and hard tissue, but many of these produce acute inflammation. Charnley et al. [65] proposed coating these medical Because testosterone is not soluble in water, a method that allows its solubilization is desirable. In this way, Millership improved the solubility of this steroid in water by means of the complexation of the testosterone with 4-sulphonic calix[n]arenes. The author measured the solubility in the presence and absence of several calixarenes, such as 4-sulphonic calix [4]arenes, 4-sulphonic calix [6]arenes, and 4-sulphonic calix [8]arenes below the critical micelle concentration, and in all cases found the formation of soluble complexes 1:1, with 4-sulphonic calix [8]arene being the most soluble. This was explained by the fact that the steroid does not enter into the cavity of 4-sulphonic calix [4]arenes, due to their small size, while it can enter comfortably into the cavity of 4-sulphonic calix [6]arenes and 4-sulphonic calix [8]arene. He also found that the pH of the solution influences the shape of the calixarene because of the effect of the hydrogen bonding of the hydroxyl groups. Experiments conducted in phosphate buffers at pH 5.8, 7.3, and 10 demonstrate that the greatest solubility is reached at pH 7.3, which suggests that the conformations adopted by calixarene at this pH favor the process of complexation. The complex constants at pH 7.3 for complexes testosterone-4-sulphonic calix [4]arenes, testosterone-4-sulphonic calix [6]arenes, and testosterone-4-sulphonic calix [8]arenes were 26± 22, 346 ± 39, and 156 ± 9, respectively [65].
Biomaterial Modification with Calixarenes to Avoid Allergy or Infection
There are several implantable medical devices useful for the repair of soft and hard tissue, but many of these produce acute inflammation. Charnley et al. [65] proposed coating these medical devices with a natural anti-inflammatory that consists of a hormone produced by the organism that does not generate allergy or infection but that can easily be synthesized in the laboratory. In this study, the hormone was α-melanocyte-stimulating (Figure 9), and was immobilized onto medical device surfaces with C-tetra(octyl)calixresorcin [4]arene. The results obtained by the authors indicated that the α-melanocyte hormone retains its anti-inflammatory properties and suggested that this strategy could be useful in the manufacture of new materials [66].
Biomolecules 2019, 9, x FOR PEER REVIEW 9 of 17 devices with a natural anti-inflammatory that consists of a hormone produced by the organism that does not generate allergy or infection but that can easily be synthesized in the laboratory. In this study, the hormone was α-melanocyte-stimulating (Figure 9), and was immobilized onto medical device surfaces with C-tetra(octyl)calixresorcin [4]arene. The results obtained by the authors indicated that the α-melanocyte hormone retains its anti-inflammatory properties and suggested that this strategy could be useful in the manufacture of new materials [66]. Figure 9. Chemical structure of α-melanocyte hormone.
Other applications similar to this one have been reported with proteins, which are important for the manufacture of synthetic organs, drug delivery systems, and biosensing, among other uses. The aim is to achieve an efficient interaction between the proteins and the solid materials for the production of nano-and biomaterials. In this context, the calixarenes have been shown to be useful as linking agents. For example, Keskinates et al. [66] used a calixarene tetraester (5,11,17,23-tetra-tertbutyl-25,26,27,28-tetramethoxycarbonylmethoxy-calix [4]arene) to achieve a good interaction between nanofibers of polyacrylonitrile (PAN) or poly(methyl methacrylate) (PMMA) and a green fluorescent protein. The containment of calixarene on the fiber where the binding of protein was highest was 50%. The usefulness of the modified fiber is as an adsorbent membrane for the removal of proteins in aqueous solution; therefore the authors suggest that PAN or PMMA fibers containing calixarene are a promising new and inexpensive material for protein purification [67].
Combination of Calixarenes and Cyclodextrins to Improve the Solubility of an Anthelmintic Drug
Niclosamide ( Figure 10) is an oral anthelmintic drug used worldwide to treat parasitic infections. It is active against beef and dog tapeworms. In addition, it is useful for treating diseases caused by these parasites, such as cancer, metabolic diseases, artery constriction, endometriosis, and rheumatoid arthritis, among others [68]. In spite of its broad clinical application, the efficacy of niclosamide can be affected by its very low solubility in water (230 ng/mL). Studies conducted by Yang et al. have shown a great increase in the aqueous solubility niclosamide by means of complexation with 4-sulphonate-calix [6]arene and hydroxypropyl-β-cyclodextrin. The advantage of the use of water-soluble calixarenes is that they provide a hydrophobic environment to include the drug and a hydrophilic head (sulphonate groups) that allow its solubilization. The combination of the two macrocycle types provides both the properties of cyclodextrins and those of micelles. Another advantage is that the p-sulphonate-calix [4, 6, and 8]arenes do not exhibit acute toxication when they are injected into mice, and exhibit no activity in the Ames test. In addition, the inner cavity diameter is variable, between 3.0, 7.6, and 11.7 Å for calix [4]arene, calix [6]arene, and calix [8]arene, respectively, and between 5.7, 7.8, and 9.5 Å for α, β, and γ-cyclodextrins. Furthermore, calixarenes are highly flexible molecules, while the cyclodextrins are quite rigid molecules. All of the above ensures good versatility of the binding of the two macrocycles with the anthelmintic drug. Other applications similar to this one have been reported with proteins, which are important for the manufacture of synthetic organs, drug delivery systems, and biosensing, among other uses. The aim is to achieve an efficient interaction between the proteins and the solid materials for the production of nano-and biomaterials. In this context, the calixarenes have been shown to be useful as linking agents. For example, Keskinates et al. [66] used a calixarene tetraester (5,11,17,23-tetra-tert-butyl-25,26,27,28-tetramethoxycarbonylmethoxy-calix[4]arene) to achieve a good interaction between nanofibers of polyacrylonitrile (PAN) or poly(methyl methacrylate) (PMMA) and a green fluorescent protein. The containment of calixarene on the fiber where the binding of protein was highest was 50%. The usefulness of the modified fiber is as an adsorbent membrane for the removal of proteins in aqueous solution; therefore the authors suggest that PAN or PMMA fibers containing calixarene are a promising new and inexpensive material for protein purification [67].
Combination of Calixarenes and Cyclodextrins to Improve the Solubility of an Anthelmintic Drug
Niclosamide ( Figure 10) is an oral anthelmintic drug used worldwide to treat parasitic infections. It is active against beef and dog tapeworms. In addition, it is useful for treating diseases caused by these parasites, such as cancer, metabolic diseases, artery constriction, endometriosis, and rheumatoid arthritis, among others [68]. In spite of its broad clinical application, the efficacy of niclosamide can be affected by its very low solubility in water (230 ng/mL). Studies conducted by Yang et al. have shown a great increase in the aqueous solubility niclosamide by means of complexation with 4-sulphonate-calix [6]arene and hydroxypropyl-β-cyclodextrin. The advantage of the use of water-soluble calixarenes is that they provide a hydrophobic environment to include the drug and a hydrophilic head (sulphonate groups) that allow its solubilization. The combination of the two macrocycle types provides both the properties of cyclodextrins and those of micelles. Another advantage is that the p-sulphonate-calix[4, 6, and 8]arenes do not exhibit acute toxication when they are injected into mice, and exhibit no activity in the Ames test. In addition, the inner cavity diameter is variable, between 3.0, 7.6, and 11.7 Å for calix [4]arene, calix [6]arene, and calix [8]arene, respectively, and between 5.7, 7.8, and 9.5 Å for α, β, and γ-cyclodextrins. Furthermore, calixarenes are highly flexible molecules, while the cyclodextrins are quite rigid molecules. All of the above ensures good versatility of the binding of the two macrocycles with the anthelmintic drug. The authors indicate that there is an increase of solubility of niclosamide with an increase in equal molar concentrations of calixarene and cyclodextrin, but when the concentrations reach 0.005 M, the solubility decreases, indicating the precipitation of an insoluble complex at higher concentrations. The process of complexation was explained by hydrogen bonding between the hydroxylic groups of calixarene and the oxygen atom and the nitrogen atom of niclosamide. Hydrophobic interactions also can be suggested between the hydrophobic cavity of calixarene and the hydrophobic molecule of niclosamide. On the other hand, the solubility of niclosamide also increases with the cyclodextrin, because these molecules establish hydrogen bonding with the hydroxyl groups on the exterior of the cyclodextrin. In general terms, the great increase of solubility of niclosamide is attributed to the additivity of solubility to 4-sulphonato-calix [6]arene and 2-hydroxypropyl-β-cyclodextrin [69].
Drug Delivery Systems Based on Calixarenes
Calixarenes can also be useful as nanocarriers in the form of inclusion complexes, micelles, hydrogels, vesicles, liposomes, and nanoparticles. For example, calixarenes have attracted attention in medicine for the treatment of cancer, because they can respond to multiple stimuli, are stable, can avoid nonspecific cell uptake, and are able to reach the target on tumor sites in order to effect the treatment. Calixarenes are ideal for applications in delivery systems, because they have shown good biocompatibility and non-cytotoxicity.
Cisplatin ( Figure 11) is an anticancer agent well known for more than The host-guest complex displays no cytotoxicity and is formed by interactions of hydrogen The authors indicate that there is an increase of solubility of niclosamide with an increase in equal molar concentrations of calixarene and cyclodextrin, but when the concentrations reach 0.005 M, the solubility decreases, indicating the precipitation of an insoluble complex at higher concentrations. The process of complexation was explained by hydrogen bonding between the hydroxylic groups of calixarene and the oxygen atom and the nitrogen atom of niclosamide. Hydrophobic interactions also can be suggested between the hydrophobic cavity of calixarene and the hydrophobic molecule of niclosamide. On the other hand, the solubility of niclosamide also increases with the cyclodextrin, because these molecules establish hydrogen bonding with the hydroxyl groups on the exterior of the cyclodextrin. In general terms, the great increase of solubility of niclosamide is attributed to the additivity of solubility to 4-sulphonato-calix [6]arene and 2-hydroxypropyl-β-cyclodextrin [69].
Drug Delivery Systems Based on Calixarenes
Calixarenes can also be useful as nanocarriers in the form of inclusion complexes, micelles, hydrogels, vesicles, liposomes, and nanoparticles. For example, calixarenes have attracted attention in medicine for the treatment of cancer, because they can respond to multiple stimuli, are stable, can avoid nonspecific cell uptake, and are able to reach the target on tumor sites in order to effect the treatment. Calixarenes are ideal for applications in delivery systems, because they have shown good biocompatibility and non-cytotoxicity.
Cisplatin ( Figure 11) is an anticancer agent well known for more than 40 years. There are several examples of formulations of oxaliplatin that use liposomes and micelles as a drug delivery vehicle. Abbott et al. published a study of the complexation of a dinuclear platinum complex with p-sulphonatocalix [4]arene. The host-guest ratio was 1:1, and the binding constant was 6.8 · 10 4 M −1 . The authors indicate that there is an increase of solubility of niclosamide with an increase in equal molar concentrations of calixarene and cyclodextrin, but when the concentrations reach 0.005 M, the solubility decreases, indicating the precipitation of an insoluble complex at higher concentrations. The process of complexation was explained by hydrogen bonding between the hydroxylic groups of calixarene and the oxygen atom and the nitrogen atom of niclosamide. Hydrophobic interactions also can be suggested between the hydrophobic cavity of calixarene and the hydrophobic molecule of niclosamide. On the other hand, the solubility of niclosamide also increases with the cyclodextrin, because these molecules establish hydrogen bonding with the hydroxyl groups on the exterior of the cyclodextrin. In general terms, the great increase of solubility of niclosamide is attributed to the additivity of solubility to 4-sulphonato-calix [6]arene and 2-hydroxypropyl-β-cyclodextrin [69].
Drug Delivery Systems Based on Calixarenes
Calixarenes can also be useful as nanocarriers in the form of inclusion complexes, micelles, hydrogels, vesicles, liposomes, and nanoparticles. For example, calixarenes have attracted attention in medicine for the treatment of cancer, because they can respond to multiple stimuli, are stable, can avoid nonspecific cell uptake, and are able to reach the target on tumor sites in order to effect the treatment. Calixarenes are ideal for applications in delivery systems, because they have shown good biocompatibility and non-cytotoxicity.
Cisplatin ( Figure 11) is an anticancer agent well known for more than The host-guest complex displays no cytotoxicity and is formed by interactions of hydrogen bonding between the NH3 groups of the metal complex with sulphate groups of the calix [4]arene. The host-guest complex displays no cytotoxicity and is formed by interactions of hydrogen bonding between the NH 3 groups of the metal complex with sulphate groups of the calix [4]arene.
Controlled Release of Doxorubicin by Vesicles Based on Calixarenes
Because calixarenes are able to self-assemble, their use for the construction of vesicles is possible. Wang et al. report the construction of binary supramolecular vesicles driven by host-guest complexation between a water-soluble calixarene and viologen. The vesicle obtained can be used as a system of controlled release of doxorubicin hydrochlodire (DOX) (Figure 12), a fluorescent dye that is used in the treatment of cancer. Cell experiments show that the release of DOX by the vesicle does not affect the therapeutic effect of the drug against the cancer cell and by contrast reduces damage to normal cells. The vesicles are stable for 100 h (approximately 4 days) at room temperature. The formation of vesicles was followed by ultraviolet-visible (UV-Vis). In the absence of calixarene, the maximum absorption of the viologen is 260 nm, and the aggregate is not formed, but in the presence of the macrocycle, aggregation takes place through the formation of a complex, and the maximum absorption changes to 450 nm. The stability of the vesicle is reinforced by the electrostatic interaction of the negative groups of sulfonate groups of resorcinarene and the positive groups of viologen.
Biomolecules 2019, 9, x FOR PEER REVIEW 11 of 17 Because calixarenes are able to self-assemble, their use for the construction of vesicles is possible. Wang et al. report the construction of binary supramolecular vesicles driven by host-guest complexation between a water-soluble calixarene and viologen. The vesicle obtained can be used as a system of controlled release of doxorubicin hydrochlodire (DOX) (Figure 12), a fluorescent dye that is used in the treatment of cancer. Cell experiments show that the release of DOX by the vesicle does not affect the therapeutic effect of the drug against the cancer cell and by contrast reduces damage to normal cells. The vesicles are stable for 100 h (approximately 4 days) at room temperature. The formation of vesicles was followed by ultraviolet-visible (UV-Vis). In the absence of calixarene, the maximum absorption of the viologen is 260 nm, and the aggregate is not formed, but in the presence of the macrocycle, aggregation takes place through the formation of a complex, and the maximum absorption changes to 450 nm. The stability of the vesicle is reinforced by the electrostatic interaction of the negative groups of sulfonate groups of resorcinarene and the positive groups of viologen. The obtained vesicles can respond to external stimuli such as temperature, host-guest inclusion, and redox. In the first case, at temperatures between 5 and 70 °C the processes of assembly and disassemblyare produced, respectively. The increasing temperature produced the gradual disassembly of the vesicle. In the second case, taking advantage of the fact that the cyclodextrins can form complexes with viologen, the gradual addition of α-cyclodextrins was carried out, producing the disruption of the vesicle. In the last case, it is known that viologens can be transformed from neutral molecules into the corresponding radical cations by chemical or electrochemical means. The chemical reduction is carried out with hydrazine, where the solution color changes to purple; however, this does not produce the disruption of the vesicle, but the average diameter decreases from 308 to 153 nm. A similar effect is achieved with electrical redox, but when a reduction potential of −1.6 V vs. Ag/AgCl is applied for 30 min, disassembly of the supramolecular vesicle occurs [70].
The above discussion shows that the calixarenes are able to recognize a great diversity of chemical species, depending on the cavity size and the shape and functional groups of the macrocycle. This property is of great importance, because it allows improved bioavailability, solubility, or even activity of molecules of biological interest. In addition, the calixarenes are able to self-assemble to form micelles, vesicles, liposomes, and nanoparticles that allow the encapsulation, transport, and controlled release of drugs. Table 1 shows an overview of the aforementioned applications, where the complexation with calixarene improved the properties of the biomolecule. Also included are methods, conditions, and bibliographic references. The obtained vesicles can respond to external stimuli such as temperature, host-guest inclusion, and redox. In the first case, at temperatures between 5 and 70 • C the processes of assembly and disassemblyare produced, respectively. The increasing temperature produced the gradual disassembly of the vesicle. In the second case, taking advantage of the fact that the cyclodextrins can form complexes with viologen, the gradual addition of α-cyclodextrins was carried out, producing the disruption of the vesicle. In the last case, it is known that viologens can be transformed from neutral molecules into the corresponding radical cations by chemical or electrochemical means. The chemical reduction is carried out with hydrazine, where the solution color changes to purple; however, this does not produce the disruption of the vesicle, but the average diameter decreases from 308 to 153 nm. A similar effect is achieved with electrical redox, but when a reduction potential of −1.6 V vs. Ag/AgCl is applied for 30 min, disassembly of the supramolecular vesicle occurs [70].
The above discussion shows that the calixarenes are able to recognize a great diversity of chemical species, depending on the cavity size and the shape and functional groups of the macrocycle. This property is of great importance, because it allows improved bioavailability, solubility, or even activity of molecules of biological interest. In addition, the calixarenes are able to self-assemble to form micelles, vesicles, liposomes, and nanoparticles that allow the encapsulation, transport, and controlled release of drugs. Table 1 shows an overview of the aforementioned applications, where the complexation with calixarene improved the properties of the biomolecule. Also included are methods, conditions, and bibliographic references. The fiber modified was characterized by FTIR, TGA analysis, UV-Vis, fluorescence microscopy and SEM
•
The bindings studied were done at room temperature • The protein content was analysed via UV-Vis at 476 nm • The synthesis of calix [4]arene tetra ester derivative was carried out on diphenyl ether with formaldehyde and a basic medium [67] Niclosamide (anthelmintic drug) 4-sulphonatocalix [6]arene The complexation between calixarene-cyclodextrin and niclosamide was followed by thermal analysis. The niclosamide content was determined by HPLC Solubility studies were done in a pH 7.0 buffer at 30 • C and ionic strength of 0.5 mol/L [67] Dinuclear platinum complex p-sulphonatocalix [4]arene The complex was examined using 1 H nuclear magnetic resonance and electrospray ionization mass spectrometry, among others • The complex was formed by an equimolar mixture of dinuclear platinum complex with p-sulphonatocalix [4]arene • Its water solubility is around 4.5 mM [71] Doxorubicin p-sulphonatocalix [4]arene The nanosupramolecular binary vesicles was studied by UV-Vis, fluorescence spectroscopy, dynamic laser scattering, transmission electron microscopy, scanning electron microscopy, and surface tension
Conclusions
A wide range of substances of biological interest can be recognized by calix[n]arenes. The recognition process depends on the size, shape, polarity, type of functional groups present, aggregation state of the macrocycle, and the formation of non-covalent interactions with the guest. The design of calix[n]arenes allows for obtaining amphiphilic molecules able to form higher-order structures such as micelles, vesicles, liposomes, and nanoparticles, which have been shown to be suitable for controlled release drug delivery systems. In addition, calix [4]aneres are useful for improving the solubility, biocompatibility, stability, and bioavailability of biomolecules and molecules with biological activity, and due to the fact that they are not cytotoxic, they can be used in the manufacture of biomaterials. All the above shows the great potential of these macrocycles in pharmacology, biomaterials engineering, and medicine, among other fields. | 9,524.2 | 2019-03-01T00:00:00.000 | [
"Chemistry",
"Medicine"
] |
Defects in the Outer Limiting Membrane Are Associated with Rosette Development in the Nrl−/− Retina
The neural retinal leucine zipper (Nrl) knockout mouse is a widely used model to study cone photoreceptor development, physiology, and molecular biology in the absence of rods. In the Nrl−/− retina, rods are converted into functional cone-like cells. The Nrl−/− retina is characterized by large undulations of the outer nuclear layer (ONL) commonly known as rosettes. Here we explore the mechanism of rosette development in the Nrl−/− retina. We report that rosettes first appear at postnatal day (P)8, and that the structure of nascent rosettes is morphologically distinct from what is seen in the adult retina. The lumen of these nascent rosettes contains a population of aberrant cells protruding into the subretinal space that induce infolding of the ONL. Morphologically adult rosettes do not contain any cell bodies and are first detected at P15. The cells found in nascent rosettes are photoreceptors in origin but lack inner and outer segments. We show that the adherens junctions between photoreceptors and Müller glia which comprise the retinal outer limiting membrane (OLM) are not uniformly formed in the Nrl−/− retina and thus allow protrusion of a population of developing photoreceptors into the subretinal space where their maturation becomes delayed. These data suggest that the rosettes of the Nrl−/− retina arise due to defects in the OLM and delayed maturation of a subset of photoreceptors, and that rods may play an important role in the proper formation of the OLM.
Introduction
Development of the mammalian photoreceptor cell layer is characterized by the sequential restriction of competence of retinal progenitor cells to either rod or cone cells [1]. Cone precursor formation begins around embryonic day (E) 11 and continues to just after birth, while rod precursor formation begins around E12 and continues until ,postnatal day (P) 8 [2]. Following terminal differentiation, all photoreceptor precursors express the transcription factors cone-rod homeobox (CRX) and orthodenticle homeobox 2 (OTX2) that are critical for establishing a photoreceptor cell fate [3,4,5,6]. Photoreceptor precursors which express a threshold level of active neural retinal leucine zipper (NRL) adopt a rod cell fate, reinforced and maintained by the NRL target gene nuclear receptor subfamily 2, group E, member 3 (Nr2e3), while those below this threshold level adopt a cone cell fate [7,8,9].
The Nrl knockout mouse is a widely used model to study cone cell development and physiology in the absence of rods. In the wild-type (WT) adult mouse, rods outnumber cones by approximately 30:1 [1]. Without NRL, rod cells all differentiate into functional cone-like cells which express cone-specific proteins [7]. Since NRL acts in post-mitotic cells, the total number of photoreceptors does not change in the Nrl 2/2 retina compared to the WT retina [1,7]. Outer segments (OSs) in the Nrl 2/2 retina have flattened, stacked lamellae similar to WT cone OSs, but are shorter and have some abnormalities [7,10]. Interestingly, the Nrl 2/2 retina exhibits undulations of the outer nuclear layer (ONL), commonly referred to as rosettes [7]. These rosettes form by an unknown mechanism and are prominent in the adult retina.
One of the known causes of rosette formation in other models is a defect in the retinal outer limiting membrane (OLM) [11]. The OLM is a series of adheren junctions between photoreceptors and neighboring Müller glia that serves, among other functions, to determine apical-basal polarity in photoreceptor cells [12]. Proper formation of the OLM is critical for photoreceptor maturation, function, and vision. Mutations of Crb1, an important component of the OLM, can cause severe blinding diseases in humans including Leber congenital amaurosis and retinitis pigmentosa [13,14]. Importantly, rosette formation has been observed in instances of both genetic and chemical disruption of the OLM [15,16]. Here we present results demonstrating that rosette formation in the Nrl 2/2 retina is associated with defects in the formation of the OLM and delayed maturation of some photoreceptors. Given that the primary defect in the Nrl 2/2 retina is a lack of rods, these data suggest that rods are required for the proper formation of the OLM and proper organization of the ONL.
Animals
All experiments and animal maintenance followed protocols approved by the University of Oklahoma's Institutional Animal Care and Use Committee (IACUC), protocol number 10-090, and the guidelines set forth by the Association for Research in Vision and Ophthalmology (ARVO). The Nrl 2/2 mice were bred from founders generously provided by Dr. Anand Swaroop (National Eye Institute, Bethesda, MD). The Rds 2/2 mice were bred from founders generously provided by Dr. Neeraj Agarwal (University of North Texas Health Science Center, Fort Worth, TX). Animals of both genders were used and all animals were reared under cyclic lighting conditions (12 h L/D).
Tissue Collection
After euthanasia, eyes were enucleated, the cornea was punctured, and eyes were placed in 4% paraformaldhyde in PBS for 2 hours. All eyes were collected between 11 am and 2 pm. After 2 hours, the cornea and lens of each eye was removed and the eye cups were returned to fixative for 2 additional hours. The eyes were cryoprotected in a sequential sucrose gradient and cryosectioned as described previously [17]. Briefly, eyes were embedded in frozen Shandon M-1 embedding matrix (Thermo Electron Corporation) and 10 mm or 25 mm sections were collected (Leica Cryostat CM3050-S).
Immunohistochemistry
For immunohistochemistry (IHC), slides were treated with 100% methanol for 20 minutes at 220uC followed by a 2 minute immersion in a 1% solution of NaBH 4 at room temperature. Slides were incubated in blocking solution for 2 hours (2% normal goat serum, 5% BSA, 1% fish gelatin, 0.5% triton X-100, in either HBSS or PBS). Primary antibodies were diluted in blocking solution and sections were incubated overnight at 4uC in a dark humid chamber. After washing, secondary antibodies were diluted in blocking solution and incubated with sections for one hour at room temperature. Slides were mounted using ProLongH Gold antifade reagent with DAPI (Invitrogen, Carlsbad, CA) prior to imaging. Sections were imaged using a Hamamatsu C-4742 camera through UPlanSApo objectives (Olympus, Tokyo, Japan) on an Olympus BX62 upright microscope equipped with a spinning disk confocal unit. Image analysis was performed using Slidebook v4 software (Olympus), and ImageJ [18].
Rosette Quantification
Rosettes were quantified in 4-5 mice per genotype/age. Nascent and mature rosettes were counted across three full 10 mm thick central retinal sections (containing the optic nerve head) for each mouse and averaged to give a value for that individual animal. An individual rosette was scored as nascent if it contained a population of cells lacking subretinal IS/OS staining and mature if it did not contain any cells. Results were analyzed using two-way ANOVA for genotype and age.
TUNEL
TUNEL staining was accomplished using the In Situ Cell Death Detection Kit, TMR red (Roche, Indianapolis, IN) according to the manufacturer's specifications with the following modifications. Prior to using the kit protocol, the slides were incubated in 100% methanol for 20 minutes at 220uC. Following the completion of the TUNEL protocol, the slides were subjected to the IHC protocol described above for co-labeling experiments.
BrdU
Mice underwent three intraperitoneal injections with a 10 mg/ml solution of BrdU (Sigma-Aldrich, St. Louis, MO) dissolved in PBS at a final dose of 50 mg/g body weight. Injections occurred at 4 pm, 12 am and at 8 am and tissue was collected 24 hours after the initial injection. Eyes were collected and processed as described above. Slides underwent IHC as described above but with the initial methanol incubation replaced by a 20 minute incubation in 10 mM citrate buffer (pH 6.0) at 95uC.
Electron Microscopy
Transmission electron microscopy was performed as previously described [25,26]. Briefly, eyecups were embedded in plastic and thin sections (600-800 Å ) were collected, stained, and imaged using a JEOL 100CX electron microscope at an accelerating voltage of 60 kV.
Nascent rosette morphology is distinct from adult rosette morphology
To begin our studies of the mechanism underlying rosette formation in the Nrl 2/2 retina, we established the time course of rosette development in this model. To visualize rosettes, we stained frozen sections of isolated eyes at the ages shown in ( Figure 1A) using peanut agglutinin (PNA), a marker for the cone extracellular matrix, which decorates the area surrounding developing and mature cone ISs and OSs. At P1, PNA labeling is detected throughout the developing retina but quickly resolves into two discrete lines as the nuclear layers separate by P8 ( Figure 1A). From P8 onwards, bright staining is detected in the developing IS/ OS layer with fainter labeling detected around the developing cone terminals in the outer plexiform layer (OPL). At P8 in the WT retina, PNA labeling is detected in a straight line across the IS/OS layer, while in the Nrl 2/2 retina, breaks and infolding in the PNA labeled layer begin to be detected. These breaks highlight the initial formation of the rosettes ( Figure 1A arrowheads). Initial rosette formation is observed at P8 in the Nrl 2/2 and these early (nascent) rosettes (arrowheads) have a different morphology than the mature rosettes seen in the adult retina (arrows).
To further study this morphological difference, we undertook higher magnification confocal imaging of retinal sections labeled with the cone IS/OS marker PNA (green) and the cone OS marker S-opsin (red). Mature rosettes appear as undulations of the ONL filled with the OSs of the surrounding photoreceptors ( Figure 1C arrows). In contrast, the nascent rosettes in the developing retina contain not only early IS/OSs, but also a population of aberrant cells ( Figure 1B arrowheads). In contrast to the photoreceptor nuclei outside rosettes which direct IS/OS into the subretinal space ( Figure 1B, C asterisks), the aberrant cells inside the nascent rosettes lack IS/OS markers in the subretinal space ( Figure 1B, arrowheads) suggesting that they either do not have OSs or that their OSs are misoriented. Some aberrant cells in nascent rosettes appear to be completely isolated from the surrounding ONL ( Figure 1B, right) while in others, the aberrant cells appear to be contiguous with the ONL ( Figure 1B, left). As the rosettes grow in size from P8 to P12, they appear to contain more aberrant cells ( Figure 1A). Beginning between P12 and P15 however, the aberrant cells begin to disappear and the nascent rosettes are converted into mature rosettes which do not exhibit cell bodies in their lumen ( Figure 1C, arrows). We observe that rosettes mature in a central to peripheral gradient consistent with the previously reported pattern of retinal development [27].
To understand further the timecourse of the transition from a retina with nascent rosettes to a retina with mature rosettes, we counted the number of rosettes in the Nrl 2/2 retina exhibiting nascent morphology (defined as those rosettes which contain aberrant cells lacking subretinal IS/OS staining) and the number of rosettes exhibiting mature morphology (defined as those which do not contain aberrant cells) in 10 mm thick central retinal cross sections containing the optic nerve. The number of nascent rosettes peaks at P10-12 after which it decreases significantly ( Figure 1D, black lines). The decrease in nascent rosettes is accompanied by a significant, dramatic increase in the number of mature rosettes observed between P12 and P15 ( Figure 1D, gray lines) consistent with maturation of rosettes and conversion from the nascent to the adult morphology during this time. Rosette formation is almost finished by P15; the number of mature rosettes observed in P15 retinas is not significantly different from the number observed in P30 retinas. The conversion of nascent nucleicontaining rosettes to mature ''empty'' rosettes occurs over time; at P15 many rosettes exhibit mature morphology while others still contain some aberrant cells, but by P30, no rosettes with nascent morphology are detected. Interestingly, the maximum number of mature rosettes (observed at P30) is approximately 3 fold higher than the maximum number of nascent rosettes (observed at P10), suggesting that each nascent rosette matures into multiple adult rosettes.
In the WT mouse, the majority of the subretinal space is filled with rod OSs with only a relatively small number of cone OSs. Since cone OSs are shorter and differently shaped than rods OSs [28], and in the Nrl 2/2 retina, all photoreceptors which would usually have a rod OS structure instead adopt a cone like structure, we wanted to rule out the possibility that the altered shape of the OSs (i.e. all cones) contributed to rosette development. To address this, we analyzed the timecourse of rosette development in an Nrl 2/2 retina which exhibits abnormal cone OS morphology. We chose the previously characterized Rds 2/2 Nrl 2/2 mouse model since it exhibits OSs that are functional but structurally deficient. The OSs of the Rds 2/2 Nrl 2/2 retina are simply open membranous sacs; they have no lamellae and no defined shape [29]. We did not detect any statistically significant differences in the number of rosettes or the timing of development between the Nrl 2/2 retina and the Rds 2/2 Nrl 2/2 retina ( Figure S1), suggesting that rosette morphology and development in the Nrl 2/2 retina is independent of OS structure.
To gain a clearer understanding of the initial morphological characteristics of the nascent rosettes, we made three-dimensional reconstructions using a multi-plane confocal image taken at P8 and stained with PNA and antibodies against S-opsin. Figure 2 shows two different representative rosettes in the left and right columns; panels A and B show two different angles of each rosette (refer to axes), one with DAPI and one without, while panel C shows two representative individual planes (out of ,20) for each of the two rosettes. At P8, nascent rosettes are small enough that the structure can be reconstructed in a 25 mm image stack. In all cases we observed, the aberrant cells protrude out of the ONL, forming a mushroom-shaped clump of cells that is contiguous with the ONL (Figure 2A, arrowheads). The aberrant cells project through a gap in normal IS/OS staining, easily visualized when the DAPI channel is turned off ( Figure 2B). The protruding cells spread out over the IS/OS labeling in the subretinal space. As a result of this spreading, in some planes, the aberrant cells appear to be fully enclosed by the IS/OSs lining the rosette ( Figure 2C, plane 11-right and Figure 1B, right) while in other sections the DAPI-labeled nuclei of the aberrant cells are clearly contiguous with the DAPI-labeled nuclei of the ONL ( Figure 2C, plane 3right). Our observations suggest that all aberrant cells are contiguous with the ONL even in the larger nascent rosettes seen between P8 and P15, although this connection is not always observable in individual image planes.
The aberrant cells are not due to abnormal late proliferation of photoreceptor cells
While previous work has shown that the adult Nrl 2/2 retina has a normal number of photoreceptor nuclei preceding subsequent degeneration [7], and cell birth in the retina is completed earlier than the onset of rosettes, the appearance of the aberrant cells at P8 suggested that there could be abnormal late cell proliferation occurring which is followed by a wave of cell death as the aberrant cells disappear and the rosettes mature. To determine whether the aberrant cells result from abnormal proliferation, Nrl 2/2 pups underwent a series of 3 intraperitoneal BrdU injections at 8 hour intervals beginning one day prior to the collection timepoint. Twenty-four hours after the initial injection, eyes were harvested co-labeled with S-opsin (to highlight rosettes-green) and anti-BrdU (red). Lack of BrdU labeling indicates that cells inside nascent rosettes are not actively dividing at either P8 or P10 ( Figure S2, top). Similar results were obtained when the experiment was conducted at P12 and P15 (not shown). TUNEL assay was used to test whether the aberrant cells were being cleared through apoptosis as the rosettes matured ( Figure S3). Apoptotic cells that are TUNEL positive (red) are observed within the INL of Nrl 2/2 and WT retinas until approximately P12 to P15 ( Figure S3, arrows) consistent with normal retinal development, but no noticeable increase in apoptosis was observed inside the nascent rosettes (labeled with S-opsin-green) vs. outside at any time point ( Figure S3, arrowheads).
Rosette associated aberrant cells are underdeveloped photoreceptors
Given that the population of cells found inside nascent rosettes is contiguous with the ONL and the appearance of their nuclei, we hypothesized that these cells were simply misplaced photoreceptor cells, despite their lack of IS/OS staining, rather than aberrantly migrating inner retinal cells or some other cell type. Therefore, sections were co-labeled with the photoreceptor cell body marker recoverin (red) and the cone OS marker S-opsin (green- Figure 3). The aberrant cells in the nascent rosettes stain positive for recoverin ( Figure 3, arrows), regardless of age, confirming that they are committed photoreceptor cells.
To further understand the characteristics of these cells, we labeled retinal sections using an antibody against cone arrestin as a marker for mature cone photoreceptor cells ( Figure 4A). In the WT retina, cone-arrestin labeled cone photoreceptors are spread evenly throughout the tissue and the number of positive cells remains relatively unchanged over time. Surprisingly, cone arrestin staining within the ONL in the Nrl 2/2 retina appeared similar to WT at early (P8-10) time points, i.e. not all cones in the Nrl 2/2 express cone arrestin at these ages. These results indicate that even though all the cells of the Nrl 2/2 are developing into cone-like cells they are not doing it at the same timecourse as the native cones of either the WT or the Nrl 2/2 . Furthermore, at P8 and P10, the aberrant photoreceptors were not positive for cone arrestin (Figure 4, arrowheads), confirming they are developmentally delayed and suggesting that they are not authentic cones. At P12 in the Nrl 2/2 retina, the cone arrestin staining became significantly more widespread compared to WT, and the aberrant photoreceptors begin to be cone arrestin positive (Figure 4, arrows), suggesting that these photoreceptors do mature, just at a delayed rate. By P15 all photoreceptor cells in the Nrl 2/2 appear to stain positive for cone arrestin.
Since our results suggest that aberrant cells are underdeveloped photoreceptors, we next asked whether they expressed IS/OS markers. We had previously observed that one of the distinguishing features of aberrant photoreceptors is a lack of ISs and OSs protruding into the subretinal space adjacent to the RPE ( Figure 1A-B), but we could not rule out the possibility that the IS/OS of these cells were misoriented and pointed into the rosette. We therefore closely examined IS/OS labeling in nascent and maturing rosettes, which can both be seen in P12 sections. In the nascent rosettes (P8-10 and some at P12), aberrant photoreceptors are located in the subretinal space and IS/OS labeling ends abruptly at the edge of the population of aberrant photoreceptors ( Figure 5A, arrowhead), and we observe no IS/ OS protruding from the aberrant photoreceptors. However, at P12, as the rosettes begin to mature we observe some rosettes that exhibit a transitional morphology. The aberrant photoreceptors of these rosettes still lack ISs/OSs in the region adjacent to the RPE ( Figure 5B, arrowheads) but start to exhibit PNA and S-opsin labeling protruding from the aberrant photoreceptors inside the rosette ( Figure 5B, asterisks).
To further examine whether the aberrant photoreceptors exhibit ISs/OSs in nascent rosettes, we conducted ultrastructural analysis at P8 and P21. The ultrastructure of the mature rosettes is characterized by an orderly, layered morphology ( Figure 5C). The outside of the rosette is defined by photoreceptor nuclei in the ONL. The next layer inwards comprises ISs, and the lumen of the rosette contains OSs that protrude from cells on all sides of the rosette ( Figure 5C). In nascent rosettes (P8), the outside layer is still composed of the nuclei of the ONL, but instead of having additional concentric layers of ISs and OSs, the IS layer is interrupted by the aberrant photoreceptors which protrude into Figure 5D) and spread into the subretinal space toward the RPE ( Figure 5D).
As we observed on the light level, ultrastructural examination shows that aberrant photoreceptors have no ISs or connecting cilia in the subretinal space and that the cell bodies of the aberrant photoreceptors come in direct contact with the apical surface of the RPE ( Figure 5E, a magnified view of box E from Figure 5D). Similarly, Figure 5F (showing a magnified area of box F from Figure 5D) shows that in the nascent rosette ISs protrude from cells surrounding the rosette but not from aberrant photoreceptors. To help with visualization, the individual cells have been outlined and lightly highlighted in the right panels of Figure 5E-F. For each cell, the cell body is highlighted in blue (for ONL cells) or aqua (for aberrant photoreceptors) and the inner segments are outlined in green.
In contrast to the aberrant photoreceptors, Nrl 2/2 photoreceptors in an area outside the nascent rosettes appear similar to the WT. At P8, the region between the RPE and the ONL is filled predominantly with ISs since OSs are still undeveloped in both the WT and the Nrl 2/2 retinas. This region is not significantly different between the two genotypes ( Figure 6A-B). Randomly oriented connecting cilia are observed in both the WT and the Nrl 2/2 retinas ( Figure 6A-B, arrowheads) as are nicely formed ISs showing that in areas that lack rosettes, the photoreceptors are formed properly and are maturing in a regular fashion.
These data further support the idea that the aberrantly localized cells are underdeveloped photoreceptors which do not label with typical cone markers (arrestin, PNA, S-opsin) on the same timecourse as cells outside the rosettes, but as the retina ages, the cells appear to develop leading to the resolved mature rosette phenotype.
RPE-photoreceptor associations are maintained in the Nrl 2/2 eye
Previous reports have demonstrated that retinal detachment and subsequent retinal folding/rosette formation can be associated with failure of the RPE microvilli to properly associate with the photoreceptor OSs [30]. To examine the possibility that photoreceptors in the Nrl 2/2 retina form abnormal associations with the RPE, we co-labeled retinal frozen sections from WT and Nrl 2/2 mice with antibodies against ezrin, a marker for the RPE microvilli (Figure 7, red), and Rds in WT or S-opsin in Nrl 2/2 as markers for OS (Figure 7, green). In the WT mouse, RPE microvilli surround the apical end of the OSs, visualized as yellow in Figure 7 (arrows, top). Similarly, in the Nrl 2/2 retina, comingling of ezrin and S-opsin labeling in photoreceptors adjacent to the RPE indicates that the RPE microvilli can surround Nrl 2/2 cones (Figure 7, arrows, bottom). No RPE microvilli are detected inside rosettes at any age, note clear green staining only (S-opsin) inside rosettes (Figure 7, asterisks). In contrast, in the areas between the aberrant photoreceptors and the RPE, only ezrin staining is present (Figure 7, arrowheads) consistent with a lack of OSs on the aberrant cells. These results suggest that when cone OSs and RPE microvilli are present, associations are formed. When photoreceptors fail to properly orient their OSs towards the RPE no association is observed. These results imply that rosettes in the Nrl 2/2 are not formed as a result of any inherent inability of cone OSs to associate with the RPE. This point is further supported by the structural analysis and developmental timecourse of the Rds 2/2 Nrl 2/2 retina which shows no change in rosette structure or number in response to dramatic changes in OS structure ( Figure S1).
Defects in OLM formation are associated with rosette formation
Improper photoreceptor layer development can also be due to defects in the formation of OLM [11,16]. The OLM is composed of adherens junctions between photoreceptors and neighboring Müller glia, and it is possible that in the absence of rods, these junctions do not form properly. To test this hypothesis we labeled retinal sections with antibodies against two components of OLM junctions, Zo-1 and beta-catenin [31] (Figure 8, red). In the WT retina, the OLM is visualized as a solid line distinctly separating the ONL from the inner segment (IS) layer ( Figure 8A-B, top). In the Nrl 2/2 retina, the OLM forms and regions lacking rosettes appear normal. However, in areas exhibiting nascent rosettes, a gap in the OLM can be seen ( Figure 8A, arrows). The aberrant photoreceptors protrude through the gap in the OLM ( Figure 8A, bottom, arrowheads) into the IS/OS region causing displacement of the surrounding nuclei and the formation of rosettes. A similar labeling pattern is observed when beta-catenin is used to decorate the OLM ( Figure 8B). Furthermore, no OLM staining is detected surrounding the aberrant photoreceptors. In the adult Nrl 2/2 retina, the OLM appears more normal with labeling that lines the rosettes, clearly dividing the entirety of the ONL from the IS layer.
Given the gaps in OLM that we see by immuno labeling with OLM markers, we wanted to examine the ultrastructure of the OLM where it terminates at the edge of the population of aberrant photoreceptors. This region was highlighted in brackets in Figure 5D and is shown again in Figure 9A for reference. A magnified view of this area is shown in Figure 9B (magnification of box from Figure 9A). For clarity, cells the right panel have been outlined; ONL cell bodies-blue, aberrant photorecptor cell bodies-aqua, ISs-green, OLM junctions-red. Each photoreceptor cell in the ONL that lines the rosette forms OLM junctions with its neighboring cells (Figure 9, red). However, this pattern of junctions and ISs is not preserved in the ring of cells surrounding the base of the population of aberrant cells. One such cell is labeled with an asterisk in Figure 9. The nucleus of this cell is found with the other aberrant photoreceptors, not in the ONL, and it is misoriented. The cell body protrudes from its nucleus towards the ONL and it forms an OLM junction with one ONL cell (black arrowhead, Figure 9), but on the other side forms no junctions (white arrowhead, Figure 9). Furthermore, although the cell body of this first aberrant photoreceptor protrudes past the To analyze the integrity of the OLM in the Nrl 2/2 retina, sections were stained with antibodies specific for Zo-1 and beta-catenin (red), components of the OLM junctions, and S-opsin (green) to label OSs. A. In the WT retina, the Zo-1 staining forms a straight continuous line. In the Nrl 2/2 retina, gaps in OLM staining (arrows) can be seen that correspond spatially and temporally with the appearance of nascent rosettes and aberrant photoreceptors (arrowheads). B. Beta-catenin staining can be seen at the OLM as well as in the OPL and INL. In the Nrl 2/2 retina, gaps in OLM staining (arrows) can be seen that correspond with the appearance of nascent rosettes and aberrant photoreceptors (arrowheads). OLM: outer limiting membrane, ONL: outer nuclear layer, INL: inner nuclear layer. Scale bar, 20 mm. doi:10.1371/journal.pone.0032484.g008 nucleus, it does not exhibit an inner segment. This pattern of abnormal junctions and no ISs for the cells surrounding the edge of the population of aberrant photoreceptors is consistent among all observed rosettes at P8 and is also observed in nascent rosettes at P12. These observations imply that the aberrant photoreceptors retain some ability to form cell-cell adhesion complexes but nevertheless fail to form proper interactions with all neighboring cells. This likely leads to the abnormal protrusion of aberrant photoreceptors through the OLM and IS layer into the subretinal space.
Since OLM defects are observed in the developing Nrl 2/2 retina it is important to establish the localization of the Müller glia cells that help form the OLM. In order to visualize the Müller cells we stained WT and Nrl 2/2 retinal sections with antibodies against the Müller cell marker aquaporin 4 (AQP4) [32,33]. In the WT retina at all time points observed the AQP4 staining is visible from the inner limiting membrane to the OLM. At the OLM the AQP4 staining can be seen to end in a distinct line that corresponds to the position where the Müller glia cells take part in the formation of the OLM (Figure 10). In the Nrl 2/2 retina the Müller glia cells also terminate in a distinct line that corresponds to the location of the OLM (Figure 10). In the region where the population of aberrant photoreceptors connect to the ONL, there is no Müller cell labeling, consistent with the lack of OLM staining ( Figure 10, asterisks). Furthermore, at P8 and P10 there is no Müller glia staining within the population of aberrant photorecptors (arrows) suggesting that the Müller cell processes have not successfully penetrated into this group of cells. At P12 however clear staining for Müller glia cell bodies is evident within the population of aberrant photoreceptors (Figure 10, arrowhead), further reinforcing the fact that the aberrant photoreceptors are in the process of finishing their maturation process at this time. By P15, Müller cell staining can be seen from the inner limiting membrane to the OLM.
Discussion
Here we show that rosette formation starts at P8 in the Nrl 2/2 retina when photoreceptor nuclei begin to abnormally protrude into the subretinal space and induce ONL infolding (Fig. 11A). These aberrant photoreceptors are not actively dividing or undergoing apoptosis, rather they are underdeveloped and their abnormal localization is associated with gaps in the OLM. By P12-15 the aberrant photoreceptors are in the process of maturing, leading to the adult rosette phenotype lacking this population of underdeveloped cell bodies. This maturation is accompanied by the establishment of a largely continuous OLM, but overall ONL folding remains until retinal degeneration and thinning of the ONL begin later in the life of the animals [34].
While rosettes are seen in multiple genotypes and are a symptom of a variety of genetic retinal diseases, the root cause in each individual genotype is varied and it is important that each example of rosettes observed is examined independently [11,30,35]. Rosettes are reported in the naturally occurring Nr2e3 mutant mouse model (rd7) and were previously thought to be due to over-proliferation of S-cones [35]. Recent work however has demonstrated that the excess S-cones in the rd7 mouse are due to a failure of a population of early born rod photoreceptor precursors to suppress cone specific genes [36]. While the reason for rosette formation was not directly studied in the rd7, both Nrl and Nr2e3 are within the same hierarchal developmental pathway, and our data here indicating that delayed maturation of a subpopulation of photoreceptors results in rosettes in the Nrl 2/2 suggests that a similar process could be at work in the rd7 mouse model. Rosettes are also observed in the nonfunctional Crumbs 1 (CRB1) mutant mouse model (rd8) [11,16]. Without CRB1, a component of adherens junctions, photoreceptors fail to properly form the OLM, a defect which leads to rosette formation [11,16]. Finally, rosette formation has been shown to be associated with aberrant interaction between the RPE microvilli and photoreceptor OSs. For example, knockdown of the chloride intracellular channel 4 (CLIC4) has been shown to prevent proper association between OSs and the RPE microvilli, in turn causing localized retinal detachment and the formation of rosettes [30]. Taken together these results emphasize the need to approach the appearance of rosettes in each genotype and determine their individual root cause on a case by case basis.
The data presented here support the hypothesis that rosettes in the Nrl 2/2 mouse result from defects in the OLM and the presence of an underdeveloped population of photoreceptors. This observation is consistent with previous work by Dr. Janet Blanks' group showing that chemical disruption of Müller cells during retinal development in the WT mouse results in the same protrusion of photoreceptor cell nuclei into the subretinal space that we observe in nascent Nrl 2/2 rosettes [15]. Although chemical disruption of Müller cells results in gross alterations in Müller cell morphology, the resultant nuclei protrusion/rosette formation is consistent with the phenotype observed in cases of targeted OLM defects, e.g. in the case of CRB1 mutations [11,16]. Here we show that Müller cell processes fail to extend into the population of aberrant photoreceptors found in nascent rosettes, consistent with the fact that Müller cells are required for the formation of the OLM.
These observations however, lead to the question, why does the cone-dominant retina exhibit defects in OLM formation? We have previously demonstrated that gross development and radial alignment of Müller cells in the Nrl 2/2 retina is normal [26], an expected outcome given that the NRL transcription factor is not involved in Müller cell differentiation. However, we here show that at early stages in development, Müller cells are not found in the regions where the underdeveloped photoreceptors are inducing rosette formation. While the exact cause of the OLM disruption in the Nrl 2/2 retina remains unclear, one possibility is that rods are required for proper OLM development. The adherens junctions which comprise the OLM consist of components expressed by both the photoreceptor cells and the Müller cells. Rods and cones may contribute different components, and in the absence of rod-mediated adherens junctions, the OLM does not form properly. We suggest that without rods, OLM formation is perturbed and some of the interactions which are required for proper photoreceptor positioning are lost. It is unclear why disruptions in the OLM resulting from a lack of rods would appear as localized disruptions, resulting in rosettes, instead of a global failure of OLM formation however, this localized rosette phenotype is consistent with observations from previous models [11,15,16]. For example, chemical disruption of the Müller cells and genetic disruption of the OLM (rd8) would both be expected to cause defects in the entire OLM, but actually result in a localized disruption in the OLM similar to what we observe here in the Nrl 2/2 retina [11,15,16]. In addition to being mislocalized, initially the aberrant photoreceptors are underdeveloped compared to other photoreceptors; they do not exhibit staining typical of ISs/Oss, no ISs/ OSs are observed on the ultrastructural level and they lack important markers of development such as cone arrestin. This phenotype is likely tied to the observed OLM defect as OLM components and regulators such as CRB1 are critical for the development of photoreceptor cell polarity [12]. In addition, mutational studies have shown that Crumbs genes play a role in regulating the IS membrane size in photoreceptor cells and have been shown to impact cilia formation in renal epithelial cells [37]. The lack of proper cell-cell adhesions between the aberrant photoreceptors and their neighboring cells may perturb their ability to determine apico-basal polarity and likewise inhibit their ability to properly develop ISs and OSs.
Our results suggest that the aberrant photoreceptors found inside nascent rosettes do mature with time, i.e. they eventually form cell-cell junctions, IS, and OSs, and thus contribute to the growth of a mature rosette ( Figure 11). Our data on the variety of rosette phenotypes seen at P12 support this idea. This ''delayedmaturation'' hypothesis would result in multiple mature rosettes for every nascent rosette (Figure 11), and is consistent with our data showing that the number of mature rosettes in the adult Nrl 2/ 2 is ,3-fold higher than the peak number of nascent rosettes.
Critically, our findings are relevant to many retinal diseases. Defects in specific OLM components have been previously associated with retinal degenerations, but our data here, indicating that the absence of rods results in OLM abnormalities, may further our understanding of the mechanisms underlying rod-cone degenerations. Many diseases which affect rods, such as retinitis pigmentosa, subsequently lead to the loss of cone cells [38]. This trend is also observed in animal models; for example in VPP mice carrying rhodopsin mutations and exhibiting a retinitis pigmentosa phenotype, rapid loss of rod cells is followed by a slow degeneration of cone photoreceptor cells [39]. In these cases the loss of rod cells is often due to a mutation in a rod specific protein (such as rhodopsin) but the precise mechanism is poorly understood. Recent data suggests the loss of rods leads to a secondary loss of cones due to a combination of retinal detachment, oxidative stress and activation of microglia [34]. Our data show that rods are critical for proper OLM formation and maintenance which suggests that in addition to these factors another mechanism underlying secondary cone degeneration may be OLM defects.
These data may also be relevant for the development of effective cell-based therapies for retinal degenerations. It has been shown that when chemical disruption of the OLM is used in conjunction with cell therapy in WT mice, greater integration of the transplanted photoreceptor cells occurs [40]. Our data suggest that OLM defects may be a part of the normal pathology of a variety of retinal diseases in cases where rods are lost. This facet should be taken into account when considering the optimal experimental design and timing of delivery for cell therapy.
Here we present the first report demonstrating that the Nrl 2/2 retina exhibits defects in the OLM, and that this defect is associated protrusion of underdeveloped photoreceptors into the subretinal space leading to rosette formation. Even though the OLM defect is detected for only a short window during development of the mouse retina, it results in rosettes a persistent rosette phenotype. Importantly these data suggest that rod cells (which are absent in the Nrl 2/2 retina) are required for proper OLM formation. Figure S1 Rosette development is not different in the Rds 2/2 Nrl 2/2 compared to the Nrl 2/2 . Left-Retinal sections at the indicated ages were labeled with Alexa 488 conjugated peanut agglutinin. Small nascent rosettes (arrowheads) and mature rosettes (arrows) are observed. NBL: neuroblast layer, IPL: inner plexiform layer, ONL: outer nuclear layer, OPL; outer plexiform layer, INL: inner nuclear layer. Scale bar 20 mm Right-Rosettes were counted in 3 sections/eye, 4-5 eyes per age/ genotype. Shown are means 6SEM. Black lines-nascent rosettes, Gray lines-mature rosettes. Figure S3 Aberrant photoreceptors are not cleared through apoptosis. A. TUNEL staining (red) was performed and followed by staining of OSs using S-opsin antibody (green). TUNEL positive cells can be seen in the INL as part of normal retinal development (arrows). All aberrant photoreceptors are negative for TUNEL staining at all time points examined (arrowheads). As previously noted, at P15 both nascent rosettes (containing aberrant photoreceptors) and mature rosettes are observed. We hypothesized that aberrant photoreceptors at this timepoint would be the most likely to be undergoing apoptosis if cell death is the mechanism underlying transition from the nascent rosette phenotype to the mature rosette phenotype. However, we did not detect any apoptotic cells in the nascent rosettes remaining at P15 (asterisks). ONL: outer nuclear layer, INL: inner nuclear layer. Scale bar, 20 mm. (TIF) | 9,034.2 | 2012-03-12T00:00:00.000 | [
"Biology"
] |
Heteroscedastic Personalized Regression Unveils Genetic Basis of Alzheimer’s Disease Stratified by Cognitive Level
In contemporary medical research, patient heterogeneity plays a pivotal role in comprehending intricate diseases such as Alzheimer’s disease and various forms of cancer. Specifically, in the genomic analysis of Alzheimer’s disease, individual patients may exhibit unique causal mutations that significantly influence their therapeutic trajectory. Conventional models that share numerous parameters across all individuals struggle to discern this heterogeneity and identify the influential factors for individuals. To tackle this challenge, we propose an innovative approach called Heteroscedastic Personalized Regression (Het-PR) to estimate the heterogeneity across samples and obtain personalized models for each sample. We demonstrate the effectiveness and robustness of Het-PR through both simulation and real data experiments. In the simulation experiment, we show that Het-PR outperforms other state-of-the-art models in capturing inter-sample heterogeneity. In the real data experiment, we apply Het-PR to Alzheimer’s data and show that it can identify persuasive selected genetic factors for each individual patient. Interestingly, our results suggest that there might be different associative SNPs for AD patients stratified by different cognitive levels. Author summary In medical research, it has been observed that causes of a disease vary significantly among individuals, especially when looking at complex diseases like Alzheimer’s disease and cancer. For Alzheimer’s disease, obesity, age, gender, and depression may play different roles across different patients. When studying the genes of Alzheimer’s patients, we find that each person might have their own unique genetic changes that can affect their treatment. For example, Alzheimer’s patients with different genetic mutations may respond differently to the same treatment. Traditional research methods often miss these individual differences and can’t always pinpoint important personalized factors for each patient, because they usually use one model for all patients. To better understand these differences, we’ve introduced a new method, Heteroscedastic Personalized Regression (Het-PR), which generates a personalized model for each individual. Our experiments show that Het-PR is more effective than other leading methods in identifying these patient differences and recognizing Alzheimer’s genetic basis for each patient through both simulation and real data experiments. When we used Het-PR on real Alzheimer’s data, it helped us spot key genetic factors for each patient. Additionally, in our study, we excitedly find that different genetic markers in Alzheimer’s patients are possibly based on their cognitive abilities. Software for Heteroscedastic Personalized Regression is available in https://github.com/rong-hash/Het-PR.
Introduction
The prevalence of heterogeneous diseases, such as Alzheimer's and cancer, is increasing across patients [1,2].For example, Alzheimer's disease is influenced by various factors including gender, mid-life obesity, depression, and others, which contribute to significant disparities in patients' clinical manifestations [3,4].Among all latent factors, genetic basis, such as SNP, is one of the important heterogeneous factors that influence disease progress, treatment response, and clinical trial design.For instance, APOE ϵ4 carriers may respond differently to certain treatments compared to non-carriers [5].What's more, as for the influence on disease progress, a study in 2021 showed that APOE ϵ4 carriers and individuals with abnormal baseline tau levels showed a faster decline at the group level, but also greater within-group variability [6].Consequently, understanding this heterogeneity is imperative for the development of personalized treatment strategies and the design of efficacious clinical trials.To comprehensively understand this heterogeneity, it is essential to construct a personalized profile for each patient.
While conventional models estimate a single set of common parameters for all samples, personalized regression models offer several advantages in uncovering heterogeneity: 1) These models can identify specific and informative factors relevant to individual patients, thus providing more refined insights compared to traditional machine learning models.2) Upon identifying the key contributing factors for individual patients, it becomes feasible to develop tailored therapeutic strategies that address each patient's unique characteristics.3) Personalized regression models only take a small amount of data to build a personal model for each patient.
With the incorporation of personalized effects, the model becomes more interpretable and can offer valuable insights into clinical treatment and pathological analysis.In fact, personalized models have also been shown to solve problems in various domains, including finance, society, and more [7,8].
Prior studies have illustrated the potential of diverse personalized regression models in the medical field.Conventional personalized regression models predominantly employ penalized terms to achieve personalization for individual samples.For example, Localized Lasso [9] extends the regularization term of Network Lasso [10] to facilitate personalized effects; Penalized Angular Regression accomplishes personalization by establishing distinct penalized terms for each sample [11]; Personalized Regression with Distance Matching achieves personalization by matching covariate distances and personalized parameter distances [7,8].While previous personalized models have demonstrated effectiveness in prediction tasks, their performance in feature selection tasks remains suboptimal, thus lacking sufficient support for the physiological analysis of patients.
However, these methods typically face two primary challenges: 1.The specifically designed regularization to express the relationship restricts the model's applicability to other scenarios; 2. The estimation of the model parameters for a single sample usually relies primarily on a subset of samples, which limits the statistical efficiency of the estimation process.3. The models usually need a large amount of data to achieve September 1, 2023 2/15 personalized effect, which is hard to access in the real world.
To address the above issues, we propose a novel personalized regression framework that employs heteroscedasticity (Het) assumption through a Linear Mixed Model (LMM) construction for estimating individualized parameters of each patient.This approach diverges from the conventional personalization methods that rely on regularization.In our proposed framework, we construct an LMM for each sample based on the distance among the sample features.To be more specific, given a matrix of data with dimensions n × p, where n denotes the number of samples and p represents the number of features, Personalized Regression can be trained with this data to estimate a set of parameters of dimensions n × p, for example, p-value of the importance of each feature associated with the label for feature section.By sorting the p-values and establishing a threshold, the top-ranked features can be extracted for further study with potential relationships with the labels.
To assess the efficacy of the proposed method, we carried out both simulation and real data tests.Our simulation tests indicated that our model exhibited exceptional performance in the AUCROC score, showcasing its robust variable selection capabilities.
In the real data test, we applied our model to the Alzheimer's Disease Neuroimaging Initiative (ADNI) database.We found that the Single Nucleotide Polymorphisms (SNPs) extracted for each patient displayed high interpretability and reliability.Notably, many genes corresponding to the selected SNPs had been previously validated as Alzheimer's-associated genes.Furthermore, the extracted SNPs exhibited a strong correlation with the patients' cognitive levels, underlining the potential of our model in facilitating the development of personalized medical treatment plans.
Background -Linear Regression
To elucidate the methodology underlying the Het-PR model, we begin with the notation and the simple linear regression model.
Given the dataset of (X, Y) where X ∈ R n×p and Y ∈ R n , where p < n, and (x, y) to denote one sample of (X, Y), we use subscripts to denote samples and superscripts to denote features, thus X (j) i denotes the j th feature of the i th sample.
Linear regression assumes the data generation process of where β is the p dimension coefficients, and ϵ denotes a vector of length n of residue error.
It is well known in the community that, if ϵ ∼ N (0, σ 2 ϵ ), we can have as an unbiased estimator.
Background -Personalized Regression
For Personalized Regression, instead of sharing the parameter β among all the samples, we assign a particular parameter set for each sample, which can be shown as where i represents the i th sample.
September 1, 2023 3/15 The major differences between personalized regression and linear regression are that we now assume there is a coefficient vector associated with each sample, for which we use β i to denote the coefficients associated with the i th sample.Further, if we use a matrix B ∈ R n×p to denote the coefficients for all samples, we can have However, it will be almost impossible to use the traditional MLE-based method i.e., (2) to solve the personalized regression problem (3) because now we only have one sample for every p coefficients.Fortunately, previous studies have demonstrated several solutions to estimate the coefficients in such cases with additional assumptions.One assumption that Lengerich [7] uses is that closely related samples will have similar coefficients.Therefore, the problem is formulated as follows.
where d 1 , d 2 , and d 3 are three customized distance functions and C i , C j are covariates such as demographic information.λ is another hyperparameter.With this setup, similar samples will have similar covariates because of the constraint.
As an example, we can consider a more specific form of the above formulation ( 4) with d 1 , d 2 , and d 3 assigned concrete choices: With this specific form, if we can write the main neighborhood assumption in [7] back to the data generation process, (4) estimates the parameters of the following data generation process: where we again assume ϵ ∼ N (0, σ 2 ϵ ), meaning the above problem is a homoscedasticity problem.
In a loosely speaking manner, we can consider the problem in (4) as a "heterogeneous" problem because of the variation introduced through the coefficients for different individuals, but it is homoscedastic because the error term associated with each individual is of the same variance.
Heteroscedastic Personalized Regression
In this section, we start to introduce our new method.Loosely speaking, our core idea lies in an exchange of variations introduced in linear regression.In contrast to (4), which is a heterogeneous and homoscedastic problem, we formalize the personalized regression as a homogeneous and heteroscedastic one.This new formalization offers us a more efficient manner in estimating the coefficients for each individual, and also comes with extended flexibility to immediately turn different MLE-based regression methods into their personalized version.
Design Rationale We first introduce the design rationale of our method by mismatching the sample of interest to the coefficient interest.For example, we study (X j , Y j ), but with B T i , so, we have Following our heteroscedasticity view of the problem and assuming the differences introduced from B j − B i can be modelled as a noise term, we can continue to substitute (B j − B i ) T with u j , and we can get where u j satisfies u j ∼ N (0, Iσ ij ), and σ ij is d(B i , B j ), where d is the customized distance function.For simplicity, here we use the same concrete form discussed in the background section, therefore we have If we continue with the data generation process in (6), we can immediately get σ 2 ji as The above derivation shows that we can use all the samples to estimate the coefficients B i even though Y j are generated with B j .The main disparity introduced by such a mismatching is the differences raised from X j B j and X j B i , and we consider this disparity as an error term associated with sample j.In other words, all the samples can be used to estimate the coefficients of sample j, if the error term for every individual is constructed differently.Thus, loosely speaking, we turn the personalized regression problem into a "homogeneous" but heteroscedastic problem.
Heteroscedastic Personalized Linear Regression Model With the above setup, for sample j, we have where we do not require any explicit relationship between j and i.Therefore, to estimate the coefficient of any samples (e.g., sample i) we can use all the samples together.Thus, we have: September 1, 2023 where ⊙ is element-wise product, and σ 2 is a vector of length n with its j th element to be σ 2 ji .
Conveniently, following the data generation assumption, although σ 2 ji has n elements, it only has one parameter γ, the rest are determinable through the knowledge of C.
As one might notice, ( 12) is a standard form of linear mixed model (LMM).This convenient connection back to the linear mixed model will allow us immediately to re-use all the mature linear mixed model estimation algorithms to estimate the coefficients of our model, once the error term is constructed properly for the sample of interest.In this work, we use the FaSTLMM algorithm [21] through the implementation of [22].It is also worth explicitly mentioning that, for every sample of interest, our algorithm needs to construct the error term and perform the estimation.Thus, it requires n runs if there are n samples of interest for personalized estimation.
Heteroscedastic Personalized Model Framework Another significance of our method is that for every population-based model, denoted as where f (•; Θ) is a function with Θ denoting the parameters.We can reuse our above idea to turn it into its personalized version with which allows us to estimate the personalized coefficients for sample i.
In the experiment section, we will refer to f (X; Θ) as the "base model" and explore two choices.
Simulation Test
Simulation Data Generation We first test our methods in comparison to other methods with simulation data.We start with a simpler situation, where instead of the situations where every individual has one's own set of oracle effect sizes, we group the individuals into different cohorts, and individuals of the same cohort will have similar parameters.We generate the simulation data with the following process.We first sample n samples, each with p features from a uniform distribution.We denote the data as X.These samples are evenly split into g cohorts, and each cohort will have its unique effect size vector β with the same sparsity level (k features will have non-zero effect sizes).Each of the k non-zero values of the β vector will be sampled as follows: We first sample a scalar a ∈ N (0, 1), then where s is a scaling factor controlling the effect sizes.
Samples within the same cohort will share the same effect size β.Thus, for the i th sample, the label Y i , will be where ϵ ∈ N (0, σ 2 ϵ ) and following the notation of this manuscript, we use B to denote the matrix of effect sizes β.Further, we use σ ϵ as another hyperparameter that controls the simulation process.
Finally, to test our model's behavior, we need to pass the covariates of the demographic information.At the simulation state, we simply use the oracle information of the partition information, i.e., the B, as the demographic information C. Simulation Result Since we are interested in recovering the genetic basis of the disease patterns, we evaluate these methods for the ability to recover the non-zero coefficients.We test the methods with AUCROC score of comparing estimated "importance" of coefficients to the Oracle coefficient sparsity pattern.In particular, based on the nature of the methods, the "importance" are p-values [23] when available or the absolute values of the estimated coefficients when p-values are not available.As we discussed before, our Het-PR model can be equipped with different "base model", and we tested two versions, the "pl pvalue" denotes the Het-PR with a base model that is a variant of linear mixed model that can report p-value even for high-dimension case [22], and "pl beta" denotes the Het-PR with a base model of linear regression.dmr model refers to personalized regression with distance matching [7,8], par refers to penalized angular regression [11], and pr is population regression.Our results demonstrate that the Het-PR model consistently outperforms other models in accurately recovering the original sparsity pattern of the coefficients across a range of parameter values.Specifically, we conducted five repeated simulations with different random seeds for each parameter set, which varied in terms of sample size (n), number of features (p), Sparsity Level (k), and number of cohorts (BlockN um).
After setting the hyperparameters to different values and running the simulation experiment with 16 parameter sets, each parameter set with 5 random seeds, we got the overall performance distribution shown in Figure 1, and detailed model performance is shown in Figure 2. We can find our model performs better than other models in most cases.
Alzheimer's Experiment
With the success of simulation experiments, we continue to move on to apply our method to understand the genetic basis of Alzheimer's disease at a personalized level.
Alzheimer's disease (AD), one of the most lethal ailments affecting the elderly population, has been a focal point of interest for biomedical research communities over several decades.Despite the extensive efforts dedicated to understanding the disease's mechanisms, effective treatments for AD remain elusive [24][25][26][27][28], owing to the absence of a widely accepted comprehension of its pathology [29].For instance, it was long theorized that amyloid plaques in the brain were responsible for AD [30][31][32][33][34][35].However, this view has been recently contested, suggesting that plaques are merely a result of decreasing levels of soluble amyloid-beta in the brain when normal proteins transform into abnormal amyloid plaques under conditions of biological, metabolic, or infectious stress [36,37].Consequently, previous treatments targeting amyloid plaques have become a subject of debate [38].
One factor contributing to the challenge of understanding AD is its heterogeneity.
Over the years, researchers have identified several AD subtypes from various perspectives.Genetically, AD is divided into two subtypes: familial AD and sporadic AD [39][40][41][42][43].However, the genetic basis of sporadic AD is still elusive until now, and a hypothesis behind this challenge is that the genetic pattern of sporadic AD might not be universal across all populations as there are typically three or four subtypes of AD based on tau pathology and brain atrophy [44][45][46][47].
From a pathological standpoint, AD is typically classified into three or four subtypes based on tau pathology and brain atrophy [44][45][46][47].Clinically, patterns have emerged, such as the presence of amnestic syndrome, language disturbances, visuospatial skill impairments, attentional function deficiencies, and executive process and praxis disruptions [48].However, clear associations between subtypes across these different perspectives have not been established [49].Moreover, AD exhibits a significantly different prevalence between female and male patients, with women affected 1.7 times more frequently than men [4].This observation implies the existence of additional subtypes when considering gender as a covariate, although previous AD studies have seldom accounted for it.Therefore, our personalized regression model can potentially identify the genetic patterns of AD by considering its heterogeneity, thus constructing a personalized profile for each patient.We test our model on the ADNI 1 genotype data, which contains 192 patients in total, consisting of 79 patients (sporadic AD) and 113 normal controls.Each sample has 257361 SNPs.Each SNP has 3 alternative values, 0, 1, 2, representing no variation, 1 variation, and 2 variations respectively.
Due to the small sample size, in this study, we only use exon markers as X for each sample, and the rest of SNPs as C. Thus, we do not have to use covariates explicitly offered by the ADNI in our model but we can save them to evaluate whether the genetic September 1, 2023 9/15 patterns we identified are meaningful.After training on the data using Het-PR, we can get the p-value of SNP for each sample, and we consider the threshold of p-value as 0.05.We first report and analyze the results at the population level by collecting the most frequently reported SNPs across all the samples.We report the top 10 SNPs for all patients in Table 1.In Table 1, we find that 7 out of these 10 genes are strongly related to Alzheimer's [50][51][52][53][54][55][56].Furthermore, although we don't have strong evidence that the other 3 genes are related to Alzheimer's, we still find it related to brain activity.For example, for the most selected gene CFAP73, it's an important paralog of the gene CCDC42, which is specifically expressed in testis and brain [57].Overall, the result suggests that our model can identify SNPs implicated with AD.
Further, since our model is particularly designed to understand the personalized genetic basis, we further report the results for its ability in unveiling the genetic basis of each patient.For the discussion, we consider the patients whose selected SNPs are more than 10.For each patient, the top 10 genes corresponding to the SNPs are shown in Table 2.
To offer a further understanding of these results, we conduct a hieratical clustering based on the p-value of the SNPs, shown as Figure 3. Based on the clustering, we can divide the patients into two groups.Group1 is colored blue and Group2 is colored orange in Figure 3.
Interestingly, we notice that this group partition is correlated with the group created by patients' cognitive level, which can be represented by 8 authoritative targets: CDRSB, ADAS11, ADAS13, ADASQ4, MMSE, RAVLT immediate, RAVLT learning, RAVLT perc forgetting.The meaning of these 8 targets is shown in Table 3 We compare the scores between Group1 and Group2 and the result is shown in Figure 4. From the experiment, we find that eight out of eight targets show that the cognitive level of Group1 is lower than Group2, meaning that the selected SNP is September 1, 2023 10/15 strongly related to the patients' cognitive level.The p-value of the t-test is shown in Figure 4. We consider these results very interesting as these discussions might be one of the first results to show that the genetic basis of AD patients might be stratified by the cognitive level of patients.
Discussion
In this study, we present a novel model called Het-PR, designed to perform Personalized Regression through the implementation of a Linear Mixed Model.Our proposed model demonstrates superior accuracy in recovering cohorts, finding the personalized genetic basis compared to the current state-of-the-art methods, as validated through an extensive simulation study and real data experiments.In the context of Alzheimer's disease research, the Het-PR model's ability to select SNPs proves to be highly reliable, based on two key factors: (1) by verifying whether the selected SNP has previously been identified to be implicated with AD, and (2) more interestingly, showing a first attempt to dividing patients genetic basis based on cognitive levels.The findings of this study suggest that the Het-PR model has the potential to be further employed in determining the stage of Alzheimer's disease, offering a chance to the field of personalized medicine and Alzheimer's disease stage assessment.
Although we only conducted the experiment on Alzheimer's SNP data, it's worth noting that our model can be applied to any heterogeneous medical data for further research, such as pan-cancer analysis, Asthma's pathological analysis, depression analysis, etc.The Het-PR model is also potential to be applied in financial, social and other fields.
In the context of this research, several limitations must be acknowledged.First of all, the constrained size of the dataset impedes a definitive understanding of the relationship between cognitive levels and their genetic determinants.As a logical progression, it is imperative to apply the proposed methodology to an expansive Alzheimer's database to garner more comprehensive insights.Additionally, the current framework, especially concerning the expansion to σ 2 ⊙ diag(XX T ) as delineated in Equation 14, represents a foundational model.This prompts a series of academic inquiries: How might one adapt σ to accommodate individual covariates?Is there a theoretical justification for incorporating the complete matrix?These considerations, among others, necessitate rigorous exploration in subsequent research endeavors.
Fig 2 .
Fig 2. Model Performance under Different Parameters.There are 16 parameter sets in total, each shown in the title of the sub-figure.We conduct five experiments per parameter set with different random seeds.
Fig 3 .
Fig 3. Hieratical Clustering based on pvalue of SNPs, Group1 is colored blue, and Group2 is colored by orange, which is the same as Figure 4.
Fig 4 .
Fig 4. Cognitive Target Score between Group1 and Group2.For every test score, it shows that the cognitive level of Group 1 is lower than Group 2.
Table 1 .
Top 10SNPs across all patients.Alzheimer's related genes are bolded.NA means this gene is not named yet.Most influential genes are in the protein-coding type.
Table 2 .
Top 10 influential genes for each patient
Table 3 .
Cognitive Level Target.8 cognitive test scores given by ADNI | 5,580.4 | 2023-09-15T00:00:00.000 | [
"Biology"
] |
Visual fixations rather than saccades dominate the developmental eye movement test
When children have visual and/or oculomotor deficits, early diagnosis is critical for rehabilitation. The developmental eye movement (DEM) test is a visual-verbal number naming test that aims to measure oculomotor dysfunction in children by comparing scores on a horizontal and vertical subtest. However, empirical comparison of oculomotor behavior during the two subtests is missing. Here, we measured eye movements of healthy children while they performed a digital version of the DEM. In addition, we measured visual processing speed using the Speed Acuity test. We found that parameters of saccade behavior, such as the number, amplitude, and direction of saccades, correlated with performance on the horizontal, but not the vertical subtest. However, the time spent on making saccades was very short compared to the time spent on number fixations and the total time needed for either subtest. Fixation durations correlated positively with performance on both subtests and co-varied tightly with visual processing speed. Accordingly, horizontal and vertical DEM scores showed a strong positive correlation with visual processing speed. We therefore conclude that the DEM is not suitable to measure saccade behavior, but can be a useful indicator of visual-verbal naming skills, visual processing speed, and other cognitive factors of clinical relevance.
Visual deficits that are not directly related to an abnormality in the peripheral visual system, but rather to an injury in brain areas that play a role in the perception or interpretation of visual information, are often difficult to recognize (for a review, see 1 ). Especially children born with visual deficits might not realize that their perception of the world is distorted. However, clinical tools to diagnose deficiencies in visual information processing and oculomotor functioning at an early age are limited 2 . Apart from visual and or oculomotor dysfunction, the child can show attentional or cognitive deficits, higher cortical visual processing, or language retrieval issues too 3,4 .
The Developmental Eye Movement test (DEM) is a visual-verbal reading task originally developed to assess oculomotor function in children without the use of eye tracking techniques 5 . In this test, subjects first read out loud a standardized list of vertically arranged numbers and then name numbers in a two-dimensional array row by row (Fig. 1a). The vertical subtest primarily calls for small equally spaced vertical saccades, whereas the horizontal subtest requires horizontal saccades of varying magnitude. The assumption is that, by eliminating the requirement for horizontal saccades of varying amplitude, the physiological constraints on the oculomotor system are reduced in such a way that the vertical subtest is predominantly determined by automatic visual-verbal naming skills (automaticity). According to this assumption, performance on the vertical subtest can be used to factor out the effects of automaticity in the horizontal subtest. An increase in both horizontal and vertical reading time compared to age-matched controls would thus indicate a difficulty in general number naming skills, which in healthy children dominates between 64 and 90% of the time spent in horizontal DEM 6 . An abnormally increased time on the horizontal test combined with a normal performance on the vertical test (a high DEM ratio) would be characteristic of oculomotor dysfunction. However, Garzia et al. 5 did not measure eye movements or verified otherwise whether the horizontal to vertical ratio truly reflect oculomotor skills.
Therefore, the DEM's use for the assessment of oculomotor function has been questioned. Test-retest reliability of the DEM ratio is only moderate to low [7][8][9] . In addition, DEM performance does not correlate significantly with parameters of horizontal saccades measured in saccade tasks that require no numerical processing S2A). Younger children made more errors than older children did. Horizontal times for children with more than two errors were comparable to the horizontal times of age-matched children who made two errors or less ( Supplementary Fig. S2B). We therefore decided not to exclude any children based on the number of errors made.
DEM scores are age-dependent.
To determine test-validity of the digital DEM, we studied the relationship between age and DEM time in all 91 participants. In line with previous studies, younger children (6-7 years old) needed more time to complete either DEM compared to older children (11-12 years old). We found a significantly negative correlation between age and vertical time (r = − 0.79), age and horizontal time (r = − 0.81) and age and DEM ratio (r = − 0.39 all p values < 0.001; Fig. 1b). The latter suggests an increasing dominance of the naming process in horizontal DEM with age. The mean and standard deviations per age group Table 1) fell within the mean ± SD given in previous reports 5,6,8 . This indicates that the digital DEM can be used in addition to or as a replacement of the original DEM. Only the numbers of errors were lower in our study population. We think this is because we allowed children to briefly practice each subtest (with different numbers), instead of just showing an example row and explaining the task in words (Methods). We did this to ensure a correct understanding of the tasks.
Eye movements during horizontal and vertical DEM. Next, we analyzed the saccade behavior during the horizontal and vertical DEM to determine how oculomotor skills influence the time needed to accomplish these subtests. Eye-tracking data were not available for all children, as it is difficult to obtain complete data sets in this particular population. For 20/91 children, no eye movements were recorded owing to technical problems, children that could not sit still enough, or problems with detecting the eyes (e.g., due to problematic glasses). For 35/91 children, signals were lost too often for analysis (in most cases because the child moved outside the view of the eye tracker). We successfully measured eye movements for the entire duration of the horizontal DEM in 36 children (10.2 ± 1.4 years). Horizontal and vertical DEM times in this subpopulation were similar to those of the entire population (Supplementary Table S1 . Most tracking problems occurred in young children. Consequently, there is a slight bias in the data towards older children. The youngest child whose data could be included was 7.2 years old. Using the gaze position signals, we separated the horizontal DEM into different epochs: "number identification epochs" and "return sweep epochs". Number identification epochs were taken from the moment the point of gaze arrived at the beginning of a line until the point of gaze arrived at the end of a line and the eye started moving towards the beginning of another line. Return sweep epochs were taken from the moment the point of gaze started moving from a line ending towards the beginning of another line until the point of gaze arrived at the beginning of another line, with or without intermittent fixations. Figure 2a shows the horizontal displacement of the right eye of a representative child during the horizontal DEM (see Supplementary Fig. S3 for raw eye-tracking data of the same child). Note that, instead of one swift saccade to the beginning of the next line, children often made multiple smaller saccades to reach the beginning of the next line. Figure 2b shows how much time subjects spent on different task elements in relation to the total time they needed to complete the two subtests (uncorrected for naming errors). We first looked at how much time subjects needed for number identification (i.e., all number identification epochs combined) compared to making the return sweeps (i.e., all return sweep epochs combined) during the horizontal DEM. As expected, subjects www.nature.com/scientificreports/ To explore further what made some children more efficient than others, we plotted the median saccade amplitude, the median fixation duration and the number of eye movements made as a function of horizontal DEM time and age. Because developing readers make more corrective saccades during return sweeps than adults 24 , one would expect, for instance, that children with better DEM performance make fewer but larger saccades during the return sweep epochs. Saccade amplitude increased significantly with decreasing horizontal DEM time, both during return sweeps epochs ( The vertical DEM limits the use of return sweeps and horizontal saccades. We wondered how saccade behavior differed between the horizontal and vertical DEM as the test relies on the comparison between these subtests to evaluate oculomotor performance. We therefore determined the relationship between vertical DEM and the oculomotor parameters as analyzed above for horizontal DEM. In this case, we focused on the number identification periods because the vertical DEM only calls for two return sweeps, one per subtest. We found no significant association between vertical DEM time and saccade amplitude, number of eye movements, or number of backwards saccades (|r|< 0. 35 DEM and the Speed Acuity task. To assess the contribution of visual information processing time to the DEM, we compared DEM times to reaction times measured during a Speed Acuity test (SA) in 90/91 children who performed both the DEM and the SA (see Supplementary Table S2 for details). The SA measures visual processing speed for different symbol sizes 21 . In this task, children had to discriminate the orientation of a Landolt "C" as left or right by pressing the corresponding mouse button (Fig. 5a, left). Besides this discrimination task, each child also performed a visual and an auditory detection task (Fig. 5a, right). We compared the DEM performance to mean reaction time during the SA for matched symbol sizes, and found significantly positive correlations between the SA reaction times and performance in the DEM (Fig. 5b; Figure 2 shows that the total fixation time during number identification contributes strongly to horizontal DEM performance. Together with the results in Fig. 5b, this raised the question whether fixation durations correlate with SA performance. Indeed, we found a strong positive correlation between SA and median fixation duration during number identification epochs (Fig. 5c left; www.nature.com/scientificreports/ processing speed, we should see a stronger relationship to performance in a visual detection task than to performance in a similar detection task with an auditory stimulus. However, fixation durations during number identification correlated equally well to performance in the auditory detection task ( Fig. 5d; visual detection: r = 0.58, p < 0.001; auditory detection: r = 0.60, p < 0.001; difference between correlation coefficients, p = 0.90). All p-values provided in Fig. 5b-d remained statistically significant after correction for multiple comparisons.
Discussion
We combined the horizontal and vertical DEM with eye tracking and studied the relation with performance on a visual SA test. Both vertical and horizontal DEM times strongly related to the visual processing speed measured in the SA test as well as the eye fixation durations measured in the DEM. These findings agree with the results of Ayton et al. 10 and Moiroud et al. 20 for the horizontal DEM, and extend the results to the vertical DEM. Additionally, we found that faster performance on the horizontal DEM coincided with fewer, but typically larger saccades. Children who needed more time for the horizontal DEM also made more backwards and vertical saccades in the number identification epochs. By contrast, we did not find a significant relationship between the time children needed for the vertical DEM and the number of vertical or horizontal saccades that they made in this subtest. The median amplitude of saccades made during the number identification epochs of the vertical DEM was also unrelated to the time children needed for that test. However, the time that subjects spent on making saccades was very short in comparison to the time they needed to complete either subtest. Instead, most of the variation in either subtest was linked to variation in the number and duration of fixations, which strongly correlated with the participants' visual processing speed measured in the SA test. We want to emphasize that visual processing speed is probably one of many factors that influence fixation durations in the DEM. Fixation durations include components relevant to saccade planning 25 , and decreasing horizontal saccade latencies in child development 26 could explain variations in fixation duration. Fixation durations are also influenced by attentional, verbal and cognitive variables 27 . The developing visual-verbal naming process, for example, could explain decreased fixation times with age 6 . Note, however, that the correlations between visual processing speed and fixation durations remained significant with age as a covariate. A limitation of our study is that we did not asses the naming process in the absence of eye movements using, for example, a rapid serial visual presentation task 28 . However, performance on such tasks is probably influenced by visual processing speed too, making it difficult to dissect the relative contributions.
The King-Devick (K-D) test is another visual-verbal test used for clinical assessment of saccade behaviour and attention 29,30 . It is similar to the horizontal DEM in design, consisting of eight lines instead of sixteen. Performance on the DEM and K-D is indeed strongly correlated 31 and performance on the K-D test is also positively correlated to fixation duration 32 . Garzia et al. argued, however, that in order to gain knowledge about oculomotor skills, the automaticity of the visual-verbal naming component has to be eliminated 5 . For this reason, they introduced the vertical DEM, assuming that visual-verbal naming automaticity is the dominant factor determining performance on the vertical DEM. The correlation in the time that subjects spent on number identification between the horizontal and vertical DEM as well as the correlations with visual processing speed indeed support the idea that rapid automatic naming skills and visual processing speed play an important role in both tasks 10 .
Interestingly, however, if one discards the time needed for the return sweeps, subjects needed more time naming the numbers in the vertical subtest than in the horizontal subtest. This is inconsistent with the assumption that the vertical DEM accurately reflects the time spent on naming in the horizontal DEM (Table 1, and 6 ). Fixation durations were also significantly longer in the vertical DEM. This latter finding is opposite to the assumptions of Garzia et al. 5 , who thought that in the horizontal subtest "saccade latency is greater because of target unpredictability". How can this be understood? First, we note that all numbers are continuously visible, so the spatial locations of targets are predictable for the duration of the test. Second, in the vertical subtest, the spacing between successive numbers is smaller, which calls, at least in principle, for more precise saccades. The longer fixation durations in the vertical DEM could thus be a consequence of a speed accuracy tradeoff seen in nearly all goal-directed behaviors. In any case, it is known that saccade latencies increase rapidly for targets appearing within 2° of the current fixation 33,34 , and horizontal saccade latencies are significantly shorter than vertical saccade latencies 35 . Lastly, crowding could also play a more prominent role in the vertical DEM compared with the horizontal DEM 36 , thereby impeding the visual verbal naming process and increasing fixation times 6 .
Our finding that the number, direction and amplitude of saccades correlated stronger with the horizontal DEM than the vertical DEM, supports the idea that the horizontal subtest puts a stronger emphasis on saccade behaviour compared with the vertical subtest. Yet, it is incorrect to think that, because of these findings, the comparison between horizontal and vertical DEM times would allow for conclusions regarding oculomotor (dys)function. First, as the purpose of saccades is to rapidly redirect gaze, we found that the contribution of saccade time (the sum of saccade durations) to the total time spent on number identification is small, and almost negligible compared to the overall DEM time. It is not surprising, therefore, that saccade parameters measured under different task conditions did not correlate significantly with horizontal DEM performance 10,19,20 . Second, visual search can take place with or without eye movements [37][38][39] , as humans are capable of shifting attention without redirecting their eyes. The observed differences in saccade behavior could therefore reflect different visual search strategies imposed by the spatial layout of the numbers, rather than physiological differences between horizontal and vertical saccade programming. Indeed, in the horizontal DEM, but not the vertical DEM, neighboring distractors in the direction orthogonal to the reading direction compete for selection with the target number. This explains the larger number of saccades made in the orthogonal direction during horizontal DEM compared to vertical DEM, whereas the number of backwards saccades was comparable. Finally, the comparison between horizontal and vertical DEM to factor out automaticity implicitly assumes that vertical saccades are generally unaffected in subjects with oculomotor problems. The neurophysiological organization of the saccadic www.nature.com/scientificreports/ system does not justify this assumption 40 . Specific lesions in the brain stem and cerebellum can selectively affect horizontal saccades as the circuits controlling the horizontal and vertical components of saccades are largely separated at these levels, but this is not the case at the cortical and subcortical planning stages where saccades are represented in topographically organized maps 41,42 . An important factor for the difference between vertical and horizontal DEM times, and hence to deviations of the DEM ratio from 1, is the time needed for the 15 return sweeps in horizontal DEM. In line with previous research 24 , we found that good performers spent less time on the return sweeps as they made larger saccades with fewer intermittent fixations to accomplish them. The significant contribution of return sweeps to the horizontal DEM time also explains why other studies have found that the horizontal subtest correlates stronger with reading performance than the vertical subtest and the DEM ratio 10 . Finally, we noticed poor, non-significant correlations of the return sweep parameters with age. Because return sweep times contribute to determining the DEM ratio, these findings partly explain why the DEM ratio showed a weaker correlation with age than the horizontal and vertical DEM times.
An advantage of our new stereoscopic eye-tracking methodology is that it simplifies the in vivo calibration to a one-point fixation task 23 . Even so, the success rates were limited, thereby limiting the statistical power of the eye movement analyses. For Pearson's correlations and paired t tests with a large effect size (ρ = 0.5), and a power (1 − β) of 0.80 with a significance level (α) of 0.05, the sample sizes required is 26 and 27 participants, respectively 43,44 . This means that the 28 children for whom we had complete eye movements records in both subtests were enough to reliably detect and reject large effect sizes (with Type I and II errors of α = 0.05 and β = 0.2, respectively) but not medium to small effect sizes (ρ = 0.3 and ρ = 0.1). We did not adjust these results for multiple comparison because we did not want to miss any possible association between the DEM and oculomotor skills. Yet it is possible that some correlations between saccade parameters and vertical DEM performance are too weak to be picked-up in the present study.
Considering that eye-tracking equipment no longer needs to be expensive nor invasive, is the DEM still relevant for clinical diagnostics of oculomotor problems? An advantage of the DEM is its simplicity in administering the test. The comparison between horizontal and vertical DEM provides some indirect clues about oculomotor skills in healthy children, but the results do not warrant conclusions about possible pathology. To accurately diagnose oculomotor problems, abnormal DEM scores should be followed up by more specialized diagnostic measurements [45][46][47][48][49] . That said, horizontal and vertical DEM times are significantly associated with literacy and numeracy scores 13,14 , reading rate 15,16 , and speed of visual processing (this study and 10 ). We therefore believe that the DEM offers a possibility to recognize anomalies in these domains at a relatively young age. For example, children with Cerebral Visual Impairments show delays in the processing of visual information 1,22 , that might also be reflected in reduced performance on the vertical and/or horizontal DEM.
Taken together, our results provide new insights in the age-dependent oculomotor behavior of children during the two subtests of the DEM. Saccade behavior is different between the horizontal and vertical DEM, but these differences do not warrant the conclusion that the DEM can assess saccade skills. Instead, the picture emerges that the oculomotor signature of DEM performance is visual fixation; a process linked to visual-verbal naming skills, visual processing speed, visual search strategies and other cognitive factors that guide fixation.
Methods
Participants. Ninety-one children (9.4 ± 2.0 years) were recruited. Inclusion criteria were age 5 to 12 years old, normal birth weight (> 2500 g), birth at term (> 36 weeks), no perinatal complications, no complaints of slow visual processing, crowded VA of 0.1 logMAR or better, and normal development. Testing occurred at the children's primary schools. Children with glasses wore them during all tests. For details, see 21 .
Informed consent was obtained in writing from the parents of all participants. The study was approved by the local ethics committee Commissie Mensgebonden Onderzoek regio Arnhem-Nijmegen, The Netherlands (protocol NL48708.091.14), and conducted according to the principles of the Declaration of Helsinki.
Developmental eye-movement test (DEM). The DEM (Fig. 1a) was administered at ~ 65 cm. Children were asked to maintain that viewing distance without head restraint.
Children first practiced with a pre-test to familiarize them with the task, and to make sure that they could read numbers. The pre-test was a shortened version of each DEM subtest with randomized ordering of the numbers. Then, children had to name the numbers of DEM A from top to bottom, one column at a time. All numbers appeared on the computer screen as soon as the experimenter pressed the space bar and disappeared when the experimenter pressed the space bar again as soon as the child named the last number. These start and stop moments were recorded by the software. DEM A was followed by DEM B, which is similar to A but with the numbers in a different order.
The numbers of the DEM C array had to be named from left to right, starting at the top left. Horizontal time was taken as the total time to name the first to the last number of DEM C. For the list of applied numbers, see Supplementary www.nature.com/scientificreports/ eyes. The one-point calibration that is required for this method was part of another task 50 . The offline gaze reconstruction combined the asynchronous data from the two cameras into gaze position signals with an average refresh-rate of ~ 500 Hz. Since the system did not only record the point of gaze (POG) on the screen, but also the three-dimensional location of the eyes, we could accurately account for head translations. The spatial accuracy of the resulting eye movement measurements was ~ 0.7°. The spatial resolution was better than 0.2° and the sampleto-sample noise was less than 0.05°.
Speed acuity test (SA).
The speed-acuity test 21 was administered binocularly at 5 m. Each trial consisted of a high-contrast black Landolt-C presented at the center of the computer screen against a white background. Children had to indicate, as quickly and accurately as possible, on which side, right or left, the opening of the C was located by pressing the corresponding mouse button. The stimulus was presented until the child responded. Task difficulty was manipulated by presenting nine optotype sizes (ranging from 0.43 to 1.09 logMAR), each presented 10 times in pseudo random order.
Here, we only included the reaction times to the 0.68 LogMAR stimulus, which was similar to the size of the numbers of the DEM (0.71 LogMAR).
Detection tasks. The children also performed a visual and an auditory detection task to measure the time children needed to respond to a supra-threshold stimulus 21 . In the visual detection task (20 trials), they had to press a mouse button as soon as they saw the visual stimulus (a large high-contrast black letter "O"). In the auditory detection task (20 trials), they had to press the mouse button as soon as they heard the sound (a 500 ms white noise burst of ~ 75 dBA). The eye tracker consisted of two USB 3.0 cameras (Lumenera lt225 NIR, Lumenera Corp., Ottawa, Canada, pixel size 5.5 × 5.5 μm) connected to the stimulus laptop and two 850-nm infrared illuminators (Abus TV6700, ABUS KG, Wetter, Germany) mounted on an optic rail. The lenses with manual focus and diaphragm had a focal length of 16 mm (Navitar NMV-16M23, Navitar Inc, Rochester, NY, USA). Infrared-passing filters (UV/Vis-Cut R-72; Edmund Optics Inc., Barrington, NJ, USA) that passed wavelengths > 720 nm were added on the lenses to block light in the visible spectrum 23 . Procedure. Children first participated in the Freiburg visual acuity test to assess crowded VA (for details, see 21,52 ), followed by SA, both administered digitally at 5 m distance. Subsequently, the computer screen was moved to 65 cm distance to measure eye movements during the DEM. For a small group of children, the DEM was performed on a different day. Test results for all children can be found in Supplementary table S2. Data analyses. The offline analysis was performed and images were created using Matlab (version 2018b).
Equipment.
DEM. Total vertical time was taken as DEM A time plus DEM B time. Time to complete DEM C was adjusted for omissions and additions 53 . Repeating a whole line counts as five addition errors. Skipping one-line counts as two omission errors. The number of errors was determined by adding the number of omissions and additions We used adjusted DEM C times for regression analyses without eye tracking, and raw DEM C times otherwise. The time for DEM A and B was not adjusted for errors because of the limited number of errors made during those tests.
Speed-Acuity. For SA and detection tasks, we computed mean reaction times after removing atypically long or short reaction times. Trials were excluded from the mean if the reaction time deviated more than three times the median absolute deviation (MAD) from the median after discarding reaction times < 0.1 s. Children whose mean reaction time deviated more than three standard deviations from the mean were excluded (excluded: 0/90 in SA, 2/90 in visual task, 1/90 in auditory task). Correlation coefficients were calculated in Matlab using Pearson's correlation coefficient, and Pearson's linear partial correlation with age as a confounding variable. Multiple comparisons were done using False Discovery Rate (FDR) 54 .
Eye-tracking. We used data from the eye with the best tracking signal. Data where signals for both eyes were lost were not included. The sampling rate of the gaze position signals was variable because the two cameras ran asynchronously 23 . Therefore, the data were resampled to a fixed sampling rate of 500 Hz using linear interpolation. Prior to this interpolation, a 13-point median filter was applied, followed by zero-phase filtering with a low-pass Butterworth filter (5th order, 40 Hz cut-off) after resampling. www.nature.com/scientificreports/ Saccades were detected with custom software using velocity and acceleration threshold criteria for saccade onsets and offsets that were all set to 3 times the median absolute deviation (MAD) of the noise (typically around 20°/s and 3000°/s 2 or less). Eye velocity was taken as the vector sum of horizontal and vertical eye velocities (track velocity). The acceleration signal was the time derivative of eye track velocity. Differentiation with respect to time was performed with a 9-point noise-robust differentiator. The software checked if eye positions measured in a 20 ms time window before saccade onset were significantly different from the eye positions measured in a 20 ms time window after saccade offset. All saccade markings were visually inspected and corrected if necessary.
Horizontal saccades were all saccades with a direction below ± 45° from the horizontal meridian. All other saccades were included as vertical saccades. We did not distinguish oblique saccades. Fixation duration was determined from the time between the offset of one saccade and the onset of the next saccade. Fixation time was defined as the sum of all fixation durations in a given time window. Saccade time was defined as the summed durations of all saccades in a given time window. Occasionally, the eye-tracking signal was briefly lost due to eye blinks, which lead to a relatively long, and incorrect, fixation duration between two saccades. Therefore, fixation durations > 3 * MAD from the median were removed from the dataset.
Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request. www.nature.com/scientificreports/ Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/. | 6,833.4 | 2021-01-13T00:00:00.000 | [
"Psychology",
"Biology"
] |
Multi-modal LSTM network for anomaly prediction in piston engine aircraft
An aircraft is a highly intricate system that features numerous subsystems, assemblies, and individual components for which regular maintenance is inevitable. The operational efficiency of an aircraft can be maximised, and its maintenance needs can be reduced using an effective yet automatic AI-based health monitoring systems which are more efficient as compared to designing and constructing expensive and harder to operate engine testbeds. It has been observed that aircraft engine anomalies such as undergoing flameouts can occur due to the rapid change in the temperature of the engine. Engine oil temperature and cylinder head temperature, two measures connected to this issue, might be affected differently depending on flight modes and operational conditions which in turn hamper AI-based algorithms to yield accurate prediction on engine failures. In general, previous studies lack comprehensive analysis on anomaly prediction in piston engine aircraft using modern machine learning solutions. Furthermore, abrupt variation in aircraft sensors' data and noise result in either overfitting or unfavourable performance by such techniques. This work aims at studying conventional machine learning and deep learning models to foretell the possibility of engine flameout using engine oil and cylinder head temperatures of a widely used Textron Lycoming IO-540 six-cylinder piston engine. This is achieved through pre-processing the data extracted from the aircraft's real-time flight data recorder followed by prediction using specially designed multi-modal regularised Long Short-Term Memory network to enhance generalisation and avoid overfitting on highly variable data. The proposed architecture yields improved results with root mean square error of 0.55 and 3.20 on cylinder head and engine oil temperatures respectively averaged over three case studies of five different flights. These scores are significantly better i.e., up to 84% as compared to other popular machine learning predictive approaches including Random Forest, Decision Tree Regression, Artificial Neural Networks and vanilla Long Short-Term Memory networks. Through performance evaluation, it can be established that the proposed system is capable of predicting engine flameout 2 minutes ahead and is suitable for integration with the software system of aircraft's engine control unit.
An aircraft is a highly intricate system that features numerous subsystems, assemblies, and individual components for which regular maintenance is inevitable.The operational efficiency of an aircraft can be maximised, and its maintenance needs can be reduced using an effective yet automatic AI-based health monitoring systems which are more efficient as compared to designing and constructing expensive and harder to operate engine testbeds.It has been observed that aircraft engine anomalies such as undergoing flameouts can occur due to the rapid change in the temperature of the engine.Engine oil temperature and cylinder head temperature, two measures connected to this issue, might be affected differently depending on flight modes and operational conditions which in turn hamper AI-based algorithms to yield accurate prediction on engine failures.In general, previous studies lack comprehensive analysis on anomaly prediction in piston engine aircraft using modern machine learning solutions.Furthermore, abrupt variation in aircraft sensors' data and noise result in either overfitting or unfavourable performance by such techniques.This work aims at studying conventional machine learning and deep learning models to foretell the possibility of engine flameout using engine oil and cylinder head temperatures of a widely used Textron Lycoming IO-540 six-cylinder piston engine.This is achieved through pre-processing the data extracted from the aircraft's real-time flight data recorder followed by prediction using specially designed multi-modal regularised Long Short-Term Memory network to enhance generalisation and avoid overfitting on highly variable data.The proposed architecture yields improved results with root mean square error of 0.55 and 3.20 on cylinder head and engine oil temperatures respectively averaged over three case studies of five different flights.These scores are significantly better i.e., up to 84% as compared to other popular machine learning predictive approaches including Random Forest, Decision Tree Regression, Artificial Neural Networks and vanilla Long Short-Term Memory networks.Through performance evaluation, it can be established that the proposed system is capable of predicting engine flameout 2 minutes ahead and is suitable for integration with the software system of aircraft's engine control unit.
Introduction
Aircraft maintenance, fuel costs, increase in air traffic, competition, recession, air pollution, security, and operational difficulties are some of the challenges that the aviation industry continues to face despite its rapid expansion [1].Aircraft maintenance accounts for approximately 40% of the total operating costs of aviation systems.Having the right health monitoring in place for a given flying service will enhance operational efficiency, while decreasing the need for aircraft repair [2].
Aircraft engines must operate within specific physical constraints due to their costly maintenance [3].Although modern engines are equipped with more control variables and sensors, they remain susceptible to failures beyond regular wear and tear [4,5].Detecting minor issues early can prevent them from escalating into major problems and potentially avoid accidents [6].The frequency of technical difficulties and accidents is directly linked to the rate of aircraft manufacturing.
The Engine Test Bed (ETB) [7] simulates actual flights for real-time estimation and optimisation of flight characteristics.Flight parameter optimisation systems consist of various Line Replacement Units (LRUs) or equipment [8].In modern commercial aircraft, extensive data is continuously recorded, encompassing weather conditions, pilot preferences, and system health parameters [9].Data mining methods are applied to these vast datasets to discover valuable patterns.The primary goal of evaluating such large datasets is to enhance the aircraft's overall health and reduce airline operating costs.Alarm systems in planes are capable of promptly detecting life-threatening issues [10].
Aircraft health parameters and their fluctuations, measured through sensor readings, are gaining attention for improved estimation systems.Effective optimisation of engine parameters can lead to reduced life cycle expenses, engine overhaul costs, and enhanced engine safety [11,12].Conventional aviation engine optimisation involves using an ETB facility to analyse and improve the engine's performance in flight [13].However, ETB is a laborious solution that can only estimate and optimise aircraft engines on the ground, despite their operational differences in the air.Maintenance procedures in the current literature are typically categorised as follows, • Run-to-failure (R2F) maintenance: A simplest yet expensive maintenance approach, this is after-breakdown maintenance and repair procedure which usually results in high cost of addressing the issue after engine component(s) failure.• Preventive maintenance (PvM): Preventative maintenance, often known as scheduled maintenance prevents failures but may lead to unnecessary maintenance exercise.• Predictive maintenance (PdM): In this approach, maintenance is performed as needed, and prediction methods are utilised in planning and scheduling systems to determine the appropriate timing for actions.PdM systems employ ad-hoc or statistical inference-derived health factors.
Usually, statistical, AI, or model-based systems are used to detect and predict machinery health.Model-based methods require mechanical and theoretical knowledge of the equipment, while statistical approaches rely on mathematical backgrounds, leading to the increasing use of AI in PdM applications [14].AI outperforms statistical methods in predicting equipment failure due to its ability to learn patterns from data and conduct predictive analysis [15].Prediction algorithms based on machine learning and deep learning are capable of detecting latent data correlations and managing high-dimensional and multivariate data in complex, dynamic scenarios [16].This study explores how can these algorithms be effectively utilised to predict engine flameout in piston engine aircraft, specifically using engine oil and cylinder head temperatures from a Textron Lycoming IO-540 six-cylinder piston engine widely used in small private and trainer aircraft?Moreover, a thorough empirical assessment of contemporary machine learning and deep learning methods becomes essential to validate their applicability, particularly when dealing with limited data and substantial variations in flight patterns and, consequently, aircraft sensor readings in real-world scenarios.The contributions of this study are summarised below.
• Arrange mission data pertaining to various flight phases after organising and extracting from an aircraft's flight data recorder (FDR).• Perform a thorough pre-processing of the FDR data for cylinder head temperature and engine oil temperature.
• Analyse popular machine learning prediction approaches including Random Forests (RF), Decision Tree Regression (DTR) and Artificial Neural Networks (ANN).• Proposing a stacked Long Short-Term Memory (LSTM) network with regularised multi-modal input scheme for accurate prediction and mitigating the challenge of unreliable performance on our data exhibiting high variation.
Related work
This section summarises research, especially related to machine learning and deep learning on estimating an aircraft engine's health metrics including remaining usable life (RUL), a measure which can be used to estimate ideal time for maintenance or service [17].Mostly, the literature discusses turbojets, turbofans, turboprops, and turbo-shafts engines whereas piston engines, although widely used, are relatively less studied.The features mostly used to predict RUL and flameouts include Exhaust Gas Temperature (EGT), Cylinder Head Temperature (CHT), Engine Oil Temperature (EOT), Low Pressure Compressor (LPC), High Pressure Compressor (HPC), Low Pressure Turbine (LPT), High Pressure Turbine (HPT), N1 and N2 speed, Angle of Attack, Engine RPMs, Pitch, Roll, and Vertical Acceleration.These features are acquired using various sensors mounted in aircraft's electronic and control systems.
The prediction of RUL and flameouts due to anomalous behaviour in sensors' data is helpful in predictive maintenance of aircraft engines.In pursue of data-driven approaches, ensemble regression approach is proposed in [18] to estimate RUL where random forest and gradient boosting regression models are trained using FEMTO ball bearings data (IEEE PHM Data Challenge 2012).Likewise, in [19], a double-CNN architecture for accurate RUL prediction is presented which brings significant improvements to the prediction reliability.The recommended framework relies on feature extraction to preserve and utilise information regarding aircraft's engine health.LSTM and Gated Recurrent Units (GRU)-based Recurrent Neural Networks (RNN) are used in [20] for estimating the approximate service period of a turbofan engine with over 90% accuracy on an engine degradation simulation dataset.A similar outcome is presented for PdM using LSTM and GRU in [21] and bidirectional LSTM in [22].The utilisation of RNN in aircraft RUL estimation is not new though as previous attempts are successful in proving the prediction capabilities of RNNs [23].RF regression is employed to predict aircraft component failure as PdM in [24,25] by collecting data from various engine sensors.Similarly, a combination of CNN and LSTM is used in [26] to predict RUL using a post flight data achieving an accuracy of 99% on a selected flight dataset.In [27] 1D CNN with Monte Carlo dropout is used in a reinforcement learning manner for estimating turbofan engine RUL with impressive success rate and dropping maintenance cost by 29.3% validating the similar previous study [28,29] which utilise related deep CNNs.
Turbojet engines are also studied in detail for PdM as [26,30,31] employ LSTM and Multiple Instance Regression (MIR) on engine degradation and post-flight datasets with promising results.Similarly, LSTM with Support Vector Machines (SVM) presents a hybrid model for RUL prediction on NASA's aero-engine dataset [32].On the other hand, [33] proposes RUL estimation using multiple machine learning modules including SVM, K-Nearest Neighbour (KNN), RF, and Analysis of Variance (ANOVA) statistical approaches.In [34] a combination of 1D CNN, autoencoder, and bidirectional GRU network is used on time-series data of 60 turbojet commercial aircraft in a pursue of detecting rare engine failures.This approach primarily addresses data imbalance to achieve favourable performance.DTR and ANN are among the popular approaches used in PdM for RUL in gas turbine engines as demonstrated by [35,36] where a simulated dataset is used to prove the proposed algorithm's efficacy.In these studies of gas turbine engines, LPC, HPC, LPT, and HPT are the most commonly used features along with a few others for better prediction.
The comprehensive literature review found that on average only 3.7 articles per year appear to address PdM of aircraft engines using machine learning and deep learning.Furthermore, as mentioned earlier, piston engines are relatively less studied in the pursue of PdM and due to their simpler structural arrangement, challenges like flameout occur more frequently.Therefore, this work aims to target piston engines with CHT and EOT as features for predicting such anomalies.As the Textron Lycoming IO-540 engine is used in our study, the aircraft that come with this engine e.g., Cessna 188, 206 and several other trainer aircraft lack ejection seat system making it imperative for the pilot to initiate landing protocol or corrective measures in case of engine flameout by varying air-fuel mixture ratio, throttle and primer.In terms of PdM, the CHT and EOT long-term data can be used to detect malfunctioned engine components for replacement hence playing critical role in estimating engine's RUL.
In the context of various studies discussed earlier, utilising either traditional machine learning or contemporary deep architectures, a research gap emerges with two crucial aspects.Firstly, cutting-edge deep learning techniques such as CNN, LSTM, and GRU networks, characterised by intricate and highly nonlinear mathematical structures, sometimes struggle to demonstrate robust generalisation capabilities when confronted with related yet diverse and small data exhibiting high variability [37].Despite achieving favourable outcomes on specific datasets, their performance on data representing different flight patterns often compromises anomaly prediction accuracy due to overfitting and a lack of data-specific regularisation constraints.These architectures usually apply data augmentation approached to extend data samples and avoid overfitting.However, aircraft data involving various sensors and natural flight patterns should not be augmented with random transformations to the original data as doing so may undermine the actual sensor behaviour leading to an anomaly.On the other hand, commonly employed traditional machine learning approaches like DTR, RF, and ANNs typically face challenges in extracting complex sequential features from data, hindering their ability to effectively learn task-specific information for aircraft anomaly prediction, particularly beyond a certain time span [38].The proposed scheme focuses on resolving these critical issues, aiming to attain high accuracy in predicting piston engine flameout within a small yet highly variable dataset of real flights.This is accomplished by extracting meaningful patterns from the data through a residual LSTM setup, complemented by data signal pre-processing for effective regularisation to avoid overfitting.
Materials and methods
This section addresses the whole pipeline comprising data collection, data pre-processing, and the proposed deep architecture for engine flameout prediction.As a result, the analytical study is based on full flights as well as cruise phases to study various prediction algorithms of machine learning and deep learning.In this regard, the proposed regularised multi-modal LSTM model is compared with popular DTR, RF, standard LSTM and ANN.The nomenclature of symbols used in the equations for explaining data pre-processing and the proposed architecture is given in Table 1.
Dataset and pre-processing
The dataset used in experimentation is collected from FDR of an aircraft housing Textron Lycoming IO-540 six-cylinder piston engine.The data is acquired in a customised protocol of Garmin ™ as their equipment is used.With the help of GPS altitude, the flight phases with vital sensory data are extracted, while removing all the other unnecessary parameters of the mission profiles.Consequently, overall dataset comprises 11 flights out of which 5 flights are kept for training different algorithms, 5 are kept for testing, and remaining one for validation of optimum parametric settings in training those algorithms.The duration of flights ranges from 30 to 65 minutes.The flights reserved for training come without any flameout anomaly therefore, we call them healthy flights.
In contrast, flights in test dataset experience flameouts due to EOT and CHT exceeding their flameout threshold of 210 and 450 respectively for the engine type under study.
After extracting and visualising FDR data, three case studies are presented to train all the algorithms including the proposed.For Case Study-1 all flights are combined to form one long duration flight.In other words, the flight data is generated by concatenating 4 out of 5 flights reserved for training to generate a bigger training dataset.For Case Study-2 cruise phases from these 4 flights are isolated and then combined and for Case Study-3 data, the fifth standalone flight showing a complete mission profile with multiple cruise phases at different altitudes is used.This data segregation of flight phases is achieved using GPS altitude monitoring.All the three case studies are shown in Fig. 1.The case-wise study helps us in judging the generalisation potential of learning algorithms to favourably perform on test dataset when trained on three different scenarios including flights with takeoff and landing phases and flights with cruise-only phases.It is worth mentioning that test dataset comes naturally without data segregation except the case when models are trained in Case Study-2, the test data of all five flights also include only cruise phases, the reason of which is explained in Results and Discussion section.
It is important to note that the data from FDR is recorded at 1 Hz sampling frequency i.e., the samples in Fig. 1 corresponds to seconds in time.Since features like control surface motion, temperature, vibration, and combat sensor data are highly variable and dependent on flight trend, predicting an engine flameout is crucial.To improve model performance, selecting appropriate attributes to describe various flight stages is important as incorporating irrelevant variables into the training process can harm model performance.Feature selection enhances predictor performance, reduces overfitting and redundant data, shortens training time, and improves model quality.To achieve this, correlation method is used for feature selection as independent and dependent features have a linear relationship.This work focuses on multivariate regression and uses the correlation technique using (1) to select attributes for the different flight phases.EOT and CHT are the dependent variables in (1) presented as y, while the independent features are represented by x.We have chosen 15 most relevant features out of the sample size n of 31, including highly correlated features such as EOT and CHT.Relevant feature selection, as the final step in data preparation, helps eliminating noise and unnecessary data which supports better training.
In each of the three case studies, a selection of 15 features is made to represent the outputs of various sensors on the aircraft.These features include Cylinder Head Temperatures (CHT) for all six cylinders, True Air Speed (TAS), Indicated Air Speed (IAS), Ground Speed (GndSpd), Revolution Per Minute (RPM), Engine Fuel Flow (Fflow), Altitude GPS (ALtGPS), Altitude Mean Sea Level (ALTMSL), Altitude Indicated (ALTInd), and Engine Oil Temperature (EOT).These features are chosen as they exhibit dependency with each other.The combined data of all case studies are analysed and a correlation heat map is generated which indicates direct or indirect relation among features as shown in Fig. 2.These 15 features in the form of time-series data are employed for training all algorithms to detect engine flameout.In case of standard and the proposed multi-modal LSTM network, all those 15 features' timeseries data are taken as separate input channels (analogous to CNN's input channels).The LSTM layers then process each channel independently, allowing the network to capture the relationships and patterns among features that may help predict behaviour of CHT and EOT at the output.From Fig. 2 it is evident that EOT shows high correlation with CHT5, which in turns is directly related to other five CHTs.For this study, the temperature of the sixth cylinder (CHT6) and engine oil temperature give enough information to predict flameout.After feature selection, scaling, normalisation, and standardisation of data is done as pre-processing step using ( 2) and ( 3) to get the normalised data .Equation ( 3) limits the data in the range [0, 1].This pre-processing assumes data to be Gaussian distributed.[39].To sum up, for Case Study-1, all the waveforms of 15 features are concatenated separately (to make 15 long waveforms) with their values representing take-off, cruise and landing phases of four different flights.Similarly, in Case Study-2, only feature values of cruise phases of these four flights are concatenated, making 15 different waveforms of cruise phases, while Case Study-3 is a single long flight with take-off, landing and multiple cruise phases making 15 different waveforms for this single flight.
Deep architecture
Here, a deep multi-modal regularised LSTM network is presented for engine flameout prediction using waveforms of data features extracted from FDR.As illustrated in Fig. 3, the proposed network is based on LSTM layers arranged in two branches to extract waveform patterns from two separate input representations i.e., raw time-series waveforms of data features and their smoothed versions after applying Singular Spectrum Analysis (SSA) [40].The SSA-applied waveforms as inputs to the deep architecture act as regularisation constraint to alleviate the impulsive and noise-like behaviour of time-series data.This in turn supports the prediction capacity of LSTM architecture by avoiding overfitting on the raw signal fluctuation.As oppose to other approaches like movingaverage or exponential smoothing, SSA is a data-driven method that does not rely on preset assumptions or parameters.It extracts the underlying patterns and components directly from the data without imposing any specific mathematical model.This makes SSA more flexible and adaptable to different types of time series.In the proposed network, both branches comprise three LSTM layers each, where a residual LSTM block of two layers precedes the third layer.Outputs of the last LSTM layers from both branches are combined and passed through the dense layers to achieve predicted output.
Assuming input and output sequences for the proposed network for each data feature to be = [(1), (2), … , ( )] and = [ (1), (2)] respectively.For = 1, 2, the input 1 is raw feature time-series data, while 2 is its SSA counterpart yielding as output of the network predicting samples of EOT and CHT6 waveforms.Using weight matrices , , , , recurrent weight matrices , , , , and bias vectors , , , , the inference of the proposed network for layer = 1, … , and branch is given in ( 4)- (12), Input gate at time : Forget gate at time : Cell state at time : Output gate at time : Hidden state at time : In the proposed network, the inputs to the last LSTM layers ( = ) of each branch are where, () () represent hidden states of residual blocks.Now, for two branches ( = 1, 2) of the proposed network, The network output is In practice, = 4, = 1 and = 2 for experimentation reported here.To simplify, mathematical expressions for dropout and dense layers are dropped (see Fig. 3).
Results and discussion
The proposed architecture takes time-series training data comprising 15 features as input arranged as 120 s sliding time-window.In other words, a group of 120 s time series sequences are processed at the input layer at a time and each next sequence is 1 s shifted in a sliding window as shown in Fig. 4. The raw time series data and its smoothed SSA version for one of 15 features is shown in Fig. 5 for illustration.As previously mentioned, the training data is organised into three case studies to carefully evaluate the performance of prediction algorithms, including the proposed one.Each algorithm is trained independently using the training data specific to each case study.
For Case Study-1, the training dataset consists of the concatenation of data from four flights to form a larger dataset of 13, 657 × 120 × 15 dimension (observations × time-steps × feature dimension).Similarly, the data dimensions are 5, 396 × 120 × 15 and 5, 907 × 120 × 15 for Case Study-2 and Case Study-3 respectively.For training data, as illustrated in Fig. 4, observations are extracted using sliding windows of 120 s (two minutes) duration while test data is arranged using non-overlapping windows.The prediction on test data is made in a special manner to estimate CHT6 and EOT values 120 s ahead.To achieve this, a test data observation of 120 s length (120 samples) is applied as input to the prediction algorithms including the proposed one and sample value at 121 s is predicted.In the next iteration, the predicted sample is included in the input sequence, while the first sample is excluded.This sliding window prediction operation is continued till 240 ℎ sample for each observation in test data (shown in blue in Fig. 4) and must be completed before the arrival of the first sample of next observation.The flavours of algorithms including ANN, DTR, and RF are taken from [36] and [33].The ANN is a 7-layer fully connected neural network with 20, 15, 10, 8, 5, 4, and 3 neurons in high accuracy on test data.The results are generally inferior to the proposed regularised multi-modal LSTM network but better than DTR, RF, and ANN.Typically, LSTM-based recurrent neural networks stand out as state-of-the-art in predicting anomalies in aircraft engines, especially for turbojet and turboprop engines.Surprisingly, these advanced networks have not been applied to piston engine aircraft.This research delves into their potential for predicting flameout in piston engines, serving as the primary motivation for the proposed LSTM architecture.
The graphs and evaluation scores indicate that in Case Study-1, ANN, DTR, and RF models, in general, fail to predict the takeoff and landing phases of flights while focusing on predicting the cruise phase.This may be due to the reason that the takeoff and landing phases of aircraft in manual controlling vary for each flight, causing the models to consider these phases as outliers and unable to learn sequence pattern.Since the cruise phase is typically the longest phase of an aircraft's mission and the sensor data during this phase exhibits less variation, the prediction models prioritise predicting this phase.To effectively capture the takeoff and landing phases, a comprehensive training dataset specific to these phases is required.Due to limited data in these evaluations, ANN, DTR, and RF models result in high RMSE and MAE errors and lower 2 scores on the test data flights showing inability to yield true sequence trends.In Case Study-2, for the training data, the cruise phases are extracted from the dataset used in Case Study-1.While testing against models trained on Case Study-2, test data for all 5 flights also comprise only cruise phases.The reason is to clearly see the performances when training and testing is done only on the cruise phases.As expected, compared with the Case Study-1, the graphs and evaluation scores in Case Study-2 show reduction in the RMSE and MAE values for all prediction models with the proposed model surpassing the rest.However, for CHT6, due to very high ratio of RMSE to the variance of ground truth data relative to the same ratio in Case Study-1 and 3, the 2 value is very small.In this scenario, the DTR, RF and ANN predictions are more accurate on individual data points, leading to lower RMSE and MAE values.However, the model might not be capturing the overall variability in the data as effectively as in Case Study-1 and 3, resulting in a smaller 2 value.
In Case Study-3, a dataset with a single long and a good mission profile is selected for the training.RF and ANN performance in this case study is similar where ANN outperforms RF in predicting CHT6 while lags in the case of EOT.This is due to behaviour of ANN when training data is very limited and network overfits to the dominant CHT6 values which are always higher than EOT.The proposed model's results remain favourable in this case study too.
To sum up, CHT6 and EOT values are predicted two minutes ahead using ANN, DTR, RF, and the proposed system.In a separate experiment not reported here, prediction window size was increased up to five minutes but very sharp deterioration in prediction accuracy by all the algorithms was observed.Although two minutes still pose a challenging situation for a pilot to carry out counterflameout measures, it is still a reasonable time to avoid flameouts in a simpler piston engine as compared to jet engines.Indeed, pilot's experience and situation awareness play vital role in such scenarios.In another set of experiments not reported here, we cross evaluated the case studies i.e., we took models trained on Case Study-1 and 3 and tested on the test data of Case Study-2, which only comprises cruise phases.The performances of models trained on the larger data of Case Study-1 outperformed the models trained on the data of Case Study-3.
In this investigation, various prediction algorithms are employed, including the proposed one, to anticipate flameout events two minutes in advance using CHT6 and EOT.All the algorithms, including the suggested architecture, exhibit diminishing performance when forecasting beyond the 2-minute mark, and accuracy sharply declines after 5 minutes, rendering these systems impractical for predictive analysis and maintenance.As indicated in the results, the predicted CHT6 and EOT waveforms are integrated into a recursive prediction process, where an initial input of 120 samples is utilised to estimate the subsequent 120 samples.Notably, small prediction errors, when accumulated during the estimation of samples beyond 120 seconds, contribute to an overall deterioration in prediction accuracy.
Upon careful examination, it is determined that, with a 120 s sliding time window sequence, the proposed flameout prediction pipeline achieves inference completion for predicting the next 120 seconds of CHT6 and EOT in 0.21 seconds on the Nvidia ® Titan X GPU and 0.59 seconds on Jetson AGX Xavier, respectively.This makes the system well-suited for real-time applications on edge computing platforms and its potential deployment on the aircraft embedded system under the DO-178C standard titled "Software Consideration in Airborne Systems and Equipment Certification".Additionally, considerations for software criticality levels (DAL A to E) and the qualification of software development tools under DO-330 must be taken into account.
Conclusion
In this study, a multi-modal regularised LSTM architecture is introduced for the prediction of engine oil temperature and cylinder head temperature in the Textron Lycoming IO-540 six-cylinder piston engine.The forecasting of these features is deemed crucial for the anticipation of anomalies such as engine flameout, providing pilots with the necessary time to manipulate air-fuel mixture, throttle, or primer settings.Additionally, this predictive model is considered a valuable tool for off-air predictive maintenance of Bias vector of Long Short-Term Memory layer Cell state of Long Short-Term Memory network Forget gate of Long Short-Term Memory network cell Hidden state of Long Short-Term Memory network cell Total number of layers in Long Short-Term Memory network Number of data samples Output gate of Long Short-Term Memory network cell and output sequences of the proposed network Recurrent weight matrix of Long Short-Term Memory network Weight matrix of Long Short-Term Memory network Independent data variable in correlation measure Input sequence to Long Short-Term Memory network Dependent data variable in correlation measure Network output of Long Short-Term Memory network Data sample normalised with mean and standard deviation Greek Symbols First layer of residual/skip connection Second layer of residual/skip connection Mean Sigmoid activation function in Long Short-Term Memory network Subscripts Long Short-Term Memory Network branch number Superscripts Same as Same as
Fig. 1 .
Fig. 1.Case Studies: From left to right, concatenation of four complete flights (Case Study-1), cruise phases (Case Study-2) and one complete multi-altitude cruise flight (Case Study-3).Altitude is given in mean sea level (MSL).
Fig. 5 .
Fig. 5. Illustration of raw time-series and its smoothed version using SSA.
Fig. 6 .
Fig. 6.Prediction results on test Dataset-1 with (a), (b), (c), and (d) represent outcomes of DTR, RF, ANN, and the proposed method respectively for the three case studies each.
Fig. 7 .
Fig. 7. Prediction results on test Dataset-2 with (a), (b), (c), and (d) represent outcomes of DTR, RF, ANN, and the proposed method respectively for the three case studies each.
Fig. 10 .
Fig. 10.Prediction results on test Dataset-5 with (a), (b), (c), and (d) represent outcomes of DTR, RF, ANN, and the proposed method respectively for the three case studies each.
Table 1
Nomenclature of Symbols.
Table 2
Performance evaluation averaged on all five test datasets for three case studies. | 7,219.8 | 2024-01-01T00:00:00.000 | [
"Engineering",
"Computer Science"
] |
Three level atom optics via the tunneling interaction
Three level atom optics (TLAO) is introduced as a simple, efficient and robust method to coherently manipulate and transport neutral atoms. The tunneling interaction among three trapped states allows to realize the spatial analog of the stimulated Raman adiabatic passage (STIRAP), coherent population trapping (CPT), and electromagnetically induced transparency (EIT) techniques. We investigate a particular implementation in optical microtrap arrays and show that under realistic parameters the coherent manipulation and transfer of neutral atoms among dipole traps could be realized in the millisecond range.
The coherent coupling between two orthogonal states of a quantum system gives rise to oscillations of their probability amplitudes such as the Rabi oscillations of a two-level atom interacting with a laser field. When three instead of two levels are considered, the interaction gives rise to a much richer phenomenology. A clear example is the electric dipole interaction between a three-level atom and two laser modes, where a large number of techniques have been proposed and reported, such as the stimulated Raman adiabatic passage (STIRAP) method used to produce a complete population transfer between two internal quantum states of an atom or molecule [1], the modification of the optical properties of a medium by means of coherent population trapping (CPT) [2], and electromagnetically induced transparency (EIT) [3] phenomena. All these three-level optics (TLO) techniques have been intensively studied with applications ranging from quantum control of atoms and molecules [1,4], laser cooling [5], and slowing down light to a few meters per second [6] to non-linear optics with few photons [7].
In this letter we propose several novel techniques for the coherent manipulation of atoms among trapped states coupled via tunneling. To illustrate the basic idea let us start by considering two well separated dipole traps and one single atom in, say, the left trap. As soon as the two traps are approached and tunneling takes place, the probability amplitude for the atom to be in the left (right) trap oscillates in a Rabi-type fashion resembling the coupling of a two-level atom to a coherent field. This tunneling induced oscillation between the two traps can be used to coherently transfer atoms between traps and, in fact, it allows for a simple realization of quantum computation [8]. However, this two-level technique is not very robust under variations of the system parameters and requires precise control of distance and timing. We will introduce here a set of tools analogous to the TLO techniques to efficiently and coherently manipulate and move atoms among traps. The basic elements will be three traps and a single atom, and the atomic external degrees of freedom will be controlled through the variation of the distance between each two traps. The proposed techniques do not need an accurate experimental control of the system parameters and they will be named is the separation between left and middle (middle and right) traps. In the limit of a large separation |n L , |n M , |n R are the nth vibrational energy eigenstates of the corresponding single trap potentials.
three level atom optics (TLAO) techniques. We will consider here arrays of optical microtraps where the dipole force of a red detuned laser field is used to store neutral atoms in each of the foci of a set of microlenses [9]. We will make use of two specific features of these arrays [10]: the possibility of individual addressing each trap and detecting whether a trap is occupied, and the independent displacement of columns or rows of microtraps. We assume here that we are able to initially store none or one atom per trap at will, as has been reported in single dipole traps [11] and in optical lattices [12]. Although we require only three traps the use of columns of traps has the advantage of doing several experiments in parallel.
The three in-line dipole traps are modeled as three piece-wise harmonic potentials of frequency ω x . and the neutral atom is assumed to be in the ground vibrational state of the left trap initially, while the other two traps are empty (Fig. 1). For simplicity the temporal evolution of the distance between each two traps has been modeled with a cosine function truncated at the minimum separation. Then, the approach and eventual separation of left and middle (middle and right) traps takes a time t , and t delay ωx = 60. (c) Transfer efficiency from |0 L to |0 R as a function of the time delay between the two approaching processes.
system is composed of the vibrational ground states of all three traps, i.e., |0 L , |0 M , and |0 R , and the strength of the interaction between each two vibrational ground states is given (in the absence of the third trap) by the following tunneling "Rabi" frequency [13]: where αd is the trap separation, and α −1 ≡ h/mω x with m denoting the mass of the neutral atom. erf(.) is the error function. The temporal shaping of Ω is realized by controlling the time dependence of d(t). While Eq. (1) is useful to explore the analogies between TLO and TLAO, an exact treatment accounting also for couplings to excited vibrational states requires the integration of the Schrödinger equation. In what follows we will numerically integrate the 1D, and, eventually, the 2D Schrödinger equation to simulate the dynamics of the neutral atom in the three-trap potential.
A robust method to coherently move atoms among traps consists in extending the STIRAP technique [1] to atom optics by using the tunneling interaction. The basic idea is to use the fact that one of the three eigenstates of the three level system involves only the ground states of the two extreme traps: where the mixing angle Θ is defined as tan Θ ≡ Ω LM /Ω MR with Ω LM (Ω MR ) denoting the tunneling "Rabi" frequency between left and middle (middle and right) traps. Following Eq. (2) it is possible to transfer the atom from |0 L to |0 R by adiabatically varying the mixing angle from 0 • to 90 • , which means to approach and separate first the right trap to the middle one and, with an appropriate delay, the left trap to the middle one ( Fig. 2(a)). This counterintuitive sequence moves the atom directly from |0 L to |0 R with an almost negligible probability amplitude to be in the middle trap ground state ( Fig. 2(b)). The STIRAP signature is shown in Fig. 2(c), where the transfer efficiency from |0 L to |0 R is shown as a function of the time delay between the two approaching processes. The plateau near the optimal delay indicates the robustness of the transfer process which also is very robust under variations of the tunneling parameters, i.e., of the maximum and minimum trap separation d max and d min , and of t r and t i for each of the processes, provided that adiabaticity is maintained. Additionally, the approaching sequence can be modified to create spatial superposition states with maximum atomic coherence, i.e., with |c 0L c * 0R | = 1/2, c 0L (c 0R ) being the probability amplitude to be in state |0 L (|0 R ). The basic idea is to adiabatically following state (2) to tω x = 600). The resulting state is the spatial equivalent to the well known dark state arising in the CPT technique [2]. To prove that this state is dark, i.e., that it can be decoupled from the tunneling interaction, we approach and separate simultaneously the two extreme traps to the middle one (see Figs. 3(a) and 3(b) from tω x = 600 up to the end). Clearly, the atom remains in the dark state in spite of the tunneling interaction. Superposition dark states are very sensitive to dephasing [2] which means that they could be used in dipole trap systems to measure experimental imperfections such as uncorrelated shaking in the traps position and/or intensity fluctuations of the trapping lasers. Also, this robust coherent splitting of the atomic wave function into two half pieces together with their individual manipulation anticipates applications in atomic interferometry.
Finally, it is also possible to extend the EIT technique [3] to the atom optics case. The basic idea of EIT is to convert a medium that is opaque to a field resonant with a certain internal transition into being transparent by applying an intense driving field to an adjacent transition. In the three-trap system we will inhibit the transition from |0 L to |0 M by driving the transition |0 M ↔ |0 R via the tunneling interaction. Figs. 4(a) and 4(b) show the inhibition of the |0 L to |0 M transition in spite of the proximity between left and middle traps. As in the STI-RAP case, the plateau near the optimum delay evidences the robustness of the transition cancellation (Fig. 4(c)). This atom optics EIT technique could create conditional phase shifts with no change in the state of the system. Fig. 5 shows the results of a 2D numerical integration of the Schrödinger equation that summarizes the previously discussed ground state TLAO techniques. It is worth to note that these techniques can be also applied to excited states, which relaxes the cooling requirements for the experimental setup. In Fig. 6 two examples for the three level system consisting of the first excited vibrational states of each trap are considered: (a) the transfer of an atom from |1 L to |1 R via the STIRAP technique, and (b) the 50% coherent splitting of the atomic wavefunction between the left and the middle trap (see the plateau around tω x = 180). This effect, which is different from CPT and requires a combination of adiabatic and diabatic processes [15], is possible through a complicated variation of the dressed level structure of the first excited states when approaching the traps.
There are two important practical points for the implementation of the TLAO techniques in optical microtrap arrays: (i) the trapping frequencies must be the same for all microtraps; and (ii) the approaching process has to be adiabatic. The use of a single laser that illuminates simultaneously all microlenses assures the identity of all microtraps even under intensity fluctuations of the laser. In particular, typical trapping frequencies for microtrap arrays of 87 Rb atoms are 10 5 -10 6 s −1 in the transverse directions and 10 4 -10 5 s −1 along the laser beam direction [9, 10] which means that the traps can be adiabatically approached in the millisecond range or even faster by using optimization techniques [8,16]. In addition, the spatial analogous of the CPT technique requires a precise control of the ratio between the two relevant "Rabi" frequencies. Fortunately, even in the presence of shaking in the microtraps position the ratio between the "Rabi" frequencies can be accurately controlled since mechanical vibrations give rise to a correlated shaking. Throughout the paper we have assumed to be able to cool down the atom to the lower vibrational states of the traps. In fact, sideband cooling to a temperature below 1µK with a ground state population of 98.4% has been re-ported in optical lattices [17] with parameters very similar to the ones considered here. In this case, heating rates below 1µK/s have been estimated [18]. In the presence of decoherence from heating, shaking and spontaneous scattering, fidelities above 98% [8] can be expected for the ground-state TLAO techniques discussed here. However, it is worth to note again that all these techniques can be also applied to excited states. Note that the real trapping potentials differ from simple harmonic ones, but the three level atom optic techniques discussed here do not rely on the particular shape of the trapping potentials, provided the adiabaticity is maintained during the whole process of approaching and separating the traps.
Summarizing, we have introduced a set of robust and efficient techniques to coherently manipulate and transport neutral atoms based on three-level atom optics. These atom optics techniques correspond to the natural extension of the largely investigated STIRAP, CPT and EIT techniques used in quantum optics [1,2,3] with the interaction mediated via tunneling and controlled by the shaping of the process of varying the separation between the traps. The fact that three-level atom optics makes use of the tunneling interaction means some important differences with respect to the quantum optics case, such as the time scale of the processes being in the millisecond range, the absence of electric dipole rules, or the possible use of excited states. Applications to atomic interferometry and precision measurement have been briefly discussed and some practical considerations for the implementation in dipole trap arrays have been addressed. These three-level atom optics techniques are widely applicable also in other atom optics systems such as magnetic microtraps, optical lattices, and dipole and magnetic waveguides. | 3,001.6 | 2003-08-07T00:00:00.000 | [
"Physics"
] |
Insights into the structural dynamics and helicase-catalyzed unfolding of plant RNA G-quadruplexes
RNA G-quadruplexes (rG4s) are noncanonical RNA secondary structures formed by guanine (G)-rich sequences. These complexes play important regulatory roles in both animals and plants through their structural dynamics and are closely related to human diseases and plant growth, development, and adaption. Thus, studying the structural dynamics of rG4s is fundamentally important; however, their folding pathways and their unfolding by specialized helicases are not well understood. In addition, no plant rG4-specialized helicases have been identified. Here, using single-molecule FRET, we experimentally elucidated for the first time the folding pathway and intermediates, including a G-hairpin and G-triplex. In addition, using proteomics screening and microscale thermophoresis, we identified and validated five rG4-specialized helicases in Arabidopsis thaliana. Furthermore, DExH1, the ortholog of the famous human rG4 helicase RHAU/DHX36, stood out for its robust rG4 unwinding ability. Taken together, these results shed light on the structural dynamics of plant rG4s.
G-rich DNA/RNA can form four-strand noncanonical Gquadruplex (G4) structures, which comprise Hoogsteenbonded planar G-quartets and be further stabilized by monovalent cations such as K + or Na + (1). They are highly stable and implicated in many crucial cellular processes (1)(2)(3). Their folding dynamics are closely related to human health (4,5) and can regulate plant growth, development, and adaption (6)(7)(8). Although DNA G4s (dG4s) have been studied well from different perspectives, RNA G4s (rG4s) have only gradually received attention in recent years (1,9). Compared to dG4s, because the vast majority of RNA exists in the cell as a single strand without the restriction of complementary strands, Grich RNA is easier to fold into more thermodynamically stable rG4 structures (10,11). In addition, although dG4s can fold into different kinds of topological structures, because of the presence of 2 0 -OH, rG4s predominantly fold into parallel structures (11,12). rG4s exist in mRNA and noncoding RNA (6,9,13,14) and can fine-tune both DNA and RNA metabolism (3). Currently, research on rG4s is mainly focused on animals. As multicellular organisms, plants have evolved independently and provide oxygen and food for animals. As early as 2010, Mullen et al. predicted that rG4s might play important roles in Arabidopsis (15). Furthermore, in 2012, the folding thermodynamics of plant rG4s were also systematically studied by the same group (16). In 2015, Kwok et al. (7) were the first to show in vivo that an rG4 structure in the 5 0 UTR of Arabidopsis mRNA could regulate the ATR protein expression by blocking its translation. In 2018, Cho et al. (8) showed that an rG4 in the SMXL4/5 5 0 UTR could control phloem differentiation. In 2019, Zhang et al. (17) discovered that an rG4 in SHR mRNA could trigger RNA phase liquid-liquid phase separation. In 2020, Yang et al. reported that rG4s were widely present in Arabidopsis and rice, and they played an important role in modulating plant growth as demonstrated by rG4 sequencing and phenotypic experiments (6).
The structures of rG4s are highly dynamic in cells (18,19), and the structural dynamics regulate related cell activities (3). Thus, it is critical to understand the structural dynamics of rG4s to understand their cellular and molecular functions. Although there are a few studies on rG4 folding-unfolding intermediaries and dynamics, using both bulk (16,20) and single-molecule methods (21,22), the underlying mechanism have not been well revealed. Studying this process is more meaningful for plants because the exterior environment can influence ion concentrations and temperature in plant cells, which can modulate the folding dynamics of rG4 structures (16). In addition, the dynamics of rG4s can be modulated by rG4-binding proteins (3,9), in which helicases are confirmed to play a key role in rG4 biology via resolving these highly stable structures. Therefore, the identification of rG4specialized helicases is quite important when explaining the functions of rG4s. In humans (4,9) and yeast (23,24), many helicases have been reported to unfold rG4 structures. Although rG4s are also important in plants (6), no specialized helicases have been reported to interact with rG4s in plants. This knowledge gap limits the study of rG4 function in plants.
In this research, four plant rG4s were chosen and their folding was confirmed by CD. Afterward, different concentrations of KCl were used to induce their folding, and four folding states, including two folding intermediates (G-hairpin and G-triplex), were captured by single-molecule FRET (smFRET) and verified through substrate mutation. In addition, proteomics screening identified five rG4-specialized helicases from Arabidopsis, and their specialization was confirmed by the microscale thermophoresis (MST) method. Further, the unwinding abilities of these five helicases were investigated. These findings may improve the understanding of rG4 functions in plants.
Results
The folding of four selected rG4 sequences depends on the concentrations of KCl To investigate the properties of rG4s, four reported plant rG4-forming sequences were selected from the literature (Fig. 1A), including two two-layer rG4s named 2G (C) and 2G (AA) (6,16) and two three-layer rG4s referred to as 3G (ATR) and 3G (SMXL) (7,8). Using the CD spectra, it was further confirmed that they could properly fold into parallel G4 structures in 100 mM KCl (Fig. 1B), showing an approximately 265 nm peak, consistent with previous reports (11,12).
rG4 structures are formed in ssRNA in cells. To mimic this structural environment and monitor the folding dynamics of rG4s at the single-molecule level, the fluorescently labeled substrates of these four rG4s were designed (Fig. 2, A and D). Each substrate was constructed with an ssRNA containing Cy3 and Cy5 to the ss-rG4 junctions and modified by biotin at the 5 0 end for immobilization to the PEG-modified surface of a coverslip with streptavidin. Thus, FRET signal change could sensitively report the structural change of rG4s. Based on the principle of FRET, folded rG4 should correspond to the highest FRET value, while unfolded ssRNA should correspond to the lowest FRET value. The formation of rG4 structures is induced and stabilized by K + and Na + (1). In plant cells, the concentration of K + is two orders of magnitude higher than that of Na + (25,26); therefore, K + -induced structures are biologically more relevant. In addition, the concentration of K + is physically charged in response to cellular stress. With the increase of KCl concentration, the FRET efficiency gradually shifted toward high FRET values (Fig. 2), indicating that K + induced the formation of rG4 structures as expected (16). Compared to the three-layer rG4s, the two-layer rG4s did not appear to change much. This difference may be due to their thermal stability. Compared to the two-layer rG4s, the threelayer rG4s are always stable. This observation is consistent with a published result that the folding of two-layer rG4s need higher K + concentration (16), which can reach up to 600 mM in plant cells under drought stress (16,27).
rG4s show four folding states
The K + -induced rG4 folding provided an opportunity to study the folding intermediates and pathways of rG4s. To dissect the intermediates and dynamics of rG4s, the FRET histograms were fitted to recognize the folding states. It was found that they were well fitted by four peaks using multipeak Gaussian distributions (Fig. 3, A-D). Therefore, the representative traces were also identified as four states by hidden Markov modeling, and the transition density plots of all four rG4s were built (Fig. S1), which further confirmed the multipeak fitting. Based on the principle of FRET, it was clear that the lowest FRET state should correspond to completely unfolded sequences (ssRNA) and the highest FRET state should parallel G4 (Fig. 1B). This was consistent with the fact that, with increasing KCl concentrations, the proportion of the lowest FRET value gradually decreased; accordingly, the proportion of the highest FRET value gradually increased (16) (Fig. 3, E-H). It also made sense that the FRET values of the unfolding states gradually decreased with the length increase of single strands (Fig. S2A), and the FRET values of folding states decreased with the increasing number of tetrad layers (Fig. S2B).
Because the four rG4s were parallel structures (Fig. 1B), the remaining two states should not be other G4 topologies and thus should be the folding intermediate states of rG4s. For dG4s, G-hairpin and G-triplex have been well identified as folding intermediates (28)(29)(30)(31)(32)(33). In terms of rG4s, G-hairpin and G-triplex RNA were also proposed as folding intermediates of rG4s (20,34), even though they were not directly captured experimentally in real time. Therefore, it was easy to speculate that the intermediates were very likely to be RNA G-hairpin and G-triplex. If so, the existence of G-hairpin and G-triplex should be independent of rG4. To confirm this speculation, the last column of 3G (ATR) (Fig. 4A insert, referred to as ATRG3) was mutated to detect the formation of G-hairpin and Gtriplex. As expected, three Gaussian peaks at 0.32, 0.50, and 0.74 were captured (Fig. 4A), consistent with the representative traces ( Fig. 4B). ATRG3 can only fold into G-hairpin and Gtriplex; therefore, the peaks 0.50 and 0.74 should be G-hairpin and G-triplex, respectively. The FRET values of ATRG3 in Figure 4A were close to the lower values of 3G (ATR) in Figure 3D, strongly supporting the intermediate states were Ghairpin and G-triplex, and the missing FRET of 0.92 was further confirmed to consist of well-folded rG4 structures. Thus, the folding-unfolding pathway of rG4s ( Plant RNA G-quadruplex dynamics and unfolding containing G-hairpin and G-triplex as intermediates was proposed. A previous study reported that rG4 unfolding by DHX36 showed four states. However, the states were not assigned to specific structures (35). To the best of our knowledge, this study represents the first time that both Ghairpin and G-triplex have been directly captured experimentally during rG4 folding-unfolding processes in real time.
In early studies, at physiological K + concentrations of 140 mM in vitro, three-layer rG4s were reported to be well folded (16,19), and even the relatively unstable three-layer dG4s were also able to fold well (36). However, in the present single-molecule experiments, a considerable proportion was not fully folded (Fig. 3), which may have been averaged in bulk assays. To resolve this inconsistency, the present study designed a DNA substrate named 3G (ATR)-DNA by replacing the RNA sequence of 3G (ATR) with DNA to verify the reliability of smFRET. CD spectral analysis showed that 3G (ATR)-DNA folded into parallel G4 structures, and it was confirmed that it was less stable than RNA 3G (ATR) through melting (Fig. S3). Under the same experimental conditions, it was found that 3G (ATR)-DNA also showed four Gaussian peaks (Fig. S4, A-C), and the peak values were close to those of RNA 3G (ATR) (Fig. 3D). G-hairpin and G-triplex have been well demonstrated to be the intermediates of dG4s (28)(29)(30)(31)(32)(33), which further supports the rG4 folding-unfolding proposal in the present study (Fig. 4C). In addition, the unfolded fraction of DNA was higher than that of RNA and the completely folded fraction was less than that of RNA (Fig. S4D); this was consistent with the stability results (Fig. S3). Therefore, it was reasoned that this difference may have been caused by the flanking sequences connected to rG4s. Recently, it was found by us that the proximal ssRNA impaired the stabilities of rG4 structures in vitro (22). In addition, rG4 functions were also reported to be context dependent in vivo (37). This fact can also explains why some proteins are needed to promote rG4 folding (3,8).
Identification and confirmation of plant rG4-binding helicases
As described previously, the dynamics of rG4s induced by K + were studied, the folding intermediates were revealed, and the folding pathway was proposed. In cells, the structural dynamics can also be modulated by rG4-binding proteins, in which helicases exhibit strong rG4 unwinding activity (38). In animals, many rG4-specialized helicases have been reported (4,10) and are indispensable for explaining the functions of rG4s. However, surprisingly, no plant rG4 helicases have been reported until now. To identify plant rG4 helicases, rG4 proteomics screening was performed (Fig. 5A), in which rG4 served as bait to pull down rG4-binding proteins and ssRNA was used for comparison, followed by SDS-PAGE and mass spectrometric analysis. The comparison of spectral counts was used to determine protein abundance (24,39). Five rG4 helicase candidates were identified using this method (Fig. 5B). Three of them belonged to the DEAD helicase family, and two of them were DExH family helicases. The mass spectrometric spectra of representative peptides are shown in Fig. S5. It should be noted that the eIF4A ortholog in human cells has been reported to be a specialized rG4 helicase (40), implying the reliability of the pull-down results in the present study.
To further confirm that these five helicases were rG4 specialized, they were expressed and purified to homogeneity (Fig. S6), and their binding affinity to different kinds of RNA substrates was checked using MST (Fig. 6A). The comparable binding affinity supported their specialization in rG4 recognition (Fig. 6, B-F) because smaller EC 50 values represent greater binding affinity (29). RH3 and RH53 were SUMO- tagged in the MST experiments. To exclude the influence of the SUMO tag, its binding with these substrates was measured and no obvious interaction was detected (Fig. S7).
DExH1 and DExH15 can unfold rG4 structures in the 3 0 -5 0 direction After identifying and confirming rG4-specialized helicases, this study sought to determine whether they had rG4 unwinding ability. Gel shift has been reported to be an intuitive and reliable method to study rG4 unwinding (41). Considering this, the Cy3-labeled 3 0 tail rG4 and 5 0 tail rG4 structures were examined, and it was reported that they were unable to basepair by unlabeled ssRNA traps unless they were unfolded by helicases (Fig. 7A). The five helicases were used to unwind both 3'ss-rG4 and 5'ss-rG4. It was found that only DExH1 and DExH15 showed obvious rG4 unwinding activity (Fig. 7, B and C). None of the three DEAD helicases showed rG4 unwinding under the experimental conditions, even when their concentration reached as high as the micromolar levels (not shown). The gel shift results clearly showed that DExH1 and DExH15 could resolve 3 0 tail rG4 (Fig. 7, B and C, upper panels), but they could not unfold 5 0 tail rG4 under the same experimental conditions (Fig. 7, B and C, bottom panels), implying that they were 3 0 -5 0 helicases. This was consistent with their preferential binding of 3 0 tail rG4 in the MST experiment (Fig. 6, E and F).
As control experiments, no unwinding occurred in the absence of DExH1/DExH15 or ATP (Fig. 7, B and C, upper panel), implying that the rG4 unwinding of these two helicases is ATP dependent, and the unwinding should not result from helicase binding with the assistance of traps. Comparison made it clear that DExH1 was much more active than DExH15 in terms of rG4 unwinding (Fig. 7D).
RNA helicases can be classed into DEAD and DExH helicases, which share different unwinding mechanisms (42). The three DEAD helicases did not show rG4 unwinding ability, which may be explained from two aspects. First, some of them may need assistant proteins to unwind G4 structures together; for example, eIF4A unwinding activity can be activated by its cofactors (43). In agreement with this, eIF4H (44) and eIF4G (45) were recently reported to selectively bind rG4s. Second, some may function in cells through G4 recognition but not unfolding, like XPB (46). DExH helicases DExH1 and DExH15 are 3 0 -5 0 helicases, which is consistent with rG4 unwinding helicases DHX9 and DHX36 from humans. In addition, through homolog research using the BLAST tool against the human protein database, it was found that DExH1 was an ortholog of DHX36, sharing the highest identity (Fig. S8). DHX36 is one of the most well-studied rG4 helicases, and it has been proposed that DHX36 can resolve most rG4 in human cells (47). This further supported the results of the present study.
rG4 was unfolded by DExH1 with intermediate states of Gtriplex and G-hairpin
After checking the unwinding activity of identified rG4 helicases, the present study investigated the intermediate states of rG4s during helicase unfolding. Because DExH1 was more active than DExH15, DExH1 was selected to resolve 3G (ATR) (Fig. 8A), which was the most stable structure among four selected rG4s (Fig. 3). First, since DExH1 unwinding is ATP dependent (Fig. 7B) and Mg 2+ is indispensable for the hydrolysis of ATP by helicases, it was verified that in the presence of 100 mM KCl and 5 mM MgCl 2 , there was no significant FRET change (Fig. S9). After injecting 100 nM DExH1 and 1 mM ATP into the reaction chamber, FRET fluctuations were captured (Fig. 8B). In agreement with this finding, after 2 min of incubation, the high FRET population decreased, accompanied by an increase of the low FRET population (Fig. 8C), indicating that the rG4 structures were unwound by DExH1. The distribution could also be well fitted by multipeak Gaussian distributions (Fig. 8D), and the peaks were close to those of dynamic folding (Fig. 3D), indicating that G-triplex and G-hairpin were also intermediate states during DExH1 unwinding (Fig. 4C). This phenomenon is consistent with the translocation-based mechanism of RHAU/DHX36 (35,(48)(49)(50). Accordingly, the fraction of folded states significantly decreased after the injection of 100 nM DExH1 and 1 mM ATP (Fig. S9B).
Discussion
In recent years, the characteristics of dG4s have been reported in great detail using different methods. However, compared with dG4s, rG4s have been less well studied, especially in plants. As described previously, the present study pioneered research of the folding pathways and helicasecatalyzed unfolding of plant rG4s using single-molecule methods. These results may help elucidate their biological functions.
Under single-molecule conditions, the three-layer rG4 structures are not very well folded under physiological K + conditions (Fig. 3), as previously reported (11,16,19). First, the small number of unfolded structures may be averaged out in bulk assays. Second, the difference may result from the ssRNA linked at both ends of rG4s, as described in a recent report (22). The substrates in the present research are more biologically meaningful because they are in line with the structural environments inside cells. In accordance with this, in both animal (3) and plant (8) cells, some factors function by directly stabilizing the rG4 structures. Therefore, in addition to the unwinding mechanisms, the molecular mechanisms of rG4 protein stabilization are also worthy of further study.
Four folding-unfolding states of rG4s were detected in both K + -induced dynamics (Fig. 3) and helicase unwinding (Fig. 8), and the intermediate states were assigned to G-triplex and Ghairpin. Thus, the folding pathways of both dG4s (32,51) and rG4s are related to G-triplex and G-hairpin. The difference is that the folding of dG4s may include a variety of other G4 topologies (28,32), while there have been no reports of rG4s with mixed structures. Recently, RHAU/DHX36 was reported to unfold rG4-containing substrate showing four states using smFRET (35). However, in these experiments, the two fluorescent dyes were not labeled at both ends of rG4. One of the dyes was labeled at the end of the ssRNA tail, so the interference of the ssRNA could not be ruled out. In addition, the study did not allocate the four states to specific molecular structures.
In the present research, five plant rG4-specialized helicases were identified for the first time using proteomics screening and MST (Figs. 5 and 6). Functional prediction indicates that they may be involved in pre-mRNA splicing and translation regulation. This information will be helpful for the study of rG4 in plants. The homologous proteins of eIF4A and DExH1 in animals have been reported to function through rG4 in animals, thus demonstrating the reliability of the present experiments. It should be noted that RHAU/DHX36, the homolog of DExH1, is the most functional rG4 helicase reported to date. RHAU/DHX36 recognizes the G4 structures through the RHAU-specific motif (RSM), which is not conserved in DExH1 in terms of primary sequence (Fig. S8). There may be two possible domains in DExH1 to specially recognize rG4s. The first possible domain is the N-terminal RGG domain, which is one of three conserved rG4 recognition domains (3). The second domain is DExH1 [45][46][47][48][49][50][51][52][53][54][55][56][57][58] , which is predicted to fold α-helix similar to RSM secondary structure using AlphaFold. In addition, the position of this α-helix in DHxH1 corresponds to the position of RSM in RHAU/DHX36, meaning that their secondary structures are conserved, although the primary sequences are not. The study of the biological functions and molecular mechanisms of the identified helicases modulating rG4s in cells is an exciting research focus.
Nucleic acid preparations
The sequences of all nucleic acids are listed in Table S1. The RNA substrates used for smFRET were purchased from Dharmacon, and the remaining oligonucleotides were purchased from General Biol. The smFRET oligonucleotides were amino-modified at two specific positions and fluorescently labeled by Cy3 and Cy5 according to previously described protocols (52). Notably, Cy3 and Cy5 were mixed to label substrates at the same time; therefore, there was a 50% chance that one substrate would be labeled with two Cy3s or two Cy5s. This would not have been detectable in smFRET because FRET can only be detected when one Cy3 and one Cy5 are labeled on the same substrate. All of the annealing was performed by incubation at 95 C for 5 min and then cooled to room temperature for 3 h in the corresponding reaction buffers. For simplicity, Kn is used to represent reaction buffer in the text, where n indicates the concentration of KCl; for example, K100 represents 25 mM Tris-HCl at pH 8.0 and 100 mM KCl.
Protein expression and purification
Full-length genes of RH3, RH53, eIF4A1, DExH1, and DExH15 from Arabidopsis complementary DNA were cloned into pET28a and tagged with 6× His-SUMO at their N-terminal. Each expression vector was transformed into Escherichia coli Rosetta2 (DE3) and induced with 0.3 mM IPTG at 16 C for 16 h after the A 600 reached 0.6 at 37 C. After the high-pressure homogenization of cells and sequential centrifugation, the expressed protein in the supernatant was captured by nickel-nitrilotriacetate beads. The 6× His-SUMO tags of eIF4A1, DExH1, and DExH15 were removed by Ubiquitin-like protein-specific protease and collected for experimental comparison. In terms of RH3 and RH53, it was found that the removal of their tags decreased their stability in solution; therefore, this tag was retained for them. Finally, gelfiltration chromatography was used to polish the captured protein. All of the purified proteins were frozen in small aliquots in liquid nitrogen and stored at −80 C in a storage buffer (10 mM Tris-HCl, pH 7.9, 200 mM KCl, 1 mM DTT, and 50% glycerol (v/v)). The images of their SDS-PAGE gels are shown in Fig. S6.
CD spectropolarimetry
The CD analysis was performed on a Chirascan V100 (Applied Photophysics Ltd), equipped with a temperaturecontrolled cell holder and a temperature probe. The final concentration of substrates was 10 μM in 25 mM Tris-HCl, pH 8.0, and 100 mM KCl. The CD spectra were recorded in a range of 220 to 320 nm with a 0.5 mm path length using a quartz cell. For melting, the temperature was raised 1 C/min continuously between 20 to 98 C and spectral data were recorded every 2 min.
smFRET data acquisition and analysis
The equipment and preparation were the same as described in a previous work (28). An oxygen scavenging system (0.8% D-glucose, 1 mg/ml glucose oxidase [266,600 units/g], 0.4 mg/ ml catalase [2000-5000 units/mg], and 4 mM Trolox) was added to the reaction buffer to prevent blinking and bleaching. The reaction buffer for the rG4 dynamics experiments contained 25 mM Tris-HCl, pH 8.0, and a corresponding concentration of KCl. Then, 5 mM MgCl 2 was added to the reaction buffer for the unfolding study of DExH1. In helicase unfolding experiments, 100 nM DExH1 and 1 mM ATP flowed into the chamber simultaneously. The exposure time of 100 ms was used for all of the measurements at a constant temperature of 22 C. The FRET efficiency was calculated using I A /(I D + I A ), where I D and I A represent the intensity of the donor (Cy3) and acceptor (Cy5), respectively. smFRET histograms were generated by selecting 30 to 50 frames of each trace from over about 300 molecules. Data analyses were carried out using scripts written in R. Hidden Markov modeling was used to identify folding states (53), and the transition density plots were built using our previous method (28). All of the data fitting was performed by OriginPro 2017.
Proteomics screen
The plant extraction protocol was conducted as described in a previous work (54), and the pull-down method was conducted as previously described (24,39). Briefly, 1 mg of streptavidincoated beads (M-280 Dynabeads, Life Technologies) was washed with 200 μl protein-binding buffer (20 mM Tris-HCl, pH 7.5, 300 mM KCl, 0.1% Tween-20, 2 mM EDTA, and 2‰ RNase inhibitor (v/v) [Promega]) and resuspended to a final concentration of 5 mg/ml. One equal volume of 4 μM biotinmodified RNA was added and incubated for 20 min at 20 C. The beads were then transferred and washed with pull-down buffer (protein-binding buffer, adding cocktail [Sigma]). One gram samples of 7-day-old Arabidopsis were frozen with liquid nitrogen and ground into fine powder. Five milliliters of pulldown buffer was added and centrifuged for 20 min at 4 C and 16,000g, followed by filtration through a 0.20 μm filter. The Arabidopsis lysates (5 ml per 1 mg of beads) were incubated with the substrate-conjugated Dynabeads at 4 C for 4 h, collected by magnetic rack, washed three times using the pull-down buffer, and then eluted with 50 μl SDS-PAGE loading buffer through boiling for 5 min. After two biological repeats for each substrate, the samples were run in an SDS-PAGE gel and subjected to LC-MS/MS analysis (BGI) flowing the published protocol (24).
MST
The MST assays were carried out using a NanoTemper monolith NT.115 (29). The experimental conditions consisted of 25 mM Tris-HCl, pH 8.0, and 300 mM KCl. The concentration of the fluorescein-labeled substrate was 20 nM. The experiments were performed using 40% LED power and 40% MST, with a laseron time of 30 s and laser-off time of 5 s, at a constant temperature of 22 C. The thermophoresis signal was recorded and fitted via NTAnalysis from at least two independent experiments.
Gel shift assays
The rG4 unwinding assays were performed using gel shift according to the procedure described in a published work in three independent repetitions (41). Briefly, different concentrations of purified DExH1 and DExH15 were mixed with 160 nM rG4 substrates in 25 mM Tris-HCl, pH 8.0, 50 mM KCl, 5 mM MgCl 2 , 2 U/μl RNase inhibitor, 10% glycerol, and 1 mM DTT. The mixture was incubated for 10 min at 37 C. Then, one equal volume of 4 mM ATP was added along with 1.6 μM traps to initiate reactions and incubated for 30 min at 37 C. Then 5× stop buffer (125 mM EDTA and 50% glycerol (v/v)) and proteinase K (final concentration of 2 mg/ml) were added sequentially. After waiting for 10 min at 37 C, the reaction products were electrophoresed on a 12% native PAGE. Finally, the gels were imaged using the Cy3 channel on a ChemiDoc MP Visualization System (BioRad) and analyzed in ImageJ (National Institutes of Health).
Data availability
The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE (55) partner repository with the dataset identifier PXD031877.
Supporting information-This article contains supporting information. | 6,388.2 | 2022-06-01T00:00:00.000 | [
"Biology"
] |
Seabirds of Easter Island , Salas y Gómez Island and Desventuradas Islands , southeastern Pacific Ocean
We reviewed available information on seabirds inhabiting Easter Island, Salas y Gómez Island and Desventuradas Islands and their adjacent waters through an analysis of published and grey literature. Results obtained indicate that a total of 37 species are present in the study area and that, among the orders represented, the Procellariiformes and Charadriiformes are the dominant taxa (29 species). Moreover, the family Procellariidae is represented by 13 species and Laridae by 7 species. There has been an increase in new records over the past six years but no systematic studies have been developed. The need for further research that focuses on ecological aspects and anthropogenic impacts is critical in order to develop adequate conservation strategies.
INTRODUCTION
The study of seabirds can promote the understanding of oceanic ecosystems since this group has several attributes that make them a good indicator species (e.g., easy to detect and identify; depend entirely on marine ecosystems to obtain their food; are highly mobile and consequently integrate ecosystem variability across different spatial and temporal scales) (Ballance, 2007).Seabird distribution patterns are quite well understood (Harrison, 1987) and have usually been correlated to habitat characteristics such as sea surface temperature, salinity or productivity (Pocklington, 1979;Schneider & Brown, 1986;Schneider, 1990;Elphick & Hunt, 1993).
Oceanic islands possess a significant cultural and economic relevance to mankind throughout the world (Moller et al., 2009), but from an ornithological point of view, many have been severely degraded through the centuries by human societies (Moller et al., 2009), both directly and indirectly.Some islands have been substantially altered through a long history of seabird predation by introduced mammals such as rats, dogs, cats and foxes, among others.These introduced species have been responsible for an important proportion of the decline of seabird populations (Courchamp et al., Glynn et al., 2007Glynn et al., ). 2003;;Jones et al., 2008;Baker et al., 2014), on occasions even more so than other factors such as bycatch in fisheries and the effects of overfishing and pollution (Blackburn et al., 2004;Jones et al., 2008).
In the present day, knowledge regarding seabirds inhabiting Chilean oceanic islands: Easter, Salas y Gómez and Desventuradas islands, is rather scarce and the conservation status of these islands is depressed, as pointed out by Schlatter (1987).The Juan Fernandez Archipelago presents a more advanced level of scientific information only because NGO Oikonos conducted systematic research, filling in these knowledge gaps (see http://oikonos.org/juan-fernandez-islands-conservancy/).
In this paper we aim to update the information about the birds of these islands based on the available literature since 1914 to present and propose categories of residence for Chilean territory according to the current state of knowledge.
MATERIALS AND METHODS
The present study pertains to Easter Island (27°09'S, 109°26'W), Salas y Gómez Island (26°27'S, 105°28'W) and Desventuradas Islands (26°20'S, 79°58'W), all located in the subtropical eastern South Pacific Ocean.The degree of isolation of Easter and Salas y Gómez islands with respect to other oceanic islands is high: Pitcairn is located 2,250 km to the west, Galapagos 3,872 km to the northeast, Juan Fernandez, 3,140 km to the southeast, Desventuradas Islands, 2,900 km to the east and the closest land mass, South America, some 3,700 and 3,400 km to the east, respectively (Fig. 1).
This paper includes a bibliographic review from the libraries of both public and private universities in Valparaíso and Santiago, the "William Mulloy" library of the "Father Sebastian Englert" Anthropological Museum on Easter Island, private libraries of national and international ornithologists and the archives of the provincial office of the Corporación Nacional Forestal (CONAF) on Easter Island.A literature search was done with Google Scholar (see http://scholar.google.cl,accessed march-october 2012) using the keywords: bird*, Easter Island, Salas* y Gómez and Desventuradas.Additional articles were found though references cited in articles resulting from our searches.
Taxonomy, scientific and common English names used in this review are those proposed by the South American Classification Committee (SACC) of the American Ornithologist Union (AOU) (Remsen et al., 2012).Statuses are proposed by the authors according to the information gathered in this review.Regular: birds regularly observed throughout the year and with several records in different years; Occasional: birds with few observations on the islands which are not regularly; Breeder: birds with reported breeding colonies across different years; Undetermined: birds without enough information to be included in a previous category.
Additionally, the Jaccard Similarity Coefficient was calculated to express the degree of similarity between sites (Pielou, 1975;Magurran, 1988) based on the number of species present at each one of them.The coefficient was obtained using the number of breeding species on each island, according to the following expression: where a = number of species breeding in site A, b = number of species breeding in site B and c = number of species breeding in both sites A and B.
RESULTS AND DISCUSSION
Our results show that seabirds present in the study area are composed of 4 orders, 9 families and 37 species.Easter Island encompasses 4 orders, 7 families and 25 species.Salas y Gómez Island includes 4 orders, 6 families and 16 species and Desventuradas Islands include 4 orders, 8 families and 22 species.Table 1 details the seabird species recorded for each island.
Among the registered orders, the Procellariiformes (e.g., petrels and shearwaters) and Charadriiformes (e.g., sandpipers, noddies and terns) are the most noteworthy, with 17 and 12 species, respectively (Fig. 2).However, under a family level analysis, there is a clear predominance of Procellariidae, with 13 species, followed by Laridae, with 7 species, together accounting for 54.1% of the overall seabird species richness in the study area (Fig. 3).Seabird representatives from these two orders are quite consistent with those reported for other oceanic islands (e.g., Williams, 1960;Benton & Spencer, 1995;Thibault & Bretagnolle, 1999), as they are characterized by high mobility and widespread distributions (Carboneras, 1992).
When comparing seabirds present within and between each island in the study area, we observed that Easter Island holds 89.3% (25 spp.) of the total reported species, Salas y Gómez Island 57.1% (16 spp.) and Desventuradas Islands 59.5% (22 spp.).The Jaccard Similarity Coefficient obtained shows that Easter Island and Salas y Gómez Island share a significant number of species (0.64) while Salas y 2).
The values obtained are significantly higher than those reported by Schlatter (1987) because this study includes a greater number of species for all the islands in this study.Schlatter (1987) reported 14 species on Easter Island.Later, Marin & Cáceres (2010) reported a total of 19 species for Easter Island, while this study updated the list up to 25 species.The case for Salas y Gómez Island is similar: Schlatter (1987) reported 9 species, Vilina & Gazitúa (1999) 14 species and this work increases that number of species to 16.For Desventuradas Islands, Millie (1963) reported 8 species, Schlatter (1987) 4).
The results presented here indicate that, while there has been an increase in the number of species recorded in the study area, not much else has been achieved in B 19,22,25 B 14,15,18,23 B 3,6,13,14,20 Thus it is understandable that Easter Island presents a higher number of new species recorded than the other islands, given that it is the only one inhabited by people and an increasing number of tourists are visiting the island for birwatching purposes (Jaramillo et al., 2008).As a national park, trained personnel in charge of protecting wildlife provide regular observation effort for seabirds on their patrols.Salas y Gómez Island is uninhabited, landing by sea is very difficult and the only regular visits to the island Access to the Desventuradas Islands is completely restricted due to the presence of a Chilean navy base, so the approach to the islands and its adjacent waters has been prohibited since the early 1980's.An easing of restrictions in the early 2000's opportunistically allowed scientists to visit the islands to gather information, but no systematic research has been conducted since the 1960's (Millie, 1963;Bahamonde, 1966Bahamonde, , 1974)).
Although Easter Island provides optimal conditions for developing systematic studies on some particularly accessible seabird populations, these have not yet been undertaken, mainly because of the lack of funding agencies interested in funding mid to long term studies in this remote and costly location.This trend is starting to be reverted by joint efforts of public institutions, universities, NGOs and local islanders.Involving the Chilean Navy and other stakeholders in promoting studies in the relatively unperturbed marine park around Salas y Gómez Island, the most important and isolated breeding site for several oceanic seabirds in the central Pacific, and in the restricted Desventuradas Islands will allow to scientists to have a better view of population sizes and breeding ecology for these species.Subsequently, studies involving interactions with fisheries, effects of marine debris and impacts of introduced species, among other topics, are the next logical step to fill in these knowledge gaps.
In conclusion, although knowledge regarding seabirds inhabiting the Chilean oceanic islands remains scarce, important contributions have been made in recent years.It is imperative to generate joint efforts among institutions and funding to develop appropriate and efficient conservation strategies.
AKNOWLEDGMENTS
We thank Pedro Lazo Hucke and Enrique Tucki (CONAF-Easter Island) for their support and valuable information on the birds of Easter and Salas y Gómez islands and to Dr. Alejandro Simeone for his critical reviews, which greatly improved this manuscript.Finally, we thank PEW Environment Trust, through the Global Ocean Legacy-Easter Island program, for their financial support for fieldwork in Easter Island.
Figure 1 .
Figure 1.Location of Easter Island, Salas y Gómez Island and Desventuradas Islands in relation to the closest islands in the South Pacific and South America (modified from Glynn et al., 2007).
Figure 2 .
Figure 2. Number of families and species for each of the seabird orders recorded in the study area.
Figure 3 .
Figure 3. Seabird species richness by family recorded in the study area.
Table 1 .
Seabirds recorded in Easter Island, Salas y Gómez Island and Desventuradas Islands and its proposed residency status for Chilean territory (RE: regular, OC: occasional, B: breeder, U: undetermined).Conservation status according to International Union for the Conservation of Nature (IUCN, 2013) and the Ministry of the Environment, Chile (MMA, 2009): EN: endangered, VU: vulnerable, LC: least concern.
Table 3 .
Number of seabird species reported by scientific papers and bibliographic reviews for Easter Island, Salas y Gómez Island and Desventuradas Islands.
Table 4 .
Sightings of new species for Chile and/or Easter Island and Salas y Gómez Island.This means access for both scientists and civilians is difficult.Given that the distance from Easter Island is ca.400 km, there are no vessels based on Easter Island that can cover the distance round trip without refueling at sea.Further research is urgently needed on this island because since 2010, it corresponds to the largest marine park in Chile known as Motu Motiro Hiva (Ministerio de Economía D.S. N°235/2010). | 2,514.4 | 2014-10-10T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
Tumor Microenvironment and Genes Affecting the Prognosis of Temozolomide-Treated Glioblastoma
Glioblastoma (GBM) is the most frequent primary brain tumor in adults and has a poor prognosis due to its resistance to Temozolomide (TMZ). However, there is limited research regarding the tumor microenvironment and genes related to the prognosis of TMZ-treated GBM patients. This study aimed to identify putative transcriptomic biomarkers with predictive value in patients with GBM who were treated with TMZ. Publicly available datasets from The Cancer Genome Atlas and Gene Expression Omnibus were analyzed using CIBERSORTx and Weighted Gene Co-expression Network Analysis (WGCNA) to obtain types of highly expressed cell types and gene clusters. Differentially Expressed Genes analysis was performed and was intersected with the WGCNA results to obtain a candidate gene list. Cox proportional-hazard survival analysis was performed to acquire genes related to the prognosis of TMZ-treated GBM patients. Inflammatory microglial cells, dendritic cells, myeloid cells, and glioma stem cells were highly expressed in GBM tissue, and ACP7, EPPK1, PCDHA8, RHOD, DRC1, ZIC3, and PRLR were significantly associated with survival. While the listed genes have been previously reported to be related to glioblastoma or other types of cancer, ACP7 was identified as a novel gene related to the prognosis of GBM. These findings may have potential implications for developing a diagnostic tool to predict GBM resistance and optimize treatment decisions.
Introduction
Glioma is a tumor originating from neuroglial cells of the brain and spinal cord, and is known for its poor prognosis [1]. Glioma is classified into four malignancy grades based on histological criteria, such as atypia and necrosis. Among them, glioblastoma (GBM) is the most frequent and malignant grade 4 tumor [2]. Surgery is the most effective treatment for GBM, albeit the challenging location and capability to infiltrate into surrounding healthy tissue make it difficult [3]. In addition, remnants must be treated with chemotherapy and radiation therapy post-surgery [4]. For chemotherapy of GBM, the cytotoxic anticancer drug Temozolomide (TMZ) is the drug of choice. Unfortunately, due to the highly heterogeneous and mutation-prone nature of GBM, more than half of patients do not respond to TMZ, limiting the median survival to 12-15 months [5].
TMZ resistance is a significant obstacle that must be tackled for the successful treatment of GBM. Previous studies have shown that glioma stem cells and the MGMT repair system play key roles in TMZ resistance, and that PI3K/AKT, Wnt/b-catenin, and JAK/STAT pathways are also involved [6][7][8]. However, limited research has been conducted on GBM transcriptome data to investigate tumor microenvironments and genes that affect prognosis in GBM patients treated with TMZ [9]. 2 of 11 In this study, we aimed to identify putative transcriptomic biomarkers with predictive value in TMZ-treated GBM patients using GBM single-cell RNA (scRNA) sequencing data, as well as clinical information and gene expression data from publicly available databases. To take the tumor microenvironment into account, we used Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts (CIBERSORTx) to identify highly expressed cell types, and employed Weighted Gene Co-expression Network Analysis (WGCNA) to construct gene modules that are highly correlated with these cell types. We also used Differentially Expressed Genes (DEG) analysis on short and long survival groups, and intersected the resulting lists to obtain a candidate gene list. Finally, survival analysis was performed to analyze these genes as biomarkers.
scRNA Sequencing Data: GEO Dataset
Glioblastoma single-cell RNA (scRNA) sequencing dataset GSE162631 was downloaded from the Gene Expression Omnibus (GEO) database. This dataset included sequences of tumor core and peripheral tissue from 4 GBM patients, which were processed using magnetic-activated cell sorting to separate endothelial cell marker CD31-positive cells.
Bulk RNA Sequencing Data and Clinical Data: TCGA Dataset
Public clinical data and gene expression data of patients meeting the criteria were downloaded from The Cancer Genome Atlas (TGCA) database using R packages "TCGAbiolinks" and "SummarizedExperiment" [10][11][12][13]. Data from patients who have undergone only TMZ chemotherapy were collected. A total of 35 available samples in the TCGA-GBM cohort were selected for further analysis. Since the survival information was right-censored data, preprocessing was performed under the assumption of exponential distribution using the mean residual lives to estimate the survival time of censored data before use, except for the final survival analysis. The patients were divided into 3 survival groups: "short survival group" with survival of fewer than 180 days, "medium survival group" with survival from 180 to 730 days, and "long survival group" with survival of more than 730 days based on two points of discontinuation ( Figure S1, Table S1a). The short and long survival groups were compared in further analyses.
scRNA Sequencing Data Analysis
The R package "Seraut" was used to analyze 8 scRNA sequence data from the GSE162631 dataset [14]. The original data contained 120,218 cells. The percentages of mitochondria and ribosomal RNA were calculated using the "PercentageFeatureSet" function. Cells were filtered in if they had genes greater than 200, a count of RNA per cell greater than 500, and a percentage of mitochondrial read less than 20%. After filtering, a total of 112,359 cells remained. Next, each of the sequencing data were normalized through log normalization, then merged using the "IntegrateData" function.
The Merged data were scaled using the "ScaleData" function, then dimension was reduced using the "RunPCA" function with the first 2000 highly variable genes screened through the "FindVariableFeatures" function. Subsequently, the top 50 principal components were selected to conduct additional dimension reduction using the UMAP method. We used the "FindNeighbors" and "FindClusters" functions with resolution = 1.0, resulting in 30 clusters. Finally, we used the "FindAllMarkers" function with logfc = 0.5, minpct = 0.35 to find the top 10 marker genes of each cluster. Marker genes were screened using the corrected p-value under 0.05, then we used cellKb database (https://www.cellkb. com/ (accessed on 19 October 2022)) to identify each cluster by cell type [15].
CIBERSORTx
CIBERSORTx (https://cibersortx.stanford.edu/ (accessed on 19 October 2022)) is a machine learning algorithm developed by Stanford, which accurately estimates the relative proportions of cell subsets in tissue bulk RNA sequencing data based on the input of scRNA sequencing data matrix [16][17][18]. The difference in cell ratio of each sample was visualized using "ggplot2". Based on the gene signature matrix from the result of CIBERSORTx, cells with high average expression were further studied to learn how intracellular gene expression characteristics affect prognosis.
WGCNA
WGCNA creates a weighted correlation network to identify modules based on gene expression data and finds modules with the highest correlation to the trait data [19]. The "blockwiseModules" function was used with power = 7, minModuleSize = 30, maxBlockSize = 3000, mergeCutHeight = 0.3.
DEG and Functional Pathway Analysis
DEG analysis was performed between the short and long survival groups [20][21][22]. Next, we overlapped the gene modules that had the highest correlation with each selected cell from CIBERSORTx with genes upregulated in either short or long survival groups from DEG analysis. Functional pathway analysis was performed using the "clusterProfiler" package [23,24].
Survival Analysis
Survival analysis was performed using the overlapping genes from the previous step to select significant genes. Analysis and visualization were performed using "survival", "survminer" packages [25]. First, univariate analysis was performed to select genes with p-value < 0.05, and then analyzed with Cox proportional-Hazard survival analysis using the stepwise variable selection method, resulting in a regression equation and verification of the significance of the variables. "My.stepwise.coxph" function from the "My.stepwise" package was used for variable selection, with sle = 0.05, sls = 0.05.
Highly Expressed Cell Types and Related Gene Modules
Thirty cell clusters were acquired from scRNA sequencing analysis and annotated with cell types (Figure 1). The top five marker genes in each cluster are available in Table S1b. By applying CIBERSORTx on the scRNA sequencing GEO dataset and TCGA-GBM RNA expression of 32 GBM samples, the relative proportions of cell types were obtained. There were seven types of cells with a prevalence higher than 5%, and these cell types were selected for further analysis of their gene expression patterns. The selected cell types were Dendritic cell_C6, Inflammatory microglial cell_C0, Inflammatory microglial cell_C1, Inflammatory microglial cell_C4, Myeloid cell_C5, Myeloid cell_C7, and Proneural glioma stem-like cell_C2, as shown in Table S1c.
WGCNA was conducted to explore the correlation between the fraction of the chosen cell groups and RNA expression from 35 TMZ-treated patients. Data from 32 samples were used, with the exclusion of three outliers. The correlation between samples and traits is demonstrated in the clustering dendrogram ( Figure 2a). Power = 7 was chosen as the appropriate soft power value (Figure 2b,c). A total of 47 modules were constructed and visualized as a cluster dendrogram and correlation heatmap (Figure 2d,e). Next, modules with the highest correlation and lowest p-value with the top seven highly expressed cells were selected ( Table 1). The highest correlated gene modules were pink, royalblue, darkmagenta, darkturquoise, black, skyblue3, and green. 1 Total number of genes in the gene module. 2 Number of genes that overlap with the differentially expressed genes in high survival group. 3 Number of genes that overlap with the differentially expressed genes in low survival group. 1 Total number of genes in the gene module. 2 Number of genes that overlap with the differentially expressed genes in high survival group. 3 Number of genes that overlap with the differentially expressed genes in low survival group.
Survival-Related Genes and Functions
To explore genes related to survival in TMZ-treated GBM patients, we conducted DEG analysis between the two groups of different survival prognoses. The cut-off criterion was set as |log2 fold change| > 1 and false discovery rate (FDR) < 0.05. A list of 752 differentially expressed genes (high DEG: 151 genes upregulated in the long survival group and low DEG: 601 genes upregulated in the short survival group) was obtained.
Functional pathway analysis revealed alterations in biological pathways based on the differentially expressed genes. Pathways related to neural development were upregulated, and those mediating immune responses were downregulated (Figure 3a). To classify the difference in pathways according to the cell types, genes from the DEG results that overlap with the gene modules identified by WGCNA were organized as shown in Table 1. This process reduced the number of candidate genes from 752 to 264.
Survival-Related Genes and Functions
To explore genes related to survival in TMZ-treated GBM patients, we conducted DEG analysis between the two groups of different survival prognoses. The cut-off criterion was set as |log2 fold change| > 1 and false discovery rate (FDR) < 0.05. A list of 752 differentially expressed genes (high DEG: 151 genes upregulated in the long survival group and low DEG: 601 genes upregulated in the short survival group) was obtained.
Functional pathway analysis revealed alterations in biological pathways based on the differentially expressed genes. Pathways related to neural development were upregulated, and those mediating immune responses were downregulated (Figure 3a). To classify the difference in pathways according to the cell types, genes from the DEG results that overlap with the gene modules identified by WGCNA were organized as shown in Table 1. This process reduced the number of candidate genes from 752 to 264. The genes were grouped according to cell type and good or bad DEG, resulting in six arbitrary subgroups, and then subjected to functional pathway analysis. In group 1, genes The genes were grouped according to cell type and good or bad DEG, resulting in six arbitrary subgroups, and then subjected to functional pathway analysis. In group 1, genes related to the plasma membrane region, cell-to-cell adhesion, upregulation of gene expression, suppression of cellular metabolism, nitrogen compound metabolism, cell death, and anatomical structure development were upregulated, and genes regarding innate immune responses, antigen binding, and hormone metabolic process were downregulated (Figure 3b). In group 2, genes related to the cell-substrate junction, cell surface, plasma external membrane composition, nitrogen compound transport, intracellular anatomical structures, and cytoplasm were upregulated, and genes regarding cytokine activity, cell morphogenesis, response to RNF, extrinsic apoptotic pathway, cytokine receptor binding, signaling receptor activity were downregulated (Figure 3c).
In groups 3 and 4, since the analysis using a threshold of significance level 0.05 did not yield any significant results, a threshold of 0.10 was used. In group 4, genes related to cation binding and catalytic activity were upregulated (Figure 3d). In group 5, genes related to nucleic acid binding, ion binding, organic cyclic compound, and heterocyclic compound binding were upregulated (Figure 3e). In group 6, no significant pathway was returned due to the low number of genes.
Survival Analysis Results
With the 264 candidate genes that were both present in DEG results and gene modules of the most prevalent cell types, survival analysis was performed to determine the genes that significantly contribute to survival in TMZ-treated patients. Firstly, univariate Cox regression was performed to select genes with p-values less than 0.05. Among the 90 selected genes, 66 genes that have a reported "external_gene_name" and have published literature related to human disease in Pubmed, the Cochrane library, or EMBASE were selected. Next, multivariate cox analysis utilizing stepwise variable selection was performed. The significance of the model and genes was evaluated by the Wald test and the p-value of each gene. Finally, the Cox proportional-hazard assumption was tested using a time-dependent Cox proportional-hazard model, including significant covariates, and confirmed that the model satisfies the assumption.
The final cox proportional-hazard model revealed that ACP7, EPPK1, PCDHA8, RHOD, DRC1, ZIC3, and PRLR were statistically significant. Based on the significant genes in univariable analysis, ACP7, EPPK1, PCDHA8, RHOD, DRC1, and PRLR were significantly associated with a short survival, while ZIC3 was associated with a long survival. Based on the significant genes of the final model, ACP7, EPPK1, PCDHA8, RHOD, and DRC1 were significantly associated with short survival, while ZIC3 and PRLR were associated with long survival when compared with the genes with positive coefficients. The hazard ratios are described in Table 2.
Discussion
In this study, a deconvolution method was used to evaluate the cell type prevalence in GBM tissue. This information was then utilized to identify genes associated with prognosis in patients with GBM who were treated with TMZ. Integrating these results with DEG analysis between long and short survival groups narrowed down the candidate genes, and the final cox model revealed ACP7, EPPK1, PCDHA8, RHOD, DRC1, ZIC3, and PRLR as putative biomarkers. These findings are congruent with the previous literature, as ZIC3, PCDHA8, PRLR, and DRC1 have been reported to affect survival in GBM or types of gliomas, and EPPK1 and RHOD have been previously connected to tumorigenesis or tumor invasion in other types of cancers. It is noteworthy that, to the authors' best knowledge, this study is the first to demonstrate that ACP7 has an influence on the survival of GBM patients.
In particular, ZIC3 was associated with longer survival, which is consistent with previous studies that have shown that ZIC3 was downregulated in malignant high-grade glioma [26]. PCDHA8, a member of the protocadherin alpha family, was associated with shorter survival. It is known to participate in neural cadherin-like adhesion, serving a key role in brain cell connection [27], and previous studies have reported that PCDHA8 is hypermethylated in gliomas [28]. Another study reported that the PCDH-gamma-A11 gene is hypermethylated in astrocytoma and inactivates cell-to-cell contact in the brain, causing astrocytoma invasion [27]. The relationship between PCDHA8 promoter methylation and its expression, and its effect on GBM survival, needs to be evaluated through further studies. As for PRLR, prolactin (PRL) has been traditionally associated with lactation and fertility, but recently it has been reported to promote tumor cell proliferation, angiogenesis, and chemoresistance [29]. PRLR expression alone was associated with shorter survival in univariate analysis, but was associated with a longer survival in the final Cox proportional-hazard model. A previous study reported that PRLR activation increased proliferation, chemoresistance, and matrix metalloproteinase activity in GBM cells [30]. However, as PRL increases when the dopamine pathway is inhibited, infiltration to the pituitary stalk region may be the underlying cause of association with poor prognosis. Present results showed that DRC1, dynein regulatory complex subunit 1, expression is associated with short survival, which was partially in line with a previous study reporting that circRNAs derived from DRC1 were upregulated in ependymomas [31]. Regarding EPPK1, epiplakin1, it is generally known to participate in epidermal growth factor signaling and cell proliferation, as well as cytoskeleton reorganization. Our data showed that EPPK1 expression is associated with short survival, which is partially coherent with previous studies that have shown that EPPK1 expression activates cell proliferation in cervical cancer and esophageal squamous cell carcinoma [32,33]. As for RHOD, ras homolog family member D, our study showed that RHOD expression is associated with short survival, which is partially in accord with a previous study that reported that the RHOD promotor was differentially methylated between pituitary adenoma and normal tissue [34]. Lastly, ACP7, acid phosphatase 7, has not been previously reported to play a role in tumorigenesis or TMZ resistance, but according to the Human Protein Atlas, ACP7 is highly expressed in head and neck cancer and lung cancer [35].
CIBERSORTx results showed that the most abundant cell types in GBM were inflammatory microglial cells, dendritic cells, myeloid cells, and glioma stem cells. Glioma stem cells (GSCs) are pluripotent cells that lead to short survival and relapse of GBM, and are known to cause TMZ resistance through their slow mitosis [8]. Microglial cells are known to function as antigen-presenting cells (APCs), recognizing tumor cells and inducing cytotoxic T cells to kill tumor cells [36,37]. Recent studies focused on microglial cell polarization to convert microglial cells to an antitumor phenotype suggest its therapeutic potential [38]. Dendritic cells have been suggested to play roles in the inflammation of the brain [39]. The marker genes of dendritic cells were ARL4C, HLA-DQA1, IL1B, HLA-DQB1, and CD70, where HLA-DQ is an MHC class 2 surface receptor of APCs, IL1B is a lymphocyte activating cytokine, and ARL4C is a GTPase that regulates cell migration [40]. CD70 is known to be activated in mature dendritic cells and to play a key role in recurrent GBM cell aggressiveness and maintenance [41]. Myeloid cells are known to contribute to the GBM microenvironment by regulating immune and therapeutic responses [42]. Genes such as DRC1, which are known to be relevant to myeloid cells and to be upregulated in ependymomas, are worth further investigation for their prognostic value [31].
The limitations of this study include the lack of comparison between patients treated and not treated with TMZ, which would have strengthened the evidence for the specificity of the identified biomarkers. As TMZ is the first-line drug for GBM, most transcriptomic datasets are derived from the tissue of patients who have already undergone TMZ treatment. Additionally, the sample size of tissue analyzed in this study was small, and there was no functional validation of the identified genes. These shortcomings highlight the need for future research, which should include larger sample sizes, as well as functional validation of the biomarkers through experimental studies. Additionally, a comparison between groups of treated and untreated patients would provide further evidence for the biomarkers identified and their potential as a prognostic indicator.
Conclusions
In conclusion, this study identified that inflammatory microglial cells, dendritic cells, myeloid cells, and glioma stem cells are highly expressed in GBM tissue. Additionally, through the use of a deconvolution method and DEG analysis, this study also identified a gene signature consisting of ACP7, EPPK1, PCDHA8, RHOD, DRC1, ZIC3, and PRLR that are associated with poor prognosis in patients with GBM who were treated with TMZ. These findings may have potential implications for developing a diagnostic tool to predict TMZ resistance in GBM patients and optimize treatment decisions. However, further research is necessary to confirm these findings and to explore the underlying mechanisms that these cell types and genes play in GBM progression.
Supplementary Materials: The following supporting information can be downloaded at https: //www.mdpi.com/article/10.3390/jpm13020188/s1. Figure S1: survival information; Table S1a: demographic information of good, middle, bad survival groups of 35 patients; Table S1b: top five highly expressed genes of each cell type identified by CIBERSORTx; Table S1c: CIBERSORTx results; and Table S1d:
Conflicts of Interest:
The authors declare no conflict of interest. | 4,688.6 | 2023-01-20T00:00:00.000 | [
"Medicine",
"Biology"
] |
COINSTAC: Decentralizing the future of brain imaging analysis
In the era of Big Data, sharing neuroimaging data across multiple sites has become increasingly important. However, researchers who want to engage in centralized, large-scale data sharing and analysis must often contend with problems such as high database cost, long data transfer time, extensive manual effort, and privacy issues for sensitive data. To remove these barriers to enable easier data sharing and analysis, we introduced a new, decentralized, privacy-enabled infrastructure model for brain imaging data called COINSTAC in 2016. We have continued development of COINSTAC since this model was first introduced. One of the challenges with such a model is adapting the required algorithms to function within a decentralized framework. In this paper, we report on how we are solving this problem, along with our progress on several fronts, including additional decentralized algorithms implementation, user interface enhancement, decentralized regression statistic calculation, and complete pipeline specifications.
Introduction
Proliferating neuroimaging data present contemporary neuroscientists with both an exciting opportunity and a cumbersome challenge. The advantages of sharing data are clear. Adding datasets to a study increases sample size, making predictions more certain, and increases diversity, allowing differences between groups to be studied. Although there is indeed an abundance of data, there exist multiple barriers to fully leverage such data. Firstly, a significant amount of existing neuroimaging data has been collected without proper provisions for post hoc data sharing. Secondly, researchers must negotiate data usage agreements (DUAs) to collaborate and build models using multiple sources of data that can be anonymized and shared. Sharing data via a DUA is advantageous in that all the variables collected can be studied. However, these DUAs may require months to complete, and the effort to obtain them could be ultimately fruitless, as researchers only know the utility of the data after they have obtained and explored it. Thirdly, even if neuroimaging data can be shared in an anonymized form, the data require a copious amount of storage, and the algorithms applied to the data require significant centralized computational resources. Fourthly, even anonymized data bears a risk of reidentification, especially for subjects who are rare because of a combination of demographic and clinical data. While centralized sharing efforts are powerful and unquestionably should continue, the community needs a family of approaches to address all the existing challenges, including decentralized models that we describe in this paper. One alternative to centralized data sharing is to perform meta-analyses utilizing existing literature to avoid the burden of negotiating DUAs and storing and processing data (Thompson et al., 2017;Thompson et al., 2014). However, meta-analyses suffer from heterogeneity among studies caused by varying preprocessing methods applied to the data and inconsistent variables collected. In addition, metaanalytic results are not as accurate as those obtained from a centralized analysis.
The Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation (COINSTAC), proposed by Plis et al., in 2016(Plis et al., 2016, solves the abovementioned problems by providing a decentralized platform by which researchers can collaboratively build statistical and machine learning models, while neither transmitting their data nor sacrificing privacy concerns, thanks to differentially private algorithms. COINSTAC can run both meta-analyses and mega-analyses via "single-shot" and "multi-shot" (iterative) computations, respectively. The COINSTAC software (currently in an early prototype) is freely available, open source, and compatible with all major operating systems (Windows, Mac OS, and Linux). It is an easy-to-install, standalone application with a user-friendly, simple, and intuitive interface. By utilizing Docker containers, COINSTAC can run computations in any programming language (including Python, R, Matlab, FORTRAN, and C++) and is easily extensible. We are also building a development community to help users create their own computations, as well.
The use of a decentralized analysis framework has many advantages. For example, decentralized analysis can move beyond meta-analysis via iteration, obtaining a solution equivalent to that of the centralized result. In addition, one can move beyond sharing summary measures-which though plausibly private can still potentially be reidentified-to a more formally private solution. Differential privacy has been touted as a solution to the data sharing and reidentification problem. Developed by Dwork et al., 2006, this approach statistically guarantees privacy and allows for sharing aggregated results without the risk of reidentification (Dwork et al., 2006).
In the past few years, we have developed many algorithms that run in a decentralized and optionally a differentially private manner. et al., 2017), joint independent component analysis (ICA) (Baker et al., 2015), and two-level differentially private support vector machine (SVM) classification (Sarwate et al., 2014). To facilitate and accelerate algorithm development, we have created COINSTAC-simulator, which allows algorithm developers to prototype and troubleshoot their algorithms before deployment to real consortia in COINSTAC.
Furthermore, we include both input and output functionality to the COINSTAC user interface. For example, the interface for regression can accept data produced by FreeSurfer, with a menu to select the region of interest (ROI) in the brain that will be used as the dependent variable in the statistical analysis. Following the analysis, COINSTAC produces a statistics table for the output of ridge regression, which calculates the global p-values and t-values in a decentralized fashion for each site in the consortium, measuring goodness of fit.
COINSTAC also enables decentralized analyses with multiple computation steps. Easy and flexible computation stacking is a built-in feature in our framework. In this paper, we demonstrate an implementation scheme for specifying and managing multiple computations. With this framework, we can incorporate local computations, such as common preprocessing brain imaging tasks, into the analysis workflow.
A common nuisance among programmers and especially nonexpert users is the assembly of an environment to run a computer program. This is a crucial step that may require upgrading an operating system and downloading and installing the latest release of software, a compiler, or a supporting library. Assembly of the environment may involve permission from IT and a substantial amount of troubleshooting, which may lead to a long delay before analysis can begin. Additionally, inconsistent machine state between computers (including operating systems, libraries, and compilers) can lead to inconsistent results from the same computation.
A popular solution to this problem is utilizing a virtual machine (VM) that contains all the dependencies needed to run a program. Because VMs are resource-intensive, many developers have switched to using containers, which are an efficient, lightweight solution to the problem of heterogeneous development environments. Containers only bundle in the supporting software needed to run the program and do not require running a full VM with its own operating system. This reduces the required amount of memory and number of CPUs.
COINSTAC encapsulates individual computations inside Docker containers (https://www.docker.com/what-docker), which are run in series in a pipeline. Containers holding computations can be downloaded and run locally, which removes the need to assemble a development environment and thus greatly reduces the time to analyze results. This solution will also allow consortium participants to run coordinated preprocessing operations that must often occur before a statistical analysis, such as FreeSurfer processing or voxelbased morphometry. We have already created a Docker container with a standalone SPM package utilizing the Matlab Compiler Runtime. The normalization and coordination of preprocessing operations reduce heterogeneity in the data, creating a solid basis for the main analyses.
Methods and use cases
Algorithms for decentralized data analysis In our previous paper (Plis et al., 2016), we demonstrated the use of decentralized gradient descent in the optimization of a basic ridge regression model. This decentralized iterative optimization process represents an analysis of virtual data pooling. The resulting model generated in this manner is equivalent to the model generated in centralized repository analysis (i.e., the metaanalysis becomes a mega-analysis).
In this paper, we apply the decentralized gradient descent methods to other more advanced algorithms in the neuroimaging domain, including t-distributed nonlinear embedding (tSNE), shallow and deep neural networks, joint ICA, and IVA. These methods are already widely used in the neuroimaging domain, but have not previously been extended to work in a decentralized framework. We demonstrate how these methods can be computed within a decentralized framework and report the algorithm performance compared to a centralized analysis.
Decentralized tSNE (dSNE).
A common method of visualizing a dataset consisting of multiple high-dimensional data points is embedding the points into a 2-or 3-dimensional space. Such an embedding serves as an intuitive exploratory tool for quick detection of underlying structure of a dataset. In 2008, van der Maaten and Hinton proposed a method named tSNE to efficiently handle this situation (Maaten & Hinton, 2008). The embeddings produced by tSNE are usually intuitively appealing and interpretable, which makes this method an attractive tool in many domains, including neuroimaging (Panta et al., 2016).
We propose a method to embed a decentralized dataset that is spread across multiple locations such that the data at each location cannot be shared with others into a 2D plane. We build the overall embedding by utilizing public, anonymized datasets. The method is similar to the landmark achievements previously used to improve computational efficiency (De Silva & Tenenbaum, 2004;Silva & Tenenbaum, 2003). However, directly copying this approach does not produce accurate results, so we introduce a dynamic modification that generates an embedding that reflects relationships among points spread across multiple locations.
The detailed algorithm diagram for decentralized multi-shot tSNE is demonstrated in Figure 1. X p and X s represent the highdimensional site data and shared data, respectively. Y p and Y s represent the low-dimensional mapping site data and shared data, respectively. The master node initializes Y s and subsequently calculates a common gradient ∇Y s (j) based on the site gradient ∇Y sp (j) for each iteration j and update Y s , accordingly. Each local node will calculate the pairwise affinities among its own dataset and the shared dataset and then update Y p by locally calculating ∇Y P (j).
With this scheme, Y s stays constant across all sites for every iteration and serves as a reference function. Meanwhile, Y s is influenced by Y p , which allows local embedding information to flow across the sites, resulting in a final map with less overlapping.
We have tested the performance of this algorithm by comparing the decentralized result with that of centralized tSNE using the quality control metric of the ABIDE dataset (Di Martino et al., 2014). The results demonstrate that the centralized and decentralized computations generate an equal number of clusters. Additionally, random splits do not affect the stability of the clusters (Saha et al., 2017). Please see Figure 2 for reference.
Decentralized neural networks.
Recently, deep learning has gained increasing attention because of its excellent performance in pattern recognition and classification, including in the neuroimaging domain (Plis et al., 2014). To enable both shallow and deep neural network computations within COINSTAC, we developed a feed-forward artificial neural network that is capable of learning from data distributed across many sites in a decentralized manner. We utilize mini-batch gradient descent to average the gradient across sites. For our purposes, each batch contains one sample per site. We then average the resulting gradients from the batch. Figure 3 shows a flow chart of the decentralized neural network algorithm. As in a stochastic gradient descent (SGD) model, we calculate the error function Q p (W i ) for each site p and ith W. Q p (W i ) represent the discrepancy between the expected result Y i from the training set and the actual result from forward propaga- Each site then sends ∇Q p (W i ) to the master node, which averages the gradient and returns the result to the sites. Each site then updates W i on the basis of the mini-batch gradient decent equation until all training data are exhausted. With the same initialization W in the master node, we find that W i is always shared across all sites, but the change in W i at each iteration is determined by the data at each site.
We use a basic neural network known as a multilayer perceptron to demonstrate the decentralized computation process, but this framework can be easily extended to other types of neural networks. We tested the performance of this model using real functional magnetic resonance imaging (fMRI) data from smokers (Fagerström Test for Nicotine Dependence dataset) (Heatherton & Kozlowski, 1992) and found that the decentralized model and pooled centralized model yielded similar classification accuracy, which vastly et al., 2017). We randomly split the data into ten local and one reference dataset. The centralized results show ten different clusters. For three random splits of decentralized computation, we also obtain ten different clusters, and the number of clusters in the embedding is stable regardless of how the data are split among sites.
Page 5 of 20 outperformed the accuracy at local, isolated sites (Lewis et al., 2017). Please see Figure 4 for reference.
Decentralized joint ICA. When shared signal patterns are anticipated to exist among datasets, joint ICA (jICA) (Calhoun et al., 2006;Calhoun et al., 2001;Sui et al., 2009) presents a solution to combine and identify shared information over multiple datasets. Although originally proposed as a method for multimodal data fusion, jICA can also implement group temporal ICA of fMRI data. In both cases, datasets are concatenated (over modalities in multimodal fusion and over subjects across time in temporal ICA) and then jointly analyzed. The jICA model is particularly et al., 2017). In this experiment, we simulated an addiction dataset with two sites. The centralized classifier (red) and decentralized neural network classifier (yellow) perform similarly, and local sites classifiers (green and aquamarine) perform poorly.
attractive for datasets where the number of observations is significantly smaller than the dimensionality of the data, as in temporal ICA of fMRI data (time points < voxels), as concatenation over datasets effectively increases the number of observations. In decentralized jICA (djICA), the datasets are stored at different sites, rendering the traditional centralized approach for concatenation ineffective. To solve this problem, we developed an implicit concatenation procedure based on the assumption that the data from each site will share the same global unmixing matrix.
A diagram of djICA is shown in Figure 5. The global unmixing matrix includes W and bias b. Using this unmixing matrix, each site estimates the independent source Z p (j) and tries to maximize the entropy function of a sigmoid transformation of Z p (j) (Y p (j)). G p (j) and h p (j) are the local gradients for W and b, respectively. The master node sums the two gradients across all sites and updates the global unmixing matrix for the next iteration until either convergence or the stopping criteria is met.
The performance of djICA has been evaluated in studies by Plis et al (Plis et al., 2016) and Baker et al (Baker et al., 2015). The results of the experiments in these two studies convincingly demonstrate that with increased sample size the quality of feature estimation increases for both pooled-data ICA and djICA. Furthermore, we have found that splitting data across sites does not degrade the results given the same global data volume. Please see Figure 6 for reference.
Decentralized IVA. When using joint ICA to decompose temporal or multimodal datasets containing a group of subjects, we make a strong assumption that the underlying source maps are identical across subjects. Clearly, it is more desirable for source maps to contain subject-specific features. IVA is an approach that allows corresponding sources from different subjects to be similar rather than identical. IVA enables the subject source maps to contain unique information, yet still be linked across different subjects (Kim et al., 2006;Silva et al., 2016).
We proposed a decentralized IVA (dIVA) method, which allows multiple institutions to not only collaborate on the same IVA problem but also spread the computational load to multiple sites, improving execution time. We use IVA with a Laplace assumption for the dependence structure of the underlying source groups (Kim et al., 2006;Lee et al., 2008). Figure 7 shows a diagram of dIVA. Specifically, dIVA optimizes the same information measure as IVA by exploiting the structure of the objective function and fitting it into a decentralized computational model. In this model, a master node (or centralized aggregator) sends requests to local sites that contain the data. The sites send only data summaries (C p , d p ) back to the aggregator, which uses them to update a matrix of norms (C) as well as the objective function (cost(j)). The aggregator sends this matrix back to the sites, which use its inverse (C 0-1 ) to apply a relative gradient update on their local data. Subsequently, the local gradients are transmitted to the master node and aggregated to calculate a global step size (α). α is then returned to the local sites to update their weights. This process is orchestrated iteratively by the local and master nodes until convergence, and results are stored at local sites. Figure 7 shows the optimization function utilized by IVA can be split across sites, allowing the bulk of the computation to be parallelized with the aid of an aggregator that collects summaries from individual sites. We have already evaluated our decentralized approach on synthetic sources, and experimental results show that dIVA provides high accuracy and significantly reduces the runtime of the method compared with a centralized computation (Wojtalewicz et al., 2017). Please see Figure 8 for reference. (Baker et al., 2015). The experiment is based on synthetic functional MRI data using a generalized autoregressive conditional heteroscedastic model (Engle, 1982;Bollerslev, 1986). The top figure shows that as the global number of subjects increases, the Moreau-Amari index (MAI) decreases for both pooled-data ICA and djICA with different principal component analysis (PCA) operations. Additionally, MAI converges for pooled-data ICA and djICA when the number of subjects increases. The bottom figure shows that number of splits in the data have no effect on MAI.
Improved COINSTAC user interface (UI)
We have improved the UI for COINSTAC by adding features that facilitate the input of brain imaging data, allow users to easily run computations, and keep users informed on the progress of the computation. To begin a collaborative, decentralized computation, a group of users that will participate in the analysis, called a consortium, must be created. This involves naming the consortium, choosing the computation, and defining the dependent and independent variables. The user who completes these steps is called the consortium owner. As shown in an example in Figure 9, the UI accepts FreeSurfer data saved in a comma-separated value (CSV) file as an input. The ROI of the brain computed by FreeSurfer is selected as the dependent variable in a ridge regression computation. Additionally, the regularization parameter (lambda), which limits overfitting in the model, is selected via a numeric field. A standard regression with no regularization is performed if lambda is given a value of zero.
Next, the consortium owner declares the covariates (independent variables) and determines their types. The UI currently allows either Boolean (True/False) or numeric covariates. Every user who participates in the consortium must then choose a local data source, such as a FreeSurfer CSV file, and map the columns in the file to the variables declared by the consortium owner. Figure 10 shows how this is accomplished in the UI.
Once all the participants in the consortium have mapped columns in their local data sources to declared variables, the computation commences. The progress of computations in multiple consortiums is displayed on the Home tab of the UI. Figure 11 shows an example of this. In the top computation, a multi-shot ridge regression is on the third iteration out of a maximum of 25 iterations.
New output statistics table with decentralized statistics computation for ridge regression Regression analysis generates an equation to describe the statistical relationship between one or more predictor variables and the response variable. Decentralized ridge regression first produces the regression coefficients for all independent variables through an iterative optimization process. However, in most cases, a researcher may not only want to know the coefficient associated with certain regressor but also the statistical significance of this coefficient and the overall goodness of fit or coefficient of determination (R 2 ) for the global model. In order to generate a standard statistical output accompanying the coefficient as in many major statistical tools, we developed a decentralized approach to calculate the t-values and goodness of fit for the global model without sharing any original data.
The decentralized R 2 calculation is demonstrated in Figure 12. First, each local node calculates the local average of dependent variable p Y and transmits it and the size of dataset N p to the master node. Then, the master node calculates the global Y and Page 11 of 20 returns it to the local node. Subsequently, every node calculates the local total sum of squares (SST p ) and sum of squared errors (SSE p ) on the basis of Y and send them to the master node. Finally, the master node aggregates SST p and SSE p across all sites to calculate the global value of R 2 .
The decentralized t-value calculation is demonstrated in Figure 13. Each local node calculates the local covariance matrix of X p and SSE p and transmits them and data size N p to the master node. The master node then aggregates cov(X p ) to generate the covariance matrix of global covariates X to allow the following calculation of the t-values. MSE represents the mean squared error of the estimated coefficient W (or β).
After generating the t-value for every covariate and intercept, we use the public distributions library on npm (https://www.npmjs. com/package/distributions) to generate the Student's t-distribution and then calculate the two-tailed p-value for corresponding t-value. Figure 14 shows an example statistical output table for ridge regression. The COINSTAC UI displays the result with summarized This output is generated using simulated freesurfer brain volume data. In the simulation, the intercept part (β 0 ) was set to a fixed amount (48466.3 for Right-Cerebellum-Cortex); the age effect(β 1 ) was selected randomly from range [-300, -100] and group(isControl) effect(β 2 ) was selected randomly from range [500, 1000] for each pseudo subject; the standard unit Gaussian noise multiplied by random index ranged from 1800 to 2200 was added subsequently.
Page 13 of 20 consortium information at the top. In the output table, we first present the global fitting parameters, following by the fitting parameters locally calculated at each site. The COINSTAC UI also provides the detailed covariate name for each β.
Complete pipeline specification COINSTAC is not only designed to apply individual computations, but also to flexibly arrange multiple computations into a pipeline. Both decentralized analyses and local preprocessing steps can be included in a pipeline. The goal of COINSTAC is to provide a shared preprocessing script that is convenient for researchers and minimizes the data discrepancies across sites that become inputs to decentralized computations.
COINSTAC concatenates multiple computations into a pipeline and uses a pipeline manager to control the entire computation flow. Figure 15 shows a pipeline specification scheme with an initial preprocessing step and a following decentralized computation. Consortium owners will be able to select the computation step and output type through connected dropdown menus. After the computation steps have been selected, all users within a consortium will be shown cascading interfaces to upload input data and set hyperparameters for each computation. Additionally, the input from the latter computation step can be linked to the output from an earlier computation step.
Once a complete pipeline has been formed, all pipeline information is transmitted to the pipeline manager. Figure 16 shows how the pipeline manager interacts with a pipeline and its internal computations. The pipeline manager controls the entire computation flow. It is responsible for piping the input data to the first computation step, caching and transferring intermediate computation output, and The pipeline manager handles the input and output of each pipeline, providing a conduit other nodes in the network. Each computation has its own schema that describes the names and types of its input and output parameters. Controllers are used to manage specific behavior in each computation in the pipeline. Each computation is encapsulated in a Docker container to improve portability among development environments.
storing the final pipeline output. An intermediate controller is added to provide fine-grained control for monitoring the iterative process between local and remote nodes for every computation. The computation schema is defined by a JavaScript object notation (JSON) structure and includes input and output specifications. A Docker container is used to encapsulate an individual computation block.
Discussion
In this paper, we reviewed our progress on the development of decentralized algorithms that can be implemented on the COINSTAC platform. Every algorithm is structured similarly in that the local gradient of the objective function is transmitted to the master node, and the master node either returns a common averaged gradient or a step size (dIVA) to update the local weights. This scheme guarantees that information is shared across all sites on every iteration in the optimization algorithm to achieve a virtually pooled analysis effect (i.e., a mega-analysis). This framework also facilitates differential privacy by allowing for the addition of noise to each local objective function. We continue to develop decentralized algorithms as described below.
Future decentralized algorithms Decentralized network gradient descent. SGD has emerged as the de facto approach to handle many optimization problems arising in machine learning, from learning classification/regression models to deep learning (Bottou, 2010;Song et al., 2013). For decentralized settings, SGD can be costly in terms of message complexity. We are currently developing approaches to limit this message complexity to enable a variety of statistical learning methods within COINSTAC. These approaches are guided by theory, but will involve developing task-specific heuristics to tune the algorithm parameters.
Nonnegative matrix factorization (NMF). NMF is another popular method for discovering latent features in data such as images, where measurements are all nonnegative (Lee & Seung, 2001). Although there has been significant work on NMF and its variants, the work on decentralized implementations is more limited, and the focus has been on improving parallelism for multicore systems (Potluru et al., 2014). Because of the message-passing nature of the COINSTAC architecture, we are developing decentralized and accelerated NMF algorithms that are optimized with gradient descent. Further extensions could allow users to find an NMF to minimize a variety of cost functions beyond squared error.
Canonical correlation analysis (CCA). One challenging task in learning from multimodal or multiview data is to find representations that can handle correlations between the two views (Sui et al., 2012;Thompson, 2005). CCA is one such method. We are currently developing privacy-preserving CCA methods, as well as determining whether decentralized, message-passing approaches will be feasible within the COINSTAC architecture.
Integration with large-scale collaborative frameworks In recent years, the ENIGMA Consortium has conducted collaborative meta-analyses of schizophrenia (van Erp et al., 2016) and bipolar disorder (Hibar et al., 2017), in which subcortical brain volumes and cortical thicknesses were compared between patients and controls, respectively. In these studies, many univariate linear regression models were created in parallel to examine group differences for different regions of the brain. ENIGMA distributes analysis software to many sites and aggregates the results to conduct a meta-analysis. The upcoming version of COINSTAC will facilitate such studies by allowing researchers to specify models that contain combinations of selected dependent and independent variables. Table 1 elaborates on this point by showing an example in which a researcher selects a group of dependent variables (right and left cerebellum cortexes) and a group of independent variables (age and isControl). One model is computed separately for each combination of dependent and independent variables. The advantage of COINSTAC is that dissemination of software and aggregation of results will be handled by our software, eliminating many manual steps. In addition, as mentioned earlier, COINSTAC enables us to run multishot regression (hence converting a meta-analysis into a mega-analysis). Finally, COINSTAC opens up the possibility of running multivariate analysis (such as SVM (Sarwate et al., 2014) or IVA), as well as incorporating differentially private analyses, which would significantly extend the current ENIGMA approach, while also preserving the powerful decentralized model. Author contributions JM helped design the architecture of COINSTAC, reviewed the decentralized algorithms, developed the statistic output table, wrote the initial draft of paper, and coordinated writing. EV was the overall technical lead, managed the COINSTAC project, and contributed to writing and proofreading the paper. AS helped develop the differentially private algorithms and additional decentralized algorithms. RK provided the pipeline specification graph and was heavily involved in COINSTAC implementation. CR and TK contributed to the detailed COINSTAC implementation. RS helped with the decentralized algorithm review. S.Panta contributed to the brain imaging data preprocessing pipeline. JT provided input on functionality aspects and served as a beta tester for COINSTAC. S.Plis proposed the decentralized data analysis system and led the algorithm development effort. VC led the team and formed the vision. All authors helped edit the manuscript. | 6,735.4 | 2017-08-18T00:00:00.000 | [
"Computer Science",
"Medicine"
] |
Research On the Pricing Model of Second-Hand Sailboats Based on GDBT Model
. Sailboats have varying values with changes in market conditions and aging. This paper analyzed the characteristics and economic characteristics of monohull along with catamaran in their respective regions. The paper also carried out Spearman correlation analysis on their listing prices. The gradient boosting decision tree model was established to calculate the feature importance of each variable on the listing price, and the accuracy of the model reached over 86%. It indicated that the model can effectively price the second-hand sailboat market. This paper used one-way ANOVA to determine whether there was a significant difference between the listing prices of monohull and catamaran in the secondary markets. There were significant differences in the listing prices of monohulled sailboats and catamarans in different geographical regions. The listing prices of monohull and catamaran in different geographical regions were significantly different, and the regional effects were inconsistent.
Background
Shipowners are more interested in second-hand ships than in the market for new ships due to considerations such as budget and delivery time. The deviation of the value judgment of the two parties to the ship trading leads to poor liquidity in second-hand sailboat trading, so the final sailboat trading price is usually determined after a sailboat broker's inspection. In terms of sailing brokers, they need to consider the value of second-hand sailing boats, investigate the models and materials of second-hand sailing boats in different regions. The sailing brokers evaluate the price of second-hand sailing boats based on various indicators that may affect their price, and make an accurate evaluation of their price to promote market trading.
Literature Review
Goulielmos [1] pointed out that the Greeks and a few maritime countries were aware of the importance of the second-hand shipping market and concluded that shipowners could predict the price of second-hand ships, during which time they could choose the highest or lowest price. Liang FANG et al. [2] studied the dynamic relationship between price and trading volume in the second-hand ship market based on the VAR-GARCH model, and concluded that there is a unilateral causal relationship between past trading volume and volatility in the return rate of all second-hand ship markets. Roar Adland [3]explored whether energy efficiency will affect the value of ships in the second-hand market, concluded that there was a negative correlation between energy efficiency and sales prices, with an elastic value of around 0.4, indicated that the actual operating energy efficiency depends on the speed of ships. Lim S S [4] used artificial neural network models outperform simple stepwise regression analysis and satisfied both statistical soundness and accuracy of results inscientific models. Floriano [5] established a model of the relationship between shipbuilder countries and second-hand ship prices and concluded that ships built in Japan and Europe often receive higher prices than other countries. Azhar A [6] builded a cost estimation model of a used ship to determine the price of a used ship andbuilds a cost estimation model of a used ship. The data was processed and analyzed by multiple methods including market price method, comparative ship and physical pricing method, and the estimated price or appraised ship is obtained from the average of the three methods. Nam H S [7] argued that the prices of used and new ships were closely related and influenced by market dynamics. The results showed that the ship's main engine type and country of construction were statistically significant in most ship categories, while other ship specific and economic factors also explained the value of second-hand ships consistent with the literature.
We possessed a batch of second-hand sailboat data. We used useful predictive factors such as the characteristics of specific sailboats and regional annual economic data to explain the search price for each sailboat. The accuracy of price estimation for each type of sailboat was discussed. The impact of regions on pricing was explained based on the established model. Then we discussed the impact of regions on sailboat prices and determined whether the regional effects were consistent for all sailboat species.
Main Work
We conducted Spearman correlation analysis between the characteristics of the two types of sailboats and the listing price respectively to determine whether the variables were significantly correlated with the listing price. The gradient boosting decision tree model was established to calculate the characteristic importance of each variable to the listing price, and to evaluate the accuracy of the model results.
We used one-way ANOVA to explore whether there was a significant difference between the listing prices of the two types of sailboats in the secondary market in different regions. Sailboat characteristics and economic data from different regions were substituted into the GBDT model developed to discuss the effect of different regions on listing prices. Kruskal-Wallis tests were conducted separately for the listing prices of the two sailboat varieties in different regions to discuss the effect of region on sailboat prices and to determine whether the region effect was consistent for all sailboat varieties.
Arrangement
This paper was organized as follows. Section I introduced the research background, literature review, research problem and analysis of this paper. Section II presented the principles and assumptions of the three models used in this paper: GBDT, one-way ANOVA, and Kruskal Wallis test principle, and proposed the basic assumptions of this article. Section III applied the GBDT model to explain the listing price of each sailboat and discussed the accuracy of price estimation. Section Ⅳ explained the impact of different regions on the price based on the establishment of the GBDT model, and judged whether the regional effect is consistent for all sailboats. Section V summarized the research questions, methods, results and implications of this paper.
Gradient Boosting Decision Tree
The gradient boosting decision tree algorithm is one of the more advanced machine learning strategies available [8]. The GBDT algorithm has excellent performance in regression and classification problems.
One-way Analysis of Variance
One-factor ANOVA is an analysis of one category of independent variables on numerical variables. Also known as one-way ANOVA, analysis of whether one factor has a significant effect on the outcome [9].
Kruskal-Wallis Test
The Kruskal-Wallis test is essentially a generalization of the Mann-Whitney U test for two independent samples to multiple independent samples, and is used to test whether the distributions of multiple aggregates are significantly different [10].
Assumption
We made the following reasonable assumptions and conditional constraints based on the actual situation to construct a more accurate mathematical model. Hypothesis 1: The pricing of used sailboats is related to the market environment in their region, excluding other unrelated factors. The price of a sailboat is influenced not only by the characteristics of the sailboat itself, but also by the supply and demand in the market.
Hypothesis 2: The selling price of a sailboat has a lot to do with its geographical location. Because market conditions and supply and demand vary geographically, this has an impact on the selling price.
Hypothesis 3: It is assumed that regional economic indicators can be reflected by GDP and GDP per capita.
Hypothesis 4: The data provided in the topic are true and reliable to a certain degree. Because the model we built is based on the data provided in the topic, Only the high validity of the data can guarantee the high reliability of the model.
Notations
The symbol description of the paper was shown in Table 1.
Spearman Correlation Analysis
We conducted Spearman correlation analysis of year, characteristics, regional economic data and listing price for used monohulled sailboats and used catamarans, respectively. The obtained correlation coefficients were shown in Table 2. These variables were significant at the 1% level, indicating that these variables explained the listing price of each used sailboat.
Modeling Steps
Step1: A gradient boosting decision tree (GBDT) regression model was established using training set data.
Step2: The feature importance was calculated using a gradient boosting decision tree established.
Step3: The established gradient boosting decision tree (GBDT) regression model was applied to training and testing data to obtain model evaluation results.
Modeling Process
Given second-hand sailboat listing price data training set: The strong learner expression of the model is
Result of the Model
We used gradient boosting decision tree (GBDT) model to regress the data, and Figure 1 showed the characteristic importance ratios of the respective variables for monohulled sailboats and catamaran. Overall, displacement was the most important factor affecting the listing price of monohulled sailboats, with an importance of 32.30%, followed by year at 18.20%. The impact of other factors was less than 10%. This suggests that the price of monohulled sailboats is mainly influenced by displacement and year, with these two factors accounting for over 50% of the cumulative importance.
For catamarans, beam was the most important factor affecting the listing price, with an importance of 36.70%, followed by year at 26.30% and LWL at 11.70%. The impact of other factors was less than 10%. It suggested that the price of catamarans was mainly influenced by beam and year, with these two factors accounting for over 60% of the cumulative importance.
From this perspective, it was clear that the price formation of second-hand sailing boats in a given year is not negligible, as the price of sailing boats changes with their aging. For monohulled sailboats, however, displacement was the most important factor influencing price formation, while for catamarans it was beam.
Model Evaluation
After repeated iterative training of GBDT model, and 2 evaluation indexes of the model were calculated based on the five-fold cross validation test set. Table 3, MAPE (mean absolute percentage error) is a percentage value. The smaller the value, the more accurate the model is. Compared with the predicted value when only the mean is used, the closer the R² value is to 1, the more accurate the model will be, indicating that the model has a good fitting effect. The R² value of the training set of monohulled sailboat and catamaran was as high as 97%, and the R² value of the test set was more than 86%, indicating that the model has a good fitting effect and a high accuracy in estimating the price of each sailboat. Figure 2 showed the prediction of Gradient Boosting Decision Tree (GBDT) on the test data.
Figure 2: Test Data Prediction Diagram of Monohulled Sailboat and Catamaran
As can be seen in Figure 5, the curves of the true and predicted values were very close to each other and predict the test data very well. It further indicated that the gradient boosting decision tree model had high accuracy in estimating the prices of each sailboat.
The Effect of Region on Listing Prices
We conducted one-way ANOVAs on the listing prices of second-hand monohulled sailboats and catamarans separately for geographic regions. The results were shown in Table 4. 407738.637 155021.207 Note: ***, **, * represent 1%, 5%, 10% significance levels, respectively It can be found that among used sailboats, the p-value of the results of the one-way ANOVA on the listing price of used sailboats in Europe, the United States and the Caribbean was less than 0.05. The results were statistically significant, indicating that there were significant differences in the listing price of single-hull sailboats in different geographic regions.
We entered the data for second-hand monohulled sailboats and second-hand catamarans from different regions (USA, Europe, Caribbean) into the mathematical model we developed and obtained the proportional importance of the characteristics of the respective variables for monohulls and catamarans from different regions.
The model evaluation results obtained using the GBDT model were shown in Table 5. The value of the training set R² and the value of the test set R² of the sailing vessels in the three regions were all over 0.97 and 0.82 respectively, indicating that the fitting effect of the model was good and the price estimation of each sailing species in the three regions was relatively accurate.
We used the GBDT model to regression the data to get the feature importance. As shown in Figure 3, in the United States, the importance of Displacement for a monohulled sailboat increased from 32.30% to 41.60%. This was because displacement was one of the most important measures of sailboat size and carrying capacity. The displacement of a sailboat was directly proportional to its carrying capacity. Catamaran length and price were generally positively correlated. It was because longer hulls could provide more space, more facilities and better performance. In terms of both monohulled sailboats and catamarans, sail area became more important. It was because sailboats were powered by wind caught by their sails. Therefore, whether a sailboat could go faster or not played a decisive role. In cases where wind direction was correctly mastered, sail area could be used to obtain greater sail power and maintain smooth sailing.
As shown in Figure 4, the importance of the price characteristics of European second-hand sailboats was basically the same as that of the whole. The importance of Draft features had increased to a certain extent for monohulled sailboats, because draft refers to the depth of the hull in the water. Draft and hull weight were positively related, so the price will increase accordingly. Generally speaking, an increase in GDP per capita will increase consumption levels in the region, so it will also had an impact on the price of monohulled sailboats. For catamarans, reasons for the increase in importance of Sail Area and Length characteristics were similar to those in the United States.
As shown in Figure 5, the importance of year features had increased significantly for monohulled sailboat in Caribbean. Year was an important factor in the price of monohulled sailboats. The price of a new sailboat was higher than that of an old one. In general, the gross domestic product of a region may have affected the price of a single sailboat. If the GDP of a region was higher, then people in that region may have had more money to buy expensive monohulls. So the price of monohulls may have been higher. Beam and length features were increasingly important for catamarans. Because catamarans were wider, which could lead to higher mooring charges and taxes. Typically, docks charged 1.5 to 2 times the docking rate for catamarans. The length of a catamaran was also proportional to the price. It was because longer hulls could provide more space, more facilities and better performance. The Kruskal-Walli's test was performed on the listed prices of monohulled sailboats and catamarans in different regions, and the results of the test analysis were shown in Table 6. The Kruskal-Walli's test results showed that based on the variable listing price, the p-value of the test was 0.000 less than 0.05 in different regions and therefore statistically significant. This indicated that there was a significant difference in the listing price of different varieties in the same region and the regional effect was inconsistent across all sailing varieties.
Among the monohulled sailboats, the Cohen's f values for the magnitude of difference were 0.821, 0.63, and 0.817 for the United States, Europe, and the Caribbean, respectively, with a large degree of difference. Among the catamarans, the Cohen's f values for the United States and the Caribbean were 0.922 and 0.541, respectively, with a large degree of difference, while the Cohen's f values for Europe were 0.399, with a moderate degree of difference.
Conclusion
This paper studied the pricing of second-hand sailboats. We used Spearman correlation analysis to determine whether the characteristics of sailboats and regional economies were significantly related to the listing price. It was found that the two were significantly related. We established the gradient boosting decision tree model to price the second-hand sailboat, with an accuracy of over 86%. When determining the price of second-hand sailboats, the manufacturing year cannot be ignored. For Monohulled sailboats, displacement and width were the most important price forming factors. For catamarans, length and width were the most important factors. We used one-way ANOVA model to explore whether there was a significant difference in the listing prices of monohulled sailboats and catamarans in different geographical regions. There were significant differences in the listing prices of monohull and catamaran in different geographical regions, and the regional effects were inconsistent. | 3,767.4 | 2023-07-24T00:00:00.000 | [
"Economics"
] |
An Experimental Multi-Target Tracking of AM Radio-Based Passive Bistatic Radar System via Multi-Static Doppler Shifts
This paper presents a description of recent research and the multi-target tracking in experimental passive bistatic radar (PBR) system taking advantage of numerous non-cooperative AM radio signals via multi-static doppler shifts. However, it raises challenges for use by multiple spatially distributed AM radio illuminators for multi-target tracking in PBR system due to complex data association hypotheses and no directly used tracking algorithm in the practical scenario. To solve these problems, after a series of key array signal processing techniques in the self-developed system, by constructing a nonlinear measurement model, the novel method is proposed to accommodate nonlinear model by using the unscented transformation (UT) in Gaussian mixture (GM) implementation of iterated-corrector cardinality-balanced multi-target multi-Bernoulli (CBMeMBer). Simulation and experimental results analysis verify the feasibility of this approach used in a practical PBR system for moving multi-target tracking.
Introduction
Passive bistatic radar (PBR) is a subset of bistate radars receiving non-cooperative transmitters of opportunity scattered by potential targets. Research on PBR has attracted extensive attention because of well-known advantages, such as no additional frequency channel allocation, lower costs, and lower probability of being detected with respect to active radars. Although, PBR systems have a long history, there are not enough operational systems.
Of all the transmitters of opportunity available in the PBR systems, very high frequency/ultra high frequency (VHF/UHF) bands represent some of the most attractive for surveillance purposes, such as analog television (ATV) [1,2], digital television-terrestrial (DTV) [3,4], frequency modulation (FM) radio [5], digital audio/video broadcasting (DAB/DVB) [6][7][8][9][10]. However, relatively little interest has been shown in the high frequency (HF) band due to the propagation complexity and low range resolution. Especially, the external illuminators in the HF band have excellent range coverage, propagation over the horizon, and stealth target detection. Some preliminary and pioneering HF-PBR works have been carried out. In the PBR system [11,12], Thomas et al. from University College London performed an analysis using the HF digital radio Mondiale (DRM)signal as transmitter of opportunity. The two-dimensional target localization, using a linear frequency modulated continuous waveform from a non-cooperative OTH radar located in Longreach, Australia, with a bandwidth of 10 kHz, is presented in [13].
In fact, compared with other opportunity illuminators in the HF band, commercial amplitude modulation (AM) broadcast signal sources have the advantages of high transmitter power, larger numbers, and wider coverage. Due to the propagation complexity and bandwidth limited, little attention is paid to AM radio signal for PBR system. The research lowest cost architectures; therefore, the main contribution of this paper is that the proposed multi-target tracking method can provide a reference for similar PBR systems.
The rest of this paper is organized as follows. The description of the AM Radio based PBR system and multi-target tracking formulation are described in Section 2. Section 3 provides the proposed IC-UT-GM-CBMeMBer filter. Simulation and field experimental implementation are given in Section 4. Finally, conclusion and possible future directions are drawn in Section 5.
System Description
Supposing the ionosphere is homogeneous and spherically symmetric, the bistatic plane geometry of an AM-radio-based sky-surface wave PBR system in the scenario can be simplified as shown in Figure 1 (two transmitters are shown explicitly simplistically). In the system, we exploited one receiving antenna array approximately 20 m above ground level located over-the-horizon (farther than 1000 km from the noncooperative transmitters generally) at Hubei province of China, which is equipped with a uniform circular array (UCA) with 16 antennas.
IC-UT-GM-CBMeMBer filter. The system described in this paper was constructed on one of the simplest and lowest cost architectures; therefore, the main contribution of this paper is that the proposed multi-target tracking method can provide a reference for similar PBR systems.
The rest of this paper is organized as follows. The description of the AM Radio based PBR system and multi-target tracking formulation are described in Section 2. Section 3 provides the proposed IC-UT-GM-CBMeMBer filter. Simulation and field experimental implementation are given in Section 4. Finally, conclusion and possible future directions are drawn in Section 5.
System Description
Supposing the ionosphere is homogeneous and spherically symmetric, the bistatic plane geometry of an AM-radio-based sky-surface wave PBR system in the scenario can be simplified as shown in Figure 1 (two transmitters are shown explicitly simplistically).
In the system, we exploited one receiving antenna array approximately 20 m above ground level located over-the-horizon (farther than 1000 km from the noncooperative transmitters generally) at Hubei province of China, which is equipped with a uniform circular array (UCA) with 16 antennas.
The direct wave (emitter-to-receiver) and the illumination wave (emitter-to-target) are reflected from the ionosphere, while the echo wave (target-to-receiver) was via lineof-sight (LOS) propagation in the surveillance area. To obtain multi-static Doppler measurements in the PBR system, some key techniques in array signal processing are summarized in a block diagram sketched in Figure 2. Similar to traditional passive radar, surveillance and reference channels are needed to receive target echoes and reference signal, respectively. The DOA of direct wave can be obtained by the multiple signal classification (MUSIC) algorithm from the reference channel, which is then used to clean the reference signal by using conventional beamforming (CBF) technology. After scanning the surveillance channel by normalized least mean square (NLMS) technology to obtain the echo signal, we calculated the cross-ambiguity function (CAF) of the direct path signal and the scattered signal to estimate range vs. Doppler shift of the targets. Finally, the time delay and Doppler shift of the targets after clutter removal were estimated. The direct wave (emitter-to-receiver) and the illumination wave (emitter-to-target) are reflected from the ionosphere, while the echo wave (target-to-receiver) was via line-of-sight (LOS) propagation in the surveillance area.
To obtain multi-static Doppler measurements in the PBR system, some key techniques in array signal processing are summarized in a block diagram sketched in Figure 2. Similar to traditional passive radar, surveillance and reference channels are needed to receive target echoes and reference signal, respectively. The DOA of direct wave can be obtained by the multiple signal classification (MUSIC) algorithm from the reference channel, which is then used to clean the reference signal by using conventional beamforming (CBF) technology. After scanning the surveillance channel by normalized least mean square (NLMS) technology to obtain the echo signal, we calculated the cross-ambiguity function (CAF) of the direct path signal and the scattered signal to estimate range vs. Doppler shift of the targets. Finally, the time delay and Doppler shift of the targets after clutter removal were estimated. Furthermore, the greatest limitation on tracking target performance in the self-developed experimental PBR system is the interference and clutter in the received signal, for example, dense direct path interference and the DOA of echo signal with very low SNR under masking effects. Although the classic DOA estimation and clutter suppression methods have been studied over the past decade [22,23], most of them are unsuitable for processing HF-AM radio signal. We adopted the method of reference [24] by building a single-snapshot virtual array signal. After extending the single-snapshot virtual array signal to multi-snapshots and the MUSIC algorithm, the clutter interference could be suppressed significantly, and the desired echo signal was enhanced simultaneously. More details of signal processing and improvement can be found in [24]. Finally, the excellent Doppler shift information of targets' corresponding time can be provided on a 2-D time vs. Doppler map.
CBMeMBer Filter
The CBMeMBer filter is first introduced to solve the aforementioned tracking problem in the PBR system in this section [21].
Similarly, there are k N measurements , each taking values in an observation space at time k. In addition, the received measurement also contains a set of missing alarms or clutter that can be modeled as a Poisson RFS k Κ . Thus, multi-target observation at time k + 1 is modeled as finite sets [26]. Furthermore, the greatest limitation on tracking target performance in the self-developed experimental PBR system is the interference and clutter in the received signal, for example, dense direct path interference and the DOA of echo signal with very low SNR under masking effects. Although the classic DOA estimation and clutter suppression methods have been studied over the past decade [22,23], most of them are unsuitable for processing HF-AM radio signal. We adopted the method of reference [24] by building a single-snapshot virtual array signal. After extending the single-snapshot virtual array signal to multi-snapshots and the MUSIC algorithm, the clutter interference could be suppressed significantly, and the desired echo signal was enhanced simultaneously. More details of signal processing and improvement can be found in [24]. Finally, the excellent Doppler shift information of targets' corresponding time can be provided on a 2-D time vs. Doppler map.
CBMeMBer Filter
The CBMeMBer filter is first introduced to solve the aforementioned tracking problem in the PBR system in this section [21].
At time k, there are N(k) target states X k = x k,1 , · · · , x k,N(k) ⊆ F (χ), which denote space of finite subsets of χ. Given a target x k at time k, it is either detected in the surveillance area with probability p D,k (x k ) and generates a Bernoulli RFS Θ k (x k ) with likelihood function g k (·|x k ) , or it is missed with probability 1 − p D,k (x k ). Given a multi-target state X k , each x k ∈ X k either continues to exist at time k + 1 with probability p S (x k+1 ) and moves to a new state x k+1 with target transition equation f k+1|k (x k ) or dies with probability 1 − p S,k (x k+1 ). Thus, given a target with state x k ∈ X k at time k, its behavior time k + 1 is modeled by the Bernoulli RFS S k+1|k (x k ), and Γ k+1 denotes the multi-Bernoulli RFS of new births at time k + 1. The multi-target state is modeled as [25].
Similarly, there are N k measurements Z k = z k,1 , . . . , z k,N k , each taking values in an observation space at time k. In addition, the received measurement also contains a set of missing alarms or clutter that can be modeled as a Poisson RFS K k . Thus, multi-target observation at time k + 1 is modeled as finite sets [26].
where Θ(x k+1 ) is a Bernoulli RFS that is generated by target state x k+1 ∈ X k+1 . A multi-Bernoulli RFS X (i) on χ is a union of a fixed number M of independent Bernoulli RFSs with existence probability r (i) and probability density p (i) , X = ∪ M i=1 X (i) . Moreover, the probability density π is [27]: The multi-target bayes recursion propagates in time [27]: where f k+1|k (·|·) is the multi-target transition density and g k+1 (·|·) is the multi-target likelihood.
Multi-Target Tracking Model
Tracking model is one of the major problems needing to be considered in the multitarget system. In this paper, we consider the target tracking scenario performed in a 2D Cartesian coordinate, with the origin point located at a single receiver antenna array; the x-axis points East and the y-axis points North. Assume that at time k, the i-th target state is represented by the state vector x is the number of targets, superscript T denotes the matrix transpose, p are the position and velocity of the target, respectively. Each target dynamic motion is followed by a nearly constant velocity model: where u k ∼ N(u; 0, Q k ) is zero-mean white Gaussian process noise with covariance Q k . We adopt: In which ∆ is the sampling interval. A 0 = 1 ∆ 0 1 , I n , and 0 n denote n × n identity and zeros matrices, respectively. In the two-dimensional surveillance area, three spatially distributed non-cooperative AM radio illuminators constantly transmit signals with a known carrier frequency f (i) c of the i-th AM radio station, i = 1 . . . 3, and the receiver places are at the original point, as illustrated in Figure 3. The direct wave and scattered waves from multiple AM broadcast stations reflected from the ionosphere (three scattered echo waves and one target are shown simplistically) reach the target and receiver simultaneously on the condition that the AM broadcast stations located at R are far away (>1000 km) from the receiver antenna array. Doppler shift measurements by i-th illumination can be generally modeled as [28]: where: → v k is the constant velocity vector of target at time k, p k = p x,k 2 + p y,k 2 , c is the is the constant velocity vector of target at time k, Multiple non-cooperative illuminators of AM radio stations (red pentacle) reflected from ionosphere received by single receiver antenna array (blue triangle) in x-y coordinate.
As the AM radio stations are typically long distance from the receiver antenna array, we approximately considered the direction of the direct wave from the i-th illuminator of AM radio station to the receiver antenna array as equal to the direction of the scattered wave from the i-th illumination to the target. According to Equation (9), the model can be approximately written as: Here, is the normalized target position relative to the receiver.
is the normalized incident direction vector of the direct wave from the i-th illuminator of the AM radio station to the receiver antenna array, which is independent on the target state and could be easily achieved from DOA estimation of the direct wave, as mentioned before.
Each Doppler-shift subset includes at most one measurement per illuminator and corresponds to the measurements made by multi-targets across all illuminators practically contaminated by false alarms and misdetections. The subsets of a partition are disjointed and comprise measurement space, which is denoted as As the AM radio stations are typically long distance from the receiver antenna array, we approximately considered the direction of the direct wave from the i-th illuminator of AM radio station to the receiver antenna array as equal to the direction of the scattered wave from the i-th illumination to the target. According to Equation (9), the model can be approximately written as: Here, → p k (i) is the normalized target position relative to the receiver. → vt k (i) is the normalized incident direction vector of the direct wave from the i-th illuminator of the AM radio station to the receiver antenna array, which is independent on the target state and could be easily achieved from DOA estimation of the direct wave, as mentioned before.
Each Doppler-shift subset includes at most one measurement per illuminator and corresponds to the measurements made by multi-targets across all illuminators practically contaminated by false alarms and misdetections. The subsets of a partition are disjointed and comprise measurement space, which is denoted as Z k is the number of the detection values, including false alarms and misdetection; z
The Proposed Multi-Target Tracking Method
The GM-CBMeMBer filter has a close-form solution under assumptions of linear Gaussian models that is difficult to implement on the nonlinear measurement models. To overcome this limitation, we extended the GM-CBMeMBer filter to a practical nonlinear measurement model by using unscented transform (UT) techniques [29]. Another straightforward extension of the single sensor CBMeMBer filters to the case of multiple illuminators can be achieved by iterating the filter update stage for each illuminator measurement set. An IC-UT-GM-CBMeMBer filter can be implemented to accommodate a multi-transmitter nonlinear Doppler model. However, this IC-CBMeMBer yields final solutions that depend on the order of the measurement set of illuminators; therefore, the development of efficient algorithms for the scenario case are left for future investigation. Hence, in this section we propose the IC-UT-GM-CBMeMBer filter for multi-target tracking in the PBR system.
IC-UK-GM-CBMeMBer Filter
We supposed that each target follows a linear Gaussian dynamical and observation mode [21], i.e., where f k+1|k (· x k ) is a transition function commonly known as Markov shift [30].N( · ; m, P) denotes a Gaussian density with mean m and covariance P, F k is the state transition matrix, Q k is the process noise covariance, g k+1 (z|x) is likelihood function, H k+1 is the observation matrix, and R k+1 is the observation noise covariance. A multi-Bernoulli RFS is characterized by a posterior distribution with parameters existence probability r (i) and probability density , which is comprised of Gaussian mixtures of the form denote the weights, means, and covariances of the j-th Gaussian component by the sample time k.
The Bernoulli filter propagates the posterior π k = r k , p k (x k ) during the whole time in "prediction" and "update" steps. This effectively means that r k and p k must be propagated.
Prediction: At time k + 1, spontaneous births are accounted for by appending a birth multi-Bernoulli RFS with components r to surviving targets. The total number of predicted hypothesized tracks is M k+1|k = M k + M Γ,k+1 . The predicted multi-target density is [21]: where: Update: In the following, based on nonlinear observations, we propose the unscented transform implementation of the IC-UK-GM-CBMeMBer filter in the update step.
At time k + 1, the updated multi-Bernoulli density π k+1 is formed by multi-Bernoulli RFS of the legacy tracks (r and measurement-corrected tracks (r U,k+1 (W), p U,k+ as follows: where: Using unscented transform extends the mean matrix and covariance matrix, respectively, We constructed a set of 2n U + 1 sigma points χ where n U is the dimension of µ k , and κ U is the scaling parameters,n U + κ U = 0.
With the i-th illuminator measurement data Z (i) k+1 , the filter is obtained by the aforementioned sequential processing of the measurement set of each illuminator with the CB-MeMBer filter corrector. The update operator Ψ k+1 is [21,31]: where a, b = χ a(x)b(x)dx denotes the inner product, and the sequential update processing is as shown where • denotes a composition.
State Extraction and Cardinality Biass
Extract multi-target states are the same as that of the GM-MB filter; for more details see [21]. The number of targets is estimated by: For completeness, the key steps of the proposed filter are summarized as a block diagram of the processing algorithm in Table 1.
construction of birth target Gaussian components using Equation (13) end update the legacy tracks for i = 1:M k+1|k r (i) each component constructs a set of sigma points and weights using Equation (22) to generate: end compute (r U,k+1 , p U,k+1 ) using Equations (18) and (19) end prune tracks end state extraction and cardinality bias using Equation (37) end
Experimental Configuration
We developed the PBR system in Huazhong University of Science and Technology by tracking a close-in civilian airplane whose working frequency band is 6-30 MHz. The system is configured to work in multi-transmitter and receiver-only mode. The experiment was carried out in December 2014, in which three AM radio broadcast stations were selected as the noncooperative transmitters, namely, Tx1, Tx2, and Tx3, respectively. The specific parameters can be obtained from the International Telecommunication Union (ITU) Radiocommunication Sector [32,33] listed in Table 2, including carrier frequency (f c ), transmitting power, distance with respect to the receiver, and so on. The ground distance between the AM radio broadcast station and the receiver antenna array is over 800 km. Thus, the transmitted signals are reflected by the ionosphere to reach targets over-the-horizon away. Figure 4 shows the geographical distribution of the illuminators and the receiver station. The noncooperative targets in the experiment are two civil aircrafts in the surveillance area with flight numbers CCAXXXX and CSNXXXX, respectively, namely, Target 1 and Target 2. The civil aircrafts parameters were broadcast by the automatic dependent surveillance-broadcast (ADS-B) system within a short interval of time. The data sets, including position, velocity, and so on, are the reference to verify the tracking method, which is recorded by a ground-based AirNav Radar Box. The two real trajectories of the civil aircrafts during the experiment are plotted in Figure 5a The noncooperative targets in the experiment are two civil aircrafts in the surveillance area with flight numbers CCAXXXX and CSNXXXX, respectively, namely, Target 1 and Target 2. The civil aircrafts parameters were broadcast by the automatic dependent surveillance-broadcast (ADS-B) system within a short interval of time. The data sets, including position, velocity, and so on, are the reference to verify the tracking method, which is recorded by a ground-based AirNav Radar Box. The two real trajectories of the civil aircrafts during the experiment are plotted in Figure 5a The noncooperative targets in the experiment are two civil aircrafts in the surveillance area with flight numbers CCAXXXX and CSNXXXX, respectively, namely, Target 1 and Target 2. The civil aircrafts parameters were broadcast by the automatic dependent surveillance-broadcast (ADS-B) system within a short interval of time. The data sets, including position, velocity, and so on, are the reference to verify the tracking method, which is recorded by a ground-based AirNav Radar Box. The two real trajectories of the civil aircrafts during the experiment are plotted in Figure 5a
Field Experimental Results
The targets are observed in the surveillance region with dimensions [−40, 40]km × [−40, 40]km. The single-target transition model is a linear Gaussian process given by Equation (11), in which ∆ = 1 s is the sampling period, and σ v = 0.1m/s 2 is the standard deviation of the process noise.
The birth process is multiBernoulli with density . The probability of target survival is p S,k = 0.95. The probability of target detection is p D,k = 0.5.
After the aforementioned signal processing, we obtained the DOA estimation of each direct wave and the Doppler shift measurement data, including the false alarms and misdetections. Figure 6 shows the detected Doppler vs. time obtained from the surveillance areas using three AM broadcast stations with the carrier frequency of 17.7 MHz, 15.37 MHz, and 15.5 MHz, respectively. The Doppler measurement sets have clutter and the missing alarm. Then, the noisy three stations Doppler-shift measurement sets Z (I) 1:80 , are passed to the tracking filter, as plotted in Figure 7. The parameters of the tracking filter are set as follows: observation noise covariance R k = σ 2 ε I 1 , where σ ε = 1Hz is the standard deviation of the measurement noise. Clutter parameter is Poisson with intensity κ k (z) = λ c Vu(z), where u(z) is a uniform probability density over the surveillance region, V = 1600 km 2 is the "volume" of the surveillance region, and the clutter intensity is λ
Field Experimental Results
The targets are observed in the surveillance region with dimensions The initial density of the target state 0 ( ) p x is the Gaussian mixture of the form: Figures 8 and 9, which show the estimated target traces and four estimated components of the state vector: px, py, vx, vy change vs time compared with the true trajectories, respectively. At ending time instants, short discontinuities occur in the tracks owing to the missing alarm of the Doppler measurement. Notice that the number of targets suffers from latency problem at the beginning of tracking in Figure 10, because the initial points are located arbitrarily. The initial density of the target state 0 ( ) p x is the Gaussian mixture of the form: Figures 8 and 9, which show the estimated target traces and four estimated components of the state vector: px, py, vx, vy change vs time compared with the true trajectories, respectively. At ending time instants, short discontinuities occur in the tracks owing to the missing alarm of the Doppler measurement. Notice that the number of targets suffers from latency problem at the beginning of tracking in Figure 10, because the initial points are located arbitrarily. Different numbers of illuminators are a problem in multi-target tracking performance when only Doppler measurements are used. To study this, the filter is implemented under the conditions of AM radio broadcast stations Ns = 2,1, corresponding to stations with serial numbers of I = {1, 2}, {1}, respectively. The optimal subpattern assignment (OSPA) is used to evaluate the tracking miss-distance. The OSPA distances (for c = 20 and p = 1) vs. time on conditions of various number of broadcast stations compared with the results on the condition of Ns = 3, I = {1, 2, 3} is shown in Figure 11. Particularly, the OSPA distances vs. the time between 6s and 32s is plotted. It can be seen that the estimated largest OSPA distances are approximately 2880 m, 2922 m, and 5712 m on the condition of Ns = 3, Ns = 2, and Ns = 1, respectively. The obvious error in the period time from k = 27 to 56 is due to the missing detections and clutter of the Doppler measurements and the number of illuminators. Therefore, we believe that the more numerous the AM radio broadcast stations that are exploited, the more accurate the tracking trajectories are. Different numbers of illuminators are a problem in multi-target tracking performance when only Doppler measurements are used. To study this, the filter is implemented under the conditions of AM radio broadcast stations Ns = 2,1, corresponding to stations with serial numbers of I = {1, 2}, {1}, respectively. The optimal subpattern assignment (OSPA) is used to evaluate the tracking miss-distance. The OSPA distances (for c = 20 and p = 1) vs. time on conditions of various number of broadcast stations compared with the results on the condition of Ns = 3, I = {1, 2, 3} is shown in Figure 11. Particularly, the OSPA distances vs. the time between 6s and 32s is plotted. It can be seen that the estimated largest OSPA distances are approximately 2880 m, 2922 m, and 5712 m on the condition of Ns = 3, Ns = 2, and Ns = 1, respectively. The obvious error in the period time from k = 27 to 56 is due to the missing detections and clutter of the Doppler measurements and the number of illuminators. Therefore, we believe that the more numerous the AM radio broadcast stations that are exploited, the more accurate the tracking trajectories are.
Simulation Results
In this subsection, the performance of the proposed method is verified via simulation under similar scenarios to those aforementioned under the situation of cross trajectories, in consideration that it usually occurs in real data processing on a 2D Cartesian coordinate. Three AM broadcast stations were chosen, the same as Table 2. The two targets' motion is assumed to be a nearly constant model adjusted for civil aircrafts, and the flight parameters are listed in Table 3. The false alarms are uniformly distributed in the field of view with range −30 Hz to 30 Hz, and the number of false alarms at each scan follows the Poisson distribution with a mean of 10. The parameters of the tracking filter are set the same as in Section 4.2. As shown in Figure 12, despite the intersection points, the two targets can follow their trajectories, respectively. In Figure 13, the OSPA metric (p = 1, c = 20) shows the track
Simulation Results
In this subsection, the performance of the proposed method is verified via simulation under similar scenarios to those aforementioned under the situation of cross trajectories, in consideration that it usually occurs in real data processing on a 2D Cartesian coordinate. Three AM broadcast stations were chosen, the same as Table 2. The two targets' motion is assumed to be a nearly constant model adjusted for civil aircrafts, and the flight parameters are listed in Table 3. The false alarms are uniformly distributed in the field of view with range −30 Hz to 30 Hz, and the number of false alarms at each scan follows the Poisson distribution with a mean of 10. The parameters of the tracking filter are set the same as in Section 4.2. As shown in Figure 12, despite the intersection points, the two targets can follow their trajectories, respectively. In Figure 13, the OSPA metric (p = 1, c = 20) shows the maintenance quality of the proposed method. However, the instantaneous peaks are observed from times k = 59 to k = 81 due to corresponding intersection point and track termination latency. The simulation results indicate that the proposed method can deal with relatively complex tracking problems.
Conclusions
In this paper, we propose a multi-target tracking filter in a self-developed PBR system by using spatially distributed multiple AM broadcast stations. Multiple non-cooperative illuminators with different carrier frequencies located over-the horizon and one receiver in the surveillance area are involved in the practical system. The direct wave and the illumination wave are reflected from the ionosphere received by a uniform circular maintenance quality of the proposed method. However, the instantaneous peaks are observed from times k = 59 to k = 81 due to corresponding intersection point and track termination latency. The simulation results indicate that the proposed method can deal with relatively complex tracking problems.
Conclusions
In this paper, we propose a multi-target tracking filter in a self-developed PBR system by using spatially distributed multiple AM broadcast stations. Multiple non-cooperative illuminators with different carrier frequencies located over-the horizon and one receiver in the surveillance area are involved in the practical system. The direct wave and the illumination wave are reflected from the ionosphere received by a uniform circular
Conclusions
In this paper, we propose a multi-target tracking filter in a self-developed PBR system by using spatially distributed multiple AM broadcast stations. Multiple non-cooperative illuminators with different carrier frequencies located over-the horizon and one receiver in the surveillance area are involved in the practical system. The direct wave and the illumination wave are reflected from the ionosphere received by a uniform circular array located over the horizon, while the echo wave (target-to-receiver) is via LOS propagation in the surveillance area. After some techniques in array signal processing, the Doppler measurement sets, including clutter and the missing alarm with corresponding time, can be collected. To overcome linear Gaussian models, we propose the tracking model and extend the GM-CBMeMBer filter to a practical nonlinear measurement model by using unscented transform (UT) techniques by iterating the filter update stage for each illuminator measurement set in this practical scenario. Three AM broadcast stations were selected as the non-cooperative illuminators. Two non-cooperative civil aircrafts were chosen as tracking targets, whose flight parameters were recorded by a ground-based AirNav Radar Box set. Considering the clutter and missing alarm in the measurement sets, the OSPA distances are acceptable. Moreover, the performance of simulation has verified the feasibility of the proposed tracking method. In future work, the unknown clutter rate and detection probability under unknown background in this practical scenario will be taken into consideration. Maneuvering target tracking is also worthy of study. | 7,226.8 | 2021-09-01T00:00:00.000 | [
"Engineering"
] |
Dot based Printout Watermarking Suitable for Practical Environment
Information leakage by printouts is common in ordinary office environment. This paper proposes a practical watermark scheme applicable with actual imaging devices to trace printouts comprising watermark patterns. We generate a watermark pattern composed of tiny yellow dots whose coordinates are irregularly located onto the intersections of virtual lattice. We exploit the relationship among the dots’ coordinates by Hough transform and discrete Fourier transform to detect the embedded watermark in the printouts. Empirical evidence from a large database of watermark images indicates the superior performance of the proposed method.
Introduction
With the development of easy-to-use printing devices, printouts are commonly used in our office environment.At the same time, the very nature of printouts, which can be easily leaked, brings into questions many of the positive aspects associated with the printouts.In fact, most of confidential information, such as CAD drawings or organization's internal documents, is disclosed with printouts, and a constantly growing number of uncovered leakages is certainly only the tip of the iceberg.
To prevent or at least trace the leakage of printout, mostly two approaches, i.e., source device forensics (1) and digital watermarking (2) , were suggested in the literature.Source device forensics, which passively investigates a given material without any prior-knowledge, has been researched for the last decade (3) .The techniques are based on the assumption that the resulting material includes inherent artifacts to designate a specific source device.Such artifacts due to the printing process, e.g., banding frequency (4) , character features (5) , or halftoning effect (6) , were revealed by numerous previous research.Unfortunately, the forensic methods mainly focused on identifying printing sources so that the techniques were not adequate to trace specific information of leakage behaviors.
Aside from the passive approaches, digital watermarking for printout, which actively inserts auxiliary information into a given document, has also been researched intensively.The earliest watermarking method for printout explicitly printed visible logos onto the documents (7) .However, scholars rarely consider it as digital watermarking since modern watermarking schemes require imperceptibility.One of the first imperceptible watermarking schemes robust to printing and scanning was proposed by Low et al. (8) .They slightly shifted text lines up or down, or words left or right from their original positions.Similarly, Borges et al. suggested adjusting the brightness of characters according to the insertion information (9) .However, these schemes were not able to detect the embedded watermark when the printout with the watermark was geometrically distorted.Recently, printer steganography, which hides unrecognizable tiny yellow dots by the naked eyes to designate printing information, was applied to numerous color laser printers (10) .It gave rise to the violation of privacy rights because various color laser printers, made by HP, Canon, or Xerox, occasionally inserted those dots without owner's permission.Inspired by the printer steganography, Briffa et al. proposed to embed a plurality of imperceptible yellow dots for documents during the printing process, yet it was weak against geometric distortions (11) .Finally, several natural language process (NLP) based watermarking schemes were proposed for printout (12,13) .They replaced chosen words with respect to a given codebook unless it did not change the meaning of the sentences.However, the nuances of the text were slightly changed although the transitions were rarely noticeable.Moreover, all of the methods mentioned above only operated with flatbed scanners with which distortions were merely occurred.
To the best of our knowledge, existing watermarking schemes for printout have technical limitations to use practically.Therefore, in this paper, we propose a novel printout watermarking scheme that overcomes the above drawbacks so that it is eventually applicable with smartphone.Specifically, we introduce watermark pattern barely perceptible by the human visual system (HVS).In addition, various kinds of attacks such as digital-analog/analog-digital (D-A/A-D) conversion, scaling, rotation, perspective projection, and cropping are considered which unintentionally occur during the detecting process with webcam or smartphone.Furthermore, the robustness to partial distortions, i.e., some of the watermark in the printout is damaged by text, is considered as well.
The rest of the paper is structured as follows.We first introduce a proposed dot based watermark generation scheme considering both imperceptibility and robustness to distortions in Section 2. Based on the watermark generated in this section, the proposed watermark detection scheme suitable for practical environment is described in Section 3. Subsequently, Section 4 reports the experimental results from a massive test setup before Section 5 concludes the paper.
Watermark Generation for Printout
As mentioned above, the practical printout watermark should satisfy both imperceptibility by the HVS model and robustness to various distortions.To satisfy the requirements, we propose a dot based watermarking scheme which highly improves the printer steganography (10) and Briffa's method (11) .To achieve the imperceptibility, the proposed method outputs irregularly spaced yellow dots of different sizes, which are tiny enough with respect to the HVS model, according to the type of information to be inserted.Since yellow is least noticeable among four colors of CMYK toners, especially accompanied by white colored blank area occupying most of the printout, using the yellow dots are adequate for printout watermark.Fig. 1 shows proposed dot symbols and the example of generated watermark which denote bit zero, bit one, and the combination of two symbols, respectively.The following subsection describes the details of the proposed generation method.
Details of Watermark Generation
As depicted in Fig. 1, we convert each bit of embedding data to a different sized yellow dot within the predefined size of white background area.We normally defines the background size as 7×7 pixels because its actual size is about 0.06×0.06cm 2 under 300 DPI printing environment, which is default setting for most of color laser printers.It is tiny enough to satisfy the requirement of imperceptibility.To minimize the perceptibility, we additionally adjust the distances among the dots which are randomly varied by the average distance threshold d.Eventually, the watermark pattern is comprised of the dot symbols sequentially located with random distances within the appropriate size of square region.Since our ultimate goal is designing a watermark detectable with smartphone, the size of watermark pattern should be restricted in terms of mobile camera's facility.In this perspective, the size of the watermark pattern is empirically chosen around 1 inch 2 , maximally capable with about 40×40 symbols.Furthermore, watermark suitable for practical environment should robust to various distortions caused by illumination, toner powder, or characters.For this purpose, we attach Reed-Solomon error correction code to the embedding data (14) .The amount of the correction code is varied with circumstances.
Watermark Detection for Printout
This section presents a watermark detection method from watermarked printout.Fig. 2 depicts an acquired image frame from a watermarked printout by Logitech C920 webcam.As shown in the Fig. 2, we can notice that the digitalized image frame undergoes several attacks such as D/A-A/D conversion, geometrical transformations, gradual lighting changes, cropping, and partial distortions by characters.Therefore, through the following subsections, we explain each step of the proposed detection scheme depicted in Fig. 3. Specifically, we detail the proposed scheme to overcome the attacks mentioned above, so that we prove the practicality of the proposed scheme.
Extraction of Dot Components
The proposed method first extracts a set of dots from a captured image frame.As described in Section 2, the proposed watermark consists of a purity of yellow dots irregularly printed onto a paper.Since the dots are marked in yellow, we can easily separate the dots by extracting the yellow objects from the given image.Hence, the image is first converted from RGB to hue-saturation-value (HSV) color space because color information is mainly included in the Hue channel (15) .Especially, the yellow component is generally depicted around 60° of Hue angle.Therefore, threshold values for Hue, Saturation, and Value are set as follows.
The other benefit of thresholding in the HSV space is shadow removal.Since the shadow, due to the illumination change in practical environment, barely has color information, it is effectively separated in the HSV space.After extracting the yellow objects, mostly assumed as watermark dots, the center coordinate (̅ , ̅) of each object is computed by dividing the x-and y-related first moments of each object by the zeroth moment values (15) .This is shown in the following equation.(1)
Correction of Perspective Projection by Hough Transform
Once the candidate coordinates for the dots are decided, the given image is restored to the original shape.The important assumption supporting to detect the watermark in the practical environment is that the dots are irregularly placed onto the intersections of virtual lattice.By revealing the relationship among the dot's coordinates, the reconstruction of the image is performed in two stages.In this section, we explain how to inversely transform perspective projection, which always happens during the image acquisition process.
We first try to find straight lines that pass through as many dots as possible.Since the coordinates for the dots are irregularly located on the intersection of the lattice, many of the straight lines, regardless of geometrical distortions, belong to either horizontal lines or vertical lines with high probability.In order to find the candidate straight lines passing through the dots, Hough transform is applied to the given coordinates (15) .The Hough transform is as follows.
for θ = 0 to π ρ = x cosθ + y sinθ accumulator[ρ, θ] = accumulator[ρ, θ] + 1 end select peaks from accumulator After the Hough transform, selected straight lines are represented by ρ and θ, which refer to the distance from the origin and the angle value of the line, respectively.Fig. 4(a) and Fig. 4(b) depict the resulting lines in terms of the Hough and the spatial space, respectively.It is noticeable that the result reveals clusters for both horizontal and vertical lines.
Perspective projected image can be restored to the corresponding original image if four matched pairs of points exist.Therefore, with the extracted lines, especially belonging to either the horizontal or vertical line group, we attempt to estimate four corner points.Specifically, we select two appropriate angles, each of them indicating either the horizontal or vertical line group, followed by two lines in each group are extracted having the minimum and the maximum distance from the origin, respectively.Consequently, four intersections by the selected two horizontal and two vertical lines are assumed the matching points for the perspective projected image, which are then used to transform the perspective projected image into the square of predefined size.Fig. 4(c) and Fig. 4(d) visualize the four intersections with the selected lines and the corrected image, respectively.
Resizing to Original Image
As described in the previous section, the distorted image is mapped into the predefined size of square.Since the intervals between the selected horizontal or vertical lines are not predictable, the aspect ratio of the restored image is not equal to one with high probability.Therefore, we inspect the periodicity between the restored dots through both x-axis and y-axis to resize the image into proper aspect ratio.Recalling watermark generation section, each dot pattern is inserted with a size of 7×7 pixels.Therefore, to correct the size of the image, the periodicity of the dots in the x-and y-axis directions is obtained followed by the interval is corrected to 7 pixels.Specifically, the periodicity of the x-axis direction is revealed by accumulating local maxima among dots through the y-axis direction, which is given by visualizes the periodicity of accumulated values in spatial domain.Afterwards, Discrete Fourier Transform (DFT) of x ̂ is applied to explicitly uncover the periodicity (16) .
𝑋 ̂= 𝐷𝐹𝑇(𝑥 ̂)
(3) Since the magnitude of ̂ is represented by the impulse train, as shown in Fig. 4(b), the reciprocal of the difference between DC and the first peak is determined as the current periodicity of the x-axis direction.The y-axis periodicity is computed similarly.Finally, the original image is estimated by resizing the given image with respect to the ratio of the computed periodicity to the predefined periodicity.
Decision of Embedded Information
This section presents a data extraction method from the watermark pattern of the reconstructed image.First of all, we determine dots' coordinates by accumulating the image followed by selecting local maxima as described in the previous section.Subsequently, the size of dot's area is computed.Dots, whose areas are less than threshold t, are considered as empty space because of noisy circumstance (e.g., illumination, shadow, toner powder, and etc.).Thereafter, we decide each dot, whose area is larger than threshold t, as one of the two symbols shown in Fig. 2 by 2-means clustering (15) .Threshold t is set as 0.4 empirically.Finally, Reed-Solomon code corrects errors during the detection process.
Experimental Results
This section reports results from an extensive series of watermark detection experiments.The setup includes many variations of watermark patterns generated by the proposed method and the corresponding detection performance.Specifically, we evaluated the detection performance along two main directions as follows.
The baseline experiments digitally simulate the performance of the proposed method under various conditions.We measured the detectability of the watermark varying the parameters mentioned in Section 2 and Section 3. Partially distorted watermark patterns with the specific parameter setting were analyzed as well.Besides, geometrical distortions, which inevitably happen during the practical acquisition process, were also simulated before the following experiments. The practical experiments investigate the performance of the proposed scheme under actual D-A/A-D conversion, i.e., printing process followed by image acquisition process with scanner and mobile.For this series of experiments, the parameter setting for watermark generation was kept constant, whereas detecting devices and the corresponding circumstances were altered diversely.If not stated otherwise, we used the following settings.A Reed-Solomon code whose length is same with the corresponding watermark was attached to the watermark so that we guaranteed 25% of error correction ability.A set of experiment, which shares same parameters, was constructed with 100 watermarked images.
Baseline Experiments (a) Various Parameter Settings without Distortions
First of all, we investigated the relationship among imperceptibility, capacity, and detectability.For this purpose, we generated numerous watermark patterns with different settings by keeping the distance parameter d, while systematically varying the amount of embedding data and the size of embedding region, resulting in altogether 16,500 runs.The imperceptibility is determined by the distance parameter d of the proposed method.The closer the gap between the dots, which means smaller d value, the stronger visual artifact is caused.On the contrary, the embedding capacity is influenced by the distance and the size of the embedding region.The size of the region should be small enough by considering image acquisition process with actual devices such as mobile and webcam.Therefore, we investigated the trade-off among imperceptibility, capacity, and detectability varying the sparseness, the amount of embedding data, and the size of the regions.Fig. 6 depicts the results of the proposed method.The results report up to 19 bytes of embedding capacity with the distance parameter d=3 and the size of 1.0 inch 2 , which satisfy requirements for capacity, imperceptibility, and detectability.Even though the imperceptibility is hard to evaluate, we empirically conclude that the distance of higher than three is sufficient for the criterion by four expert observers participated in the experiments.
(b) Robustness to Partial Distortions Subsequently, the watermark should be robust to the distortions by characters.To verify the robustness, we randomly manipulated embedded bytes in the watermark patterns fixing the size of the region as 1.0 inch 2 whereas varying the distance and the number of erroneous bytes.This series of experiments comprised in total 4800 runs.Fig. 7 shows the detection rate against partial manipulation, 3 bytes manipulation with d=4, and 15 bytes manipulation with d=3, respectively.The results revealed the robustness of proposed method against partial manipulations.
(c) Robustness to Geometrical Distortions
Eventually, we manipulated watermarked patterns geometrically, which always occurs during the image acquisition process.The patterns were systematically modified by representative Affine transformation, e.g., scaling and rotation.Separating from the Affine transformation, perspective projection was performed on the given watermark patterns.Since the projection matrix is hard to parameterize, we randomly modified both x-axis and y-axis aspect ratio in the range of 0.9 to 1.1 to generate modified patterns.Table 1 x-axis aspect ratio y-axis aspect ratio 0 0.9 ~ 1.1 0 100 100 0.9 ~ 1.1 100 100 detection rate under Affine transformation and perspective projection, respectively.Surprisingly, the proposed method demonstrated detection rate of 100% with various simulated geometrical distortions
Practical experiments
Experimental results in the following sections depict the performance of the proposed method with actual printing and image acquisition devices.Table 3 lists the devices we used.Randomly generated watermark patterns with d=4, the size of 0.8 inch 2 , and embedding capacity of 7 bytes were firstly printed onto the papers by Samsung Multi Xpress C9250ND with 300 DPI.Thereafter, the watermarks on the papers were digitalized by Epson Perfection V37 and Samsung Galaxy S7, respectively.With the obtained image frames, we analyzed the performance of the proposed method.
(a) Detection with a Scanner
The watermarked image frames were obtained by scanning the papers in the range of 0° to 30°, applied in steps of 10°.Additionally, DPI of the scanner was also varied from 300 DPI to 900 DPI, in steps of 150 DPI.Table 4 completes the detection result of the proposed method under this scenario.The result showed similar result with the previous section because the scanned images were not affected by harsh attacks, e.g., perspective projection or illumination.
(b) Detection with a Smartphone
Our last series of experiments analyzed image frames acquired from an actual smartphone's continuous video.Unlike the scanner, smartphone's camera offers limited resolution.Because higher resolution provides better image details, we targeted HD and FHD videos, the resolution of 1280×780 and 1920×1080, respectively, which are normally the highest resolution supported by ordinary smartphones.Furthermore, we modified the distance between the watermarked papers and the smartphone from 7 cm to 16 cm in steps of 3 cm because the close distance ensures the details of image frames but might cause lens distortions.Table 5 summarizes the results in this perspective.Compared to the case from a scanner, the results tended to give lower performance.The performance was strictly deteriorated due to lack of details when the distance was too far.The performance was also degraded with the distance of 7 cm.We suspect both lens distortions and the size exceeding of the watermarked region.In spite of the decreased performance, the results from midrange were acceptable applying to actual smartphone because the video frames were able to be processed continuously.
Conclusions
With the rapid progress of printing technology, we are threatened by a constantly increasing number of information leakages in forms of printouts.To trace the leakage of the printouts, in this paper we have focused on developing a practical watermark scheme applicable with actual imaging devices such as printers, scanners, and smartphones.Given that prior studies for information hiding, which print tiny dots onto papers, we further invented a novel watermarking scheme even detectable with smartphones.By exploiting the relationship among the dots' coordinates, which are irregularly placed onto the intersections of virtual lattice, we developed a watermark
Fig. 1 .
Fig. 1.Dot patterns and a watermark pattern: (a) a dot pattern denoting bit zero, (b) a dot pattern denoting bit one, and (c) a watermark pattern generated by the combination of (a) and (b).Fig. 2.An acquired image frame in practical environment (by Logitech C920).
Fig. 4 .
Fig. 4. Snapshots during the correction of perspective projection: (a) extracted lines in Hough domain, (b) the corresponding lines seen in spatial domain, (c) four selected lines for the correction, and (d) the corrected image.
Fig. 5 .
Fig. 5.A x-axis periodicity revealed by resizing procedure: (a) a periodicity of accumulated local maxima through y-axis, (b) and the corresponding impulse train in frequency domain.
Fig. 6 .
Fig.6.The shapes of watermark patterns and the corresponding detection results; (a, b, c) visualize the shapes of watermark patterns with the fixed distance threshold d=1, d=3, and d=7, respectively; (d, e, f) depict the corresponding detection rates varying the size of watermark pattern in the range of 0.8 2 to 1.6 2 inch 2 in steps of 0.2 2 inch 2 and the amount of embedded data from 3 bytes to 31 bytes in steps of 3 bytes.
Fig. 7 .
Fig. 7. Watermark detection with partial distortion; (a) depicts the detection rates with respect to the distance threshold d and the amount of byte errors; (b, c) visualize distorted watermark patterns by 3 byte errors with d=4 and 15 byte errors with d=3, respectively.
and Table 2 summarize the
Table 1 .
Watermark detection results with simulated rotations and scalings.
Table 3 .
Devices used for the experiments.
Table 4 .
Watermark Detection results from scanned watermark printouts.
Table 5 .
Watermark Detection results with image frames acquired by smartphone.Experimental results confirmed the superior performance of the approach under a variety of settings.Apart from endeavors to extract printout watermark generated by color laser printers, watermark patterns printed by black and white printers, which are inherently affected by half-toning effect, should be considered in the literature.Several expecting issues caused by adopting black and white printers, e.g., reducing perceptibility, difficulty in separating dots, and increasing complexity, should be solved as well. | 4,842.6 | 2017-01-01T00:00:00.000 | [
"Computer Science"
] |
Estimating the Transmission Risk of COVID-19 in Nigeria: A Mathematical Modelling Approach
Objectives: The potential burden of COVID-19 in sub-Saharan African might be substantially more significant than reported, and more than the existing health system can handle. Hence, in this study, we estimate and project the burden and transmission risk of COVID-19, in Nigeria, using current interventions.
Methods: Modified SEIR epidemic mathematical model was used to simulate the disease progression in weeks, for up to 19 weeks. Different situations, involving zero-intervention and varying degrees of interventions are modeled. For the intervention phase, 25% and 75% social distancing are considered, while border closure includes 80% closure of airports, seaports, and intra-state borders, using available data as of 15th May 2020.
Results: The effects of various interventions on the R0 of COVID-19 are presented. A higher percentage of social distancing appears to be more effective in controlling the spread of COVID-19 in Nigeria than border closure. Up to 131,000 persons could be infected if there are no interventions.
Conclusion: According to our results, it is easier to enforce 75% closures than 25%, as the percentage of the population complying with social distancing is higher when at least 75% of public places were closed. The minimum requirement of the population percentage that needs to comply with the social distancing advice, to weaken the epidemic can be obtained from the model.
Introduction
In late 2019, coronavirus disease 2019 (COVID- 19) emerged in Wuhan, Hubei province of China, causing a pandemic that has continued to wreak havoc, through unprecedented global spreading. As of April 3, 2020, at least 1,088,878 cases have been confirmed in over 180 countries, 200 territories, and five international ships, with a case fatality rate of 5.4% [1,2]. In Africa, all but four countries have reported cases. The first case of the continent was confirmed on February 14, 2020, in Egypt. The first confirmed case in sub-Saharan Africa was in Nigeria. Most of these cases were individuals who just arrived from Europe and the United States. There are concerns about the spreading of COVID-19 in Africa. The reason is most of the healthcare systems are inadequate, having problems such as lack of equipment, lack of funding, insufficient training of healthcare workers, and inefficient data transmission. The pandemic could also cause substantial economic issues across the countries. As of April 3, Nigeria has recorded 210 cases, with a case fatality of 1.9% [3]. The potential burden of COVID-19 in Nigeria might be substantially more significant than reported, and more than the health system can handle. In this study, we estimate the burden and transmission risk of COVID-19, using hybrid mathematical modeling.
There are still developing virologic details about the COVID-19. However, it is known to belong to the class of infections associated with humans but linked to animal origins, and this is typical of all Severe Acute Respiratory Syndrome coronaviruses (SARS-Cov). According to the International Committee on Taxonomy of Viruses [4], Coronaviruses belong to the family of Nidovirales and the sub-family of Coronavirinae, with four identified strains (alpha coronavirus, beta coronavirus, gamma coronavirus, and delta coronavirus), of which alpha and beta Coronavirus strains mostly affect the respiratory tracts of humans, with human to animal transmissions typical with the other two strains [5]. Coronavirus disease spreads primarily through contact with an infected person when they cough or sneeze. It also spreads when a person touches a surface or objects that have the virus on it and then touches their eyes, nose, or mouth. Other characteristics of the disease presentation include communal transmission. The basic reproduction number R 0 , which is generally considered low for SARS-CoV infections, however, posts higher values in COVID-19. The study in [6] identifies varying modes of transmission, including community spread, international and local travels, and some interventions (government and individual action; quarantine; restricted movements and social distancing. Others include the influence of geographic location, humidity, and temperature on the scale of the epidemic. There are questions about coronaviruses affected by these environmental factors. The study in [7] suggests high temperature and humidity may have influenced the transmission of SARS-CoV in China. The report suggests a decline in the rate of transmission of infection in relatively humid conditions and high temperatures. Similar findings provide support to previous studies on the influence of temperature on the infection rates [8,9]. Among others, persistent dry cough, fever, nausea, and vomiting, fatigue, and pneumonia were reported to be common among patients, although at varying degrees and length of contracting the virus [10]. Some of the interventions reported as addressing or mitigating the effects of disease transmission include the establishment of a contingency system; examples include monitoring, contact tracing and identification of international travelers, early diagnosis, early isolation of infected patients, government actions (including national /city lockdown, an extension of the holiday period, hospitalization and quarantine) [11,12]. Peng used the classical susceptible, exposed, infectious, and recovered model (SEIR) to estimate the duration of the COVID-19 epidemic in multiple cities in China. Similar models have been used to study the dynamics of other viral outbreaks [13,14]. The length of infection, reproduction number, and quarantine time were the parameters used to forecast the inflection point, recession time, and the magnitude of the epidemic (depending on the commencements of intervention). The duration of the outbreak in China was estimated to last between one month to four months, affected by temperature, mitigation intensity, individual, and government actions [15].
Materials and Methods
We developed a hybrid stochastic mathematical model, with some deterministic inputs, involving a modified version of S-E-I-R epidemiological modeling structure. Modeling real-life situations allow researchers to comprehensively study an otherwise complex set of events, environment or populations and their interaction with causative agent factors. Mathematical models simulate the spread of infectious diseases within a host population. They can be used to investigate mechanisms underlying disease spread, or to predict the future trajectory of an epidemic and impacts of selected control measures. Several disease models have differing assumptions about mixing patterns between hosts. Mathematical and agent-based models have proven to be very useful in studying disease dynamics and assessing treatment strategies for specific groups and general populations [16][17][18][19]. Similar structures have been used recently in investigating the dynamics of the COVID-19 [20]. The population is divided into six groups: Susceptible (S), Exposed (I), Infected (I), Hospitalized (H), Recovered (R), and Removed (D). The Susceptible group represents all members of the public that are prone to contacting the disease, due to interactions with other population sets. The Exposed group consists of individuals who have been exposed to the virus, and their infections are in a latent stage. These individuals progress to the infectious group, with zero or more symptoms, by a relative rate κ. Infected persons can move to the hospitalized group (H), by a rate α. The number of deaths is recorded in the removed (D) group. The relationships among the groups are presented in the model structure in Fig-1. The population in the model N, at time is equal to the sum of all groups. Model parameters and descriptions are presented in Table-1.
Nigeria is used as a case study in the model. The current population of Nigeria is estimated to be 204 million [21]. The prevailing situation of COVID-19 in Nigeria as of May 15 th , 2020, as released by the National Center for Disease Control [3], is used as the initial values for all model groups. Our model considered events during pre-and post-intervention scenarios, similar to the general population modeling scenarios used in [22] for the study of Ebola virus transmission. The initial scenario explores the possible behavior of viral transmission instead of public health interventions.
Parameter Estimation:
Due to the paucity of data and modeling rates, a sampling algorithm, based on Bayesian Inference, for estimating and improving model simulation accuracy is used. Reasonable parameter estimation with uncertainty analysis is important in predictions [23]. The sampling method assumes these parameters are independent and repeatedly use the previous number of daily confirmed cases to iterate in the multidimensional space composed of all parameters and obtain optimal estimation by constructing the likelihood function. Assuming ( ) represents the parameters prior distribution, then ( | ) becomes the likelihood function, representing the probability of observation sets when the parameter is . The steps are highlighted below: 1.
Select as an initial parameter value, with ( ) is chosen in space θ. 2. Sample parameters at ( ) based on the prior uniform distribution, or normal
distribution.
3. Obtain the acceptance rate, based on the ( | ( )) ( | ( )) and prior distribution. Using the acceptance level, determine if there is a change, otherwise accept the previous value. 4. Repeat step 2 until the required number of iterations is reached.
The normal distribution probability function is: The second major scenario models outbreak transmissions and intervention impacts in the postintervention phase. Two different interventions, socialdistancing, and border-closure are further modeled. Different rates of compliance with social distancing public health advice and lockdown are used to study the transmission. The model equations are presented below: The force of infection λ is calculated using the following equation: The susceptible group is exposed to the novel coronavirus through the infected and hospitalized group. As the susceptible group is exposed to the virus through two groups, we have considered two different types of transmission rates: a rate of transmission among the community members and another rate for spreading the disease between the community members and the hospitalized patients. The force of infection is calculated by combining these two different transmission rates. One is the transmission rate between community members, and the other is the transmission rate between the community members and the hospitalized people. Similar transmission rates for the general population were used in the modeling study. The incubation rate (κ) is the reverse of the incubation period.
The rate of symptom onset to the hospitalization of community members (α) is calculated by reversing the period of symptom onset to hospitalization. The basic reproduction number R 0 , which is the number of secondary cases generated by infected people, for Nigeria, is considered to be 2.42 initially. [24], while the natural mortality rate is 0.014. For studying the behavior of the disease in the post-intervention phase, three different scenarios were used. 25% and 75% of social distancing interventions were applied in first and second scenarios, respectively. In this study, 25% of social distancing implies closures of worship centers and avoidance of gatherings of 50 persons and above. 75% includes closure of public transportation and all markets, in addition to the initial conditions of 25%. In the third scenario, 75% of social distancing and 80% border closure is used to study the behavior of the COVID-19 disease outbreak model in the postintervention phase. We observed the effect of the intervention in each of these scenarios. The model parameters, description, and their values are as shown in Table-1. The susceptible and recovered equations after social distancing are applied as the intervention is given below: The equation for calculating the force of infection becomes: The effect on R 0 after social distancing is applied as an intervention is given below: (12) The susceptible and recovered equations after social distancing and border closure are applied are given below: The effect on R 0 after border closure is applied as an intervention is observed using the following equation: (15)
Results
Four different scenarios were created to aid the understanding of the outbreak dynamics in Nigeria. The first scenario corresponds to a zero-intervention situation. Data from NCDC informs the relatively high force of infection used. As seen in Fig-2, the number of infected people reaches a high value within the first two weeks. In scenarios 2, 3, and 4, interventions of varying degrees were applied to the model. These are as shown in Fig-4 and Fig-5, respectively. Fig-3 compliance with social distancing increases, R 0 gradually drops. A similar effect and relationship can be seen in Fig-4, where a tighter restriction (75%) on public gathering and social distancing along with 80% border closure is applied. Outputs of scenario 3 are as shown in Fig-4. Here, the value of R 0 drops rapidly and converges to zero earlier, than in scenario 2. This is due to a higher rate of intervention used to model the outbreak in this scenario. A faster convergence of R 0 towards zero is observed in scenario 4, as shown in Fig-4 compared to scenarios 2 and 3 ( Fig-3 and Fig-4). This can be attributed to the introduction of these combined interventions.
Discussions
To further understand the effects of R 0 on all levels of social distancing and border closure interventions, the relationship between the percentage of the population observing social distancing and the time in weeks when the value of R 0 falls below 1 is presented in Fig-5. It can be seen that the percentage of the population complying with social distancing is higher when a 75% social distancing rate is used. This implies that it is easier to enforce 75% social distancing in the population. That is, individuals tend to adhere to instructions when more centers and public transportation are closed. When the rate of social distancing applied is increased, the value of R 0 falls below 1 within fewer days. In scenario 4, where 75% of social distancing is used along with 80% border closure, the time in weeks when the value of R 0 falls below 1 is the same as scenario 3, where 75% of social distancing is applied. The minimum percentage of the population that needs to comply with social distancing, to weaken the outbreak model is obtained from the model. Scenario 4 shows the effects of the application of combined intervention, to understand the dynamics of the outbreak model. Changes in the value of R 0 for scenarios 2, 3, and 4 are shown over time in Fig-6. In all three scenarios, the value of R 0 decreases over time. It is observed that the value of R 0 for scenario 2 is larger compared to scenarios 3 and 4 at each time step time. In scenario 4, where combined intervention is applied, the value of R 0 converges to zero earlier compared to scenario 2 and scenario 3, where single intervention is applied. It is observed that the combined intervention weakens the outbreak model earlier compared to a single intervention at a different rate.
In this paper, we present the transmission pattern, risk, and estimated burden of COVID-19 in Nigeria, using mathematical modeling. Situations of zero interventions and different degrees of interventions are modeled. For the intervention phase, 25% and 75% social distancing are considered, while border closure involves 80% closure of airports, seaports, and intra-state borders. Here, 25% of social distancing implies the closure of worship centers and avoidance of gatherings of 50 persons and above. 75% includes closure of public transportation and all markets, in addition to the initial conditions of 25%. The effects of various interventions on the R0 of COVID-19 are presented. A higher percentage of social distancing appears to be more effective in controlling the spread of COVID-19 in Nigeria than border closure. Full compliance with social distancing is encouraged. According to our model, it is easier to enforce 75% closures than 25%, as the percentage of the population complying to social distancing is higher when a 75% rate is used. The minimum requirement of the population percentage that needs to abide by the social distancing advice, to weaken the epidemic can be obtained from the model. | 3,623.2 | 2020-09-05T00:00:00.000 | [
"Mathematics",
"Medicine",
"Environmental Science"
] |
Minimal residual disease and circulating tumor cells in breast cancer
Tumor cell dissemination in bone marrow or other organs is thought to represent an important step in the metastatic process. The detection of bone marrow disseminated tumor cells is associated with worse outcome in early breast cancer. Moreover, the detection of peripheral blood circulating tumor cells is an adverse prognostic factor in metastatic breast cancer, and emerging data suggest that this is also true for early disease. Beyond enumeration, the characterization of these cells has the potential to improve risk assessment, treatment selection and monitoring, and the development of novel therapeutic agents, and to advance our understanding of the biology of metastasis.
Introduction
Breast cancer (BC) is the most common cancer in women in Europe [1]. Despite surgery and adjuvant systemic therapy, many women with early BC still relapse and die of their disease. Minimal residual disease (MRD) after potentially curative surgery for BC is thought to contribute to disease relapse and to be the target of adjuvant treatment. MRD is defi ned as micrometastatic cells undetectable by conventional imaging and laboratory tests. Surrogates of MRD are tumor cells detected in the bone marrow (disseminated tumor cells (DTCs)) and peripheral blood (circulating tumor cells (CTCs)) [2]. Th e detection and characterization of DTCs/CTCs are expected to lead to personalized treatment strategies and accelerate the development of novel therapeutic agents for BC [2]. Furthermore, genotypic and phenotypic characterization of DTCs/CTCs at the single cell level may provide novel insights into the biology of tumor progression [3].
Detection methods
Th e detection of DTCs/CTCs in BC is challenging since these cells are rare, occurring at a frequency of one tumor cell per 10 6 to 10 7 mononuclear cells. To isolate DTCs/ CTCs, enrichment techniques are therefore typically applied. Th ese techniques are based either on the physical properties of the cells (for example, cell density by fi coll centrifugation or cell size by fi ltration) or on their immunological characteristics (for example, cell surface antigens of DTCs/CTCs by immuno enrichment or markers of hematopoietic cells by immunodepletion). Ficoll centrifugation was widely used in the initial clinical studies of bone marrow DTCs [4]. Currently, however, enrichment techniques incorporating immunomagnetically labeled monocolonal antibodies are more often used because they improve tumor cell recovery (recovery rates of >50% to 85%) [5,6] over fi coll enrichment (recovery rate of 40%) [7] in spiking experiments using cell lines. After the initial enrichment step, DTCs/CTCs have been detected using assays based on either antibodies (immuno cytochemistry, immuno fl uorescence) or nucleic acids (mRNA transcripts by reverse transcription PCR (RT-PCR)). Table 1 summar izes the main technologies for CTC detection in breast cancer. detection that uses microscopic ferrofl uids coated with an antibody against EpCAM to magnetically separate epithelial cells from whole blood [10]. Captured cells are stained with antibodies specifi c for cytokeratins 8, 18 and 19 (pan-CK) and CD45 (specifi c for leucocytes) and stained with 4'6-diamidino-2-phenylindole-2 (DAPI; to confi rm the presence of a cell nucleus). A CTC is defi ned as a cell staining for pan-CK and DAPI, but not for CD45. Currently, CellSearch® is the only technology that has received US Food and Drug Administration (FDA) approval for CTC detection as an aid in monitoring patients with metastatic breast, colorectal and prostate cancer [10][11][12]. Th e performance of CellSearch® for CTC detection in metastatic solid tumors has also been validated in ring studies [13,14].
Other technologies include the MagSweeper, which uses immunomagnetic separation and gently enriches target cells by 10 8 -fold from blood, eliminating cells that are not bound to magnetic particles [6]. Th is process has been shown to keep cell function intact and not to perturb rare cell gene expression [6]. CTCs have been also detected using multi-parameter fl ow cytometry, and their detection with this technology was associated with poor outcome in women with BC [15]. MAINTRAC®, another method, detects circulating epithelial tumor cells from whole unseparated blood, and uses a laser scanning cytometer after staining with anti-human epithelial and anti-CD45 fl uorescent antibodies [16]. Th is technology results in CTC counts up to 10 5 per milliliter of blood in all women with early BC, consequently raising concerns about the specifi city of the method to detect tumor cells [16].
Advances in optical technologies have also improved DTC/CTC detection. Several slide-based automated microscopic scanning devices, such as the Ikoniscope® [17] and the Ariol® system [18,19], have been applied for standardized micrometastatic cell detection and characteri zation. Another approach has been developed that uses fi ber-optic array scanning technology (FAST) for rare cell detection [20]. It has been demonstrated that FAST cytometry is capable of a 500-fold increase in speed over automated digital microscopy, with com para ble sensitivity and superior specifi city [20]. Th e combination of FAST and automated digital microscopy has allowed investigators to detect rare epithelial cells from whole unseparated blood after immunofl uorescence staining with a pan-CK antibody.
Nucleic acid-based assays
Nucleic acid-based assays have been initially hampered by false-positive results due to inability to assess tumor cell morphology, expression of target genes in normal cells, and the presence of pseudogenes (genes without protein-coding abilities) [21]. Newer quantitative assays have addressed some of these problems. To detect DTCs/ CTCs in breast cancer, nucleic acid-based assays, either as single genes or as part of multiplex assays [22][23][24][25][26][27], have mainly used CK19, mammaglobin-A (MGB1), HER2 and mucin 1 (MUC1) mRNA. Th e AdnaTest® BreastCancerSelect (AdnaGen AG, Langenhagen, Germany) is a commercially available molecular assay that utilizes immunomagnetic separation with antibodies against MUC1 and EpCAM followed by a multiplex RT-PCR for HER2, MUC1 and EpCAM [28].
Emerging detection technologies
Beyond immunomagnetic separators, microfl uidic devices have been developed for rare cell tumor capture, and these involve non-electrokinetic methods, such as immobili zation via antibody [29] and size-based sorting [30,31], or electrokinetic methods (for example, dielectro phoresis) [32]. An example of a microfl uidic platform is the 'CTC-chip' , which is capable of effi cient and selective separation of viable CTCs from peripheral whole blood samples, mediated by the interaction of target CTCs with EpCAM-coated microposts under precisely controlled laminar fl ow conditions [29]. A direct comparison between CellSearch® and two commercially available CTC-chips showed that these platforms provided similar sensitivity and yield in patient samples [33]. Stott and colleagues [34] recently reported improved sensitivity of the CTC-chip for CTC detection in patients with localized prostate cancer. Several other assays have also been developed. For example, a technique named EPISPOT (epithelial immuno spot) allows detection of viable DTCs and CTCs owing to their ability to secrete individual proteins after 48 hours of short-term culture [35]. A functional cell separation method called CAM, or the collagen adhesion matrix assay, was reported to detect CTCs with the invasive phenotype and to explore their molecular features [36]. Beyond these assays, new imaging procedures have been developed for the in vivo detection of CTCs [37].
Several investigators have also evaluated the potential utility of circulating cell-free DNA, either as a surrogate to monitor MRD [38], or as a 'liquid biopsy' for real-time monitoring of tumor mutations in cancer patients [39]. Moreover, some investigators have been able to identify patient-specifi c genomic rearrangements in plasmacirculating DNA as a way to monitor MRD [40]. Th ey employed next-generation sequencing to rapidly identify patient-specifi c genomic rearrangements in primary tumors and showed that PCR assays could reliably detect these rearrangements in plasma [40]. A recent review has summarized advances in cell-free nucleic acids (DNA, mRNA, microRNA) as potential biomarkers in cancer [41].
Critical interpretation of detection technologies
Th e diff erent technologies use diff erent enrichment and detection steps and therefore do not always detect the same CTC population (Table 1). In a study comparing two commercially available assays (CellSearch® and AdnaTest®) in the same metastatic BC patient samples, the concordance between the two assays was 64% for CTC detection and 50% for HER2-positive CTC detection [42]. Th erefore, it is important to study the clinical utility of the assay-dependent CTC detection and characterization. Moreover, most enrichment methods used by the diff erent assays are biased because they result in loss of a fraction of CTCs due to tumor cell heterogeneity. As an example, some available technologies detect only EpCAM+ CTCs (Table 1). However, it has been shown that BC cell lines with low EpCAM expression and high expression of mesenchymal markers cannot be effi ciently captured using a purely EpCAM-based mechanism [33,43,44]. Some other technologies are using enrichment based on red blood cell lysis or leukocyte depletion (CD45-negative depletion) aiming at a less biased CTC enrichment (Table 1). Another critical issue with all cell detection technologies is that blood cannot be stored and must be processed soon after it has been drawn, within up to 72 hours [13] depending on the technology used. Th ere fore, the clinical validation of CTCs depends on the availability of detection technologies in diff erent labs. Th is is a major diff erence between CTCs and biomarkers from paraffi n-embedded primary tumor blocks, for which real-time processing is not mandatory. Since all currently available platforms will continue to evolve rapidly, the challenge will be to prospectively evaluate the utility of each technology to address specifi c clinical questions.
Clinical relevance of DTCs/CTCs
CTCs and DTCs were cited for the fi rst time in the 2007 recommendations of the American Society of Clinical Oncology (ASCO) on tumor markers [45]. Recently, in the 7th edition of the American Joint Committee on Cancer Staging Manual (2010), a new M0(i+) category was proposed for TNM (tumor, node, metastasis) staging in BC [46]. Th is new category is defi ned as no clinical or radiographic evidence of distant metastases, but deposits of molecularly or microscopically detected tumor cells (no larger than 0.2 mm) in blood, bone marrow, or other non-regional nodal tissue in a patient without symptoms or signs of metastases.
Clinical relevance of DTCs
Th e inclusion of the M0(i+) category was driven at least in part by a pooled analysis of individual data from several studies, which showed that bone marrow CKpositive DTCs were detected at the time of surgery in 30.6% of 4,703 patients with invasive BC [4]. Bone marrow DTCs were signifi cantly more frequent in women with larger tumors, or tumors with higher histologic grade, hormone receptor negativity, and lymph node metastasis. In multivariate analysis, the presence of bone marrow DTCs predicted for signifi cantly higher risk of death from BC [4]. Recently, in the American College of Surgeons Oncology Group's (ACOSOG) Z0010 multicenter trial, bone marrow DTCs were identifi ed at surgery by immunocytochemistry in only 105 of 3,491 patients (3%) with clinical T1/T2 N0 M0 BC [47].
Although the DTC detection rate was very low, bone marrow DTCs still signifi cantly predicted decreased overall survival [47]. A pooled analysis of individual patient data from 676 women with stage I-III BC from three studies showed that bone marrow DTCs were detected in 15.5% of patients at a median 37-month follow-up after diagnosis [48]. Th e presence of DTCs was an independent indicator of poor prognosis and could be used to select patients for secondary adjuvant treatment strategies [48].
Clinical relevance of CTCs (CellSearch®)
Using CellSearch®, ≥5 CTCs/7.5 ml of blood were detected in 49% of 177 patients with measurable metastatic BC before a new treatment was started [10]. CTC detection was an independent predictor of progressionfree survival and overall survival [10]. Th is and other studies [49][50][51][52] have provided solid data about the adverse prognostic value when CTCs are detected by CellSearch® in metastatic BC.
Detecting CTCs in non-metastatic BC is more challeng ing because these cells occur at a very low frequency in this setting. Pierga and colleagues [53] found ≥1 CTC/7.5 ml in 23% of 97 patients before administrating neoadjuvant chemotherapy (NAC) and in 17% of 86 patients after NAC. Th e detection of ≥1 CTC/7.5 ml before NAC, after NAC, or at both time points in the above study was associated with worse distant metastasis-free survival and overall survival at a median follow-up of 36 months [54]. In another study ≥1 CTC/7.5 ml were detected in 21.6% of 213 patients before NAC and in 10.6% of 207 patients after NAC [55]. In both these studies, however, neither CTC detection before or after NAC, nor changes in CTC detection during treatment, were predictive of pathological complete response [53,55]. Rack and colleagues [56] detected ≥1 CTC/22.5 ml before the start of adjuvant treatment in 21.5% of 2,026 patients with early BC [56]. In this study, pretreatment detection of CTCs was confi rmed as an indepen dent predictor for both disease-free survival and overall survival [56]. Several other investigators have detected CTCs by CellSearch® in 9% to 38% of patients with early BC without reporting survival data [57][58][59].
Th ese diff erences in CTC detection rate in early BC could be attributed to the Poisson distribution of rare events [60], to diff erences in patient populations, sampling time points, blood volume analyzed, the use or not of fi coll enrichment before processing with CellSearch®, and diff erences in image interpretation between diff erent labs. Most women in this setting have only one detectable CTC/whole blood volume analyzed. Th erefore, in order to prospectively test potential clinical applications of CTCs in non-metastatic BC, it is important to standardize image interpretation across labs by taking into account cytomorphologic criteria.
CTCs versus DTCs
Since blood is more easily obtained than bone marrow, an important question is whether peripheral blood CTCs can be used as surrogate markers for bone marrow DTCs. In one study, peripheral blood and bone marrow were collected from 341 patients at a median follow-up of 40 months after initial surgery [61]. In this study, 8 patients were CTC+/DTC+, 26 were CTC+/DTC-, and 40 were CTC-/DTC+. Although both CTCs (10% of the patients) and DTCs (14% of the patients) were significantly associated with worse clinical outcome, DTCs were more informative than CTCs [61]. Th is and other studies [62] showed that there was no good correlation between CTC and DTC detection. However, it is not clear whether this is because CTCs and DTCs represent diff erent tumor cell populations or whether this is also related to limitations of the detection technologies used. At present there are no data to support that CTCs can replace DTCs.
Clinical relevance of nucleic acid-based assays
In early BC, initial single-center studies have reported that the detection of peripheral blood CK19 mRNA by RT-PCR after fi coll enrichment of mononuclear cells was an independent prognostic factor for reduced diseasefree survival and overall survival [63,64]. In another study, 13% of 431 early BC patients were CTC-positive according to the AdnaTest®; however, no correlation with clinical outcome was reported [65]. In metastatic BC, CTC detection by AdnaTest® was reported in 52% of 42 women and predicted therapy response in 78% of cases [66]. Finally using immunomagnetic tumor cell enrichment and a multi-marker quantitative PCR based assay, CTCs were detected in 7.9% of 733 stage I/II breast cancer patients with a median follow-up time of 7.6 years and their detection was an independent predictor of metastasis-free survival and breast cancer specifi c survival [67]. However, despite these initial results, no nucleic acidbased assay has received FDA approval nor has demonstrated utility in treating patients with BC.
Clinical trials with DTCs/CTCs
Interestingly, CTC or DTC clearance after systemic treatment has been used as an endpoint in BC clinical trials. In one single-center study, it was shown that a short course of trastuzumab (3 cycles every 3 weeks) eliminated chemotherapy-resistant CK19 mRNA-positive cells in peripheral blood or bone marrow in 20 of 30 women with stage I-IV BC [68]. Another study randomized women with stage II and III BC to NAC with or without zolendronic acid [69]. Th e primary endpoint of the trial was the number of patients with detectable DTCs at 3 months' post-treatment. At 3 months, DTCs were detected in 17 of 56 patients receiving zoledronic acid versus 25 of 53 patients who did not. Although fewer women had detectable DTCs after NAC with concurrent zoledronic acid than with chemotherapy alone, this was not the case when only women who tested DTC-positive at baseline were analyzed. A critical question is if CTC clearance can be used as a 'surrogate' for survival for regulatory purposes. Such an eff ort is ongoing and investigators are studying CTC detection by CellSearch® before and after treatment in the phase 3 registration trials of abiraterone acetate in prostate cancer [70].
Although data on the adverse prognostic value of CTC d etection by CellSearch® in metastatic BC are solid, evidence from prospective trials is needed that CTC detection can lead to changes in treatment decision and thus improve clinical outcome in metastatic BC. Such an eff ort is ongoing in a phase III trial run by the Southwest Oncology Group, which is testing the strategy of changing chemotherapy versus continuing the same chemotherapy for patients with metastatic BC who have elevated CTC levels at their fi rst follow-up assessment (ClinicalTrials.gov NCT00382018).
Identifi cation of therapeutic targets
Beyond enumeration, further characterization of DTCs and CTCs holds the promise to improve treatment outcome in women with BC. Because of the availability of anti-HER2 agents, HER2 expression was studied on DTCs [71][72][73] and CTCs [33,42,55,59,66,[74][75][76][77][78] (Table 2) and was correlated with HER2 expression on the primary tumor. In most studies, HER2 expression on DTCs/CTCs is more prevalent in women with HER2-positive BC than in women with HER2-negative BC in both non-metastatic and metastatic settings. Interestingly, among women with HER2-negative primary tumors defi ned by standard pathology and detectable CTCs, between 14% and 50% may have at least one HER2-positive CTC. However, it is not known whether the discordant cases can be attri buted to technical causes or whether there is any underlying biological explanation. Clinical testing for HER2 in the primary tumor is known to result in false-negative and false-positive results [79]. Furthermore, in most cases diff erent technologies are used to evaluate HER2 in the primary tumor and the CTCs. Beyond technical issues, functional HER2 protein up-regulation on CTCs cannot be excluded, and the acquisition of HER2 amplifi cation during the course of the disease has been suggested [74]. It was shown that four out of nine patients with metastatic BC whose primary tumors were HER2-negative and who had CTCs showing HER2 gene amplifi cation derived benefi t from trastuzumab-containing therapy [74].
Beyond HER2, several other markers have been studied on DTCs/CTCs. Markers related to angiogenesis, such as vascular endothelial growth factor (VEGF), VEGF2, and hypoxia inducible factor (HIF)-1α, were observed in CTC-positive samples from metastatic BC patients [80]. Using the CTC-chip technology to purify CTCs, epidermal growth factor receptor mutations conferring drug resistance were detected in CTCs from non-small-cell lung cancer patients who had received tyrosine kinase inhibitors [81]. Androgen receptor mutations were also identifi ed in CTC-enriched peripheral blood samples from castration-resistant prostate cancer patients [82]. In most of these studies, DTC/CTC characterization was performed in few patients in the metastatic setting, and therefore validation in independent larger patient series is required. Characterizing DTCs/CTCs in nonmetastatic tumors poses additional challenges since such cells are only rarely detected in this setting. Finally, clinical trials are needed to demonstrate that CTC characterization is important for patient management.
Identifi cation of DTCs/CTCs with 'tumor-initiating cell' phenotype
Beyond the potential for improving patient outcome, the study of DTCs/CTCs aims to lead to a better under standing of the metastatic process. Research has shown a signifi cant proportion of DTCs to be resistant to conventional chemotherapy [48]. Furthermore, using Ki67 immuno staining, most micrometastatic cells have been found to be in a non-proliferative state [83]. Interestingly, the CD44 + CD24 -/low tumor-initiating cell phenotype [84] was observed in a signifi cant number of bone marrow DTCs using triple-staining by immunocytochemistry [85]. Moreover, the CK19+/MUC1 stem cell-like phenotype was demonstrated in a signifi cant number of DTCs in BC by the EPISPOT assay [35]. Epithelial mesencymal transition markers and aldehyde dehydrogenase 1 (ALDH1) were also identifi ed in a major proportion of CTCs from patients with metastatic BC [86]. However, clinical studies are needed to associate the presence of CTCs/DTCs with tumor-initiating cell phenotype with clinical outcome in women with BC.
Tumor dormancy
An issue related to the role of DTCs in the metastatic process is determining which of them will grow into overt metastases and which will not. According to clinical studies on bone marrow DTCs, 50% to 70% of patients with detectable DTCs will not develop metastases, although even patients without DTCs may relapse and die of BC [4]. For patients who relapse without such cells detectable in their bone marrow, it is possible that the DTCs have actually settled into other organs; alternatively, lack of DTCs could be the result of sampling error or refl ect the suboptimal sensitivity of CKs as a marker for DTC detection. Indeed, tumor cell dissemination has been linked to epithelial mesencymal transition and the down-regulation of epithelial cell markers [87]. Conversely, it is possible that non-relapse in the case of patients with DTCs/CTCs can be attributed to the detection of apoptotic cells or to tumor dormancy. Interestingly, CTCs have been detected in one-third of women without clinical evidence of disease up to 22 years after mastectomy for BC [88].
Th ere is evidence that several mechanisms of dormancy exist, including cellular dormancy, in which DTCs enter a state of quiescence (G0-G1 arrest), and tumor mass dormancy, in which DTCs divide but the lesion does not grow beyond a certain size [89]. Th ere is also evidence that mechanisms regulating the switch between cellular dormancy and escape from it are related to the cross-talk between DTCs and the microenvironment [89,90]. For example, loss or absence of a surface receptor like HER2, urokinase-type plasminogen activator receptor (uPAR) or integrins that transduce growth signals from the microenvironment may result in a dormant DTC, whereas the presence of such a receptor and a permissive microenvironment may result in a proliferating DTC. Interestingly, the overexpression of HER2 [71] or uPAR [91] on DTCs was associated with poor prognosis in patients with breast and gastric cancer, respectively.
Th e mechanisms that regulate the switch between tumor mass dormancy and expansion have been suggested to be related to angiogenesis [92] and the immune system [93]. When there are limitations in blood supply and/or when there is an active immune system, for example, the micrometastasis cannot grow into overt metastasis. By contrast, a shift in favor of pro-angiogenic factors and activation of transcriptional programs that allow the recruitment of new blood vessels (angiogenic switch) or an escape of immune surveillance (immunosupression) may cause the expansion of the micrometastatic cells into macrometastasis. It is not clear how all these mechanisms operate in a given patient, or how they are infl uenced by exogenous factors like stress and diet or by host genetic factors [94].
Genomic characterization of DTCs/CTCs
Th us far, only limited information is available about the global gene expression programs that determine the fate of DTCs and CTCs. Some studies have performed molecular characterization of CTC-enriched samples and reported mRNA or microRNA expression of CTCspecifi c genes in metastatic BC [95,96]. Using single cell comparative genomic hybridization, it has been shown that bone marrow DTCs are genetically heterogeneous and display fewer genetic aberrations than primary tumor cells [97][98][99]. In addition, the most prevalent chromo somal aberrations of primary breast tumors (including 8q gain, 13q loss, 16q loss and 17p loss) have rarely been found in DTCs with abnormal karyograms isolated at the time of curative surgery [3]. Husemann and colleagues [100] provided evidence that systemic spread occurs early in BC by showing that tumor cells can disseminate from earliest breast epithelial alterations in transgenic mice and from breast ductal carcinoma in situ in women. Th ese results have led to the proposal of a parallel progression model in which tumor cells disseminate early at ectopic sites and evolve in parallel with tumor cells in the primary site [3]. Finally, beyond gene expression profi ling and comparative genomic hybridization, the characterization of DTCs/CTCs using next generation sequencing may provide new insights into the cellular programs that regulate tumor dormancy and metastasis.
Future directions
Th e characterization of DTCs/CTCs might lead to the identifi cation of targets for the design of new drugs. CTCs might also be used to accelerate drug development if ongoing or future trials demonstrate that CTC clearance is a 'surrogate' for drug effi cacy. In order to move DTCs/CTCs into clinical practice, prospective trials with innovative designs and endpoints are needed to demonstrate both clinical utility and cost-eff ectiveness. Such eff orts are currently ongoing. Because the techno logies used to detect and characterize tumor cells in peripheral blood are rapidly evolving, issues like easy access to newer technologies and standardization across laboratories will be critical for prospective validation. CTC detection and characterization have the potential to improve risk assessment and provide a 'liquid biopsy' for real-time monitoring of tumor genotype/phenotype in metastatic BC. In early BC, the presence of MRD after patients have completed standard adjuvant treatment may contribute to a better selection of patients to evaluate secondary adjuvant treatment strategies. Overall, the integration of information from both the primary tumor and MRD may eventually lead to personalized treatment strategies.
Competing interests
MI has received consultancy fees and research support from Veridex. MR has no fi nancial competing interests. | 5,931.8 | 2011-10-25T00:00:00.000 | [
"Biology",
"Medicine"
] |
Protection against Experimental Stroke by Ganglioside GM1 Is Associated with the Inhibition of Autophagy
Ganglioside GM1, which is particularly abundant in the central nervous system (CNS), is closely associated with the protection against several CNS disorders. However, controversial findings have been reported on the role of GM1 following ischemic stroke. In the present study, using a rat middle cerebral artery occlusion (MCAO) model, we investigated whether GM1 can protect against ischemic brain injury and whether it targets the autophagy pathway. GM1 was delivered to Sprague-Dawley male rats at 3 doses (25 mg/kg, 50 mg/kg, 100 mg/kg) by intraperitoneal injection soon after reperfusion and then once daily for 2 days. The same volume of saline was given as a control. Tat–Beclin-1, a specific autophagy inducer, was administered by intraperitoneal injection at 24 and 48 hours post-MCAO. Infarction volume, mortality and neurological function were assessed at 72 hours after ischemic insult. Immunofluorescence and Western blotting were performed to determine the expression of autophagy-related proteins P62, LC3 and Beclin-1 in the penumbra area. No significant changes in mortality and physiological variables (heart rate, blood glucose levels and arterial blood gases) were observed between the different groups. However, MCAO resulted in enhanced conversion of LC3-I into LC3-II, P62 degradation, high levels of Beclin-1, a large area infarction (26.3±3.6%) and serious neurobehavioral deficits. GM1 (50 mg/kg) treatment significantly reduced the autophagy activation, neurobehavioral dysfunctions, and infarction volume (from 26.3% to 19.5%) without causing significant adverse side effects. However, this biological function could be abolished by Tat–Beclin-1. In conclusion: GM1 demonstrated safe and robust neuroprotective effects that are associated with the inhibition of autophagy following experimental stroke.
Introduction
Ischemic stroke is a consequence of vascular occlusion.The result is irreversible neural tissue injury, which is a major cause of death and disability worldwide.As the world's population continues to age, the incidence of stroke is expected to grow, increasing the interest and need for novel approaches focused on improving the recovery and quality of life of stroke patients.
A key complication of ischemic stroke is neuron damage secondary to excitotoxicity, acute energy failure, or programmed cell death under ischemia/reperfusion conditions [1,2].Increasing evidence demonstrates that prolonged autophagy, a non-apoptotic route of type II programmed cell death, plays a role in cerebral ischemic injury amongst a variety of other neurologic conditions [3][4][5].Autophagy is a programmed cell survival process that generally mediates breakdown and recycling of cellular components such as long-lived proteins and damaged organelles.However, persistent stress can promote high levels of autophagy resulting in cell death [6][7][8].
Massive activation of neuronal autophagy and its associated markers, LC3 and Beclin-1, have been firmly established in a variety of focal ischemic stroke models [5,9].Another notable factor includes the gangliosides.While GM2 and GM3 are induced transiently within the ipsilateral hemisphere after the induction of ischemic injury in the mouse [10], GM1, an important component of lipid rafts, can act as a neurotrophic factor [11][12][13][14].GM1 has been shown to potentiate the action of neurotrophins and display a wide variety of central nervous system functions including promoting survival, differentiation [15], neurodegeneration [14,16,17], axon stability, and regeneration [18].A plethora of studies have suggested that GM1 may be involved in the stroke process, specifically the orchestration of cell death and subsequent neurological dysfunctions [19].However, the precise mechanism of action remains inconclusive.
In the present study, the role of GM1 in ischemic stroke and its potential regulation of neuronal autophagic activities were investigated.We hypothesized that GM1 could improve neurological outcomes via the inhibition of excessive autophagy in stroke development.Tat-Beclin-1 peptide was used to induce autophagy, and the effects of GM1 on infarction volume, mortality rate and neurological deficits were assessed.Additionally, the expression of P62, LC3, and Beclin-1 were measured as a potential mechanism of GM1 in conferring neuroprotective properties following ischemia-reperfusion injury.
Animals and treatments
All experimental procedures using rats in this study strictly followed recommendations provided by the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Bioethics Committee of Capital Medical University, Beijing, China.Sprague-Dawley male rats weighing 280~300 g were housed in a 12-hour light/dark cycle at a controlled temperature and humidity with free access to food and water.All surgery was performed under chloral hydrate anesthesia, which was specifically reviewed and approved by the ethics committee, and all efforts were made to minimize suffering.During surgery, rats were anesthetized with chloral hydrate (0.4 g/kg, i.p.), which was selected because, unlike alternative anesthetics, it has not been shown to upregulate autophagy [20].Then rats were subjected to focal cerebral ischemia by intraluminal middle cerebral artery occlusion (MCAO) with a nylon suture as previously described [21,22].After 2 hours of MCAO, the suture was withdrawn to allow reperfusion.The heart rate, blood glucose levels, and blood gases were monitored before, during, and after ischemia.Core body temperature was maintained at 37°C.Sham-operated rats underwent a similar procedure with the exception of nylon suture occlusion and reperfusion.Ten rats that died during the operation under anesthesia were not included.
Rats were monitoring hourly for the first 4 hours after the MCAO operation.The rats were then observed twice daily for the next 2 days, once in the morning and once in the afternoon.Endpoints used included heart rate, blood pressure, and temperature.Post-operative care consisted of monitoring heart rates, use of a heating pad to alleviate cold temperatures, oxygen therapy for poor ventilation, closely monitoring the incision site for hemorrhage and infection, and administration of pain killers in response to characteristic changes in behavior and/or dietary habits.Fifteen rats died post-operatively.The surviving rats were neurologically tested by a masked investigator and euthanized 72 hours after MCAO.The brains were removed for the preparation of slices or lysates.
TTC Staining
The infarct size of the ischemic cortex was measured as described previously [23].In brief, 2-mm thick coronal sections were dissected using a rat brain slicer (Matrix, ASI Instruments, Houston, TX, USA) 72 hours after MCAO.The slices were incubated in 2% 2,3,5-triphenyltetrazolium chloride (TTC) solution (Sigma-Aldrich Chemical Co., St. Louis, MO, USA) at 37°C for 30 minutes, and fixed with 4% paraformaldehyde in PBS (pH 7.4) at 4°C for 6 hours.Normal tissue stains red while infarcted tissue with absent mitochondrial enzyme activity fails to stain and appears white.The infarct volume of each section was traced and measured by manually outlining the margins of non-ischemic areas using an image analysis system (Imaging-Pro-Plus, Silver Spring, MD, USA).To account for the increase in brain volume secondary to post-ischemic brain edema, infarct size was normalized to the contralateral cortex and expressed as a percentage according to the following formula: Infart Volume ¼ Left Hemisphere Volume À Right Nonischemic Volume Left Hemisphere Volume x 100
Mortality and Neurobehavioral Deficits
Mortality was calculated at 72 hours after MCAO.Neurobehavioral deficits were assessed and scored based on the scoring system of Garcia et al in a blinded fashion [24].
Immunofluorescence Staining
Animals were deeply anesthetized with chloral hydrate and perfused transcardially with saline followed by 250ml cold 4% paraformaldehyde (PFA) in 0.1M phosphate buffer (PB; pH 7.4).Brains were dissected, post-fixed in PFA for 12~18 hours, and transferred to 30% sucrose in 0.1 M PB for at least 48 hours at 4°C.Serial sections of the entire brain were cut at 10μm thickness on a cryostat (Leica CM3050S).Double fluorescence labeling was performed as described previously [25].In brief, sections were rinsed 3 times in 0.01M PBS for 5 min each, permeabilized with 0.3% Triton X-100 in PBS for 30 min at 37°C, blocked with 1% BSA in PBS for 2 hours at room temperature, and incubated with the following primary antibodies: mouse anti-Beclin-1, and rabbit anti-LC3-II (Santa Cruz Biotechnology, Inc., CA, USA) at a concentration of 1:50~1:200 diluted in 0.01M PBS in a humidified chamber overnight at 4°C.Sections were then washed with 0.01M PBS and incubated for 2 hours with the secondary antibodies (anti-mouse IgG labeled with Alexa Fluor-488, and anti-rabbit IgG labeled with Alexa Fluor-568, 1:200, Jackson ImmunoResearch Laboratories, Inc., West Grove, PA, USA) at room temperature.Tissue sections were washed 3 times in 0.01M PBS for 5 min each, and then incubated with 2μg/ml Hoechst 33258 (Sigma Aldrich, Inc., St-Louis, MO, USA) for 10 min at room temperature to contain nuclei.Primary antibody incubation was omitted in some sections as a negative control.Images were acquired using an OLYMPUS BX51 microscope.
Western Blotting
Animals were sacrificed for tissue harvest at 72 hours after MCAO.Tissue from the brain cortex, including the penumbra area, were dissected and immediately frozen in liquid nitrogen and stored at -80°C.Samples were homogenized in RIPA buffer (Santa Cruz) with protease inhibitor cocktail (Sigma) and incubated on ice for 30min.Lysates were centrifuged at 14000 x g for 25min at 4°C.Total protein concentrations were then measured using the Bradford assay (Santa Cruz).Supernatants were mixed with an equal volume of 2x Laemmli buffer (Santa Cruz) and denatured at 95°C for 5min.Protein samples (50 μg) were loaded on polyacrylamide gels, electrophoresed, and transferred to 0.45μm nitrocellulose membranes (Bio-Rad).The membranes were then blocked with 5% milk (Sigma) in 1xTBST for 2 hours at room temperature, followed by incubation with the primary antibodies (mouse anti-P62, rabbit anti-LC3-II/ LC3-I or mouse anti-Beclin-1, Santa Cruz) at 1:1000 dilution in 3% milk overnight at 4°C.The membranes were then washed 3x5min in 1xTBST and probed with the secondary antibodies (goat anti-rabbit IgG, goat anti-mouse IgG, Santa Cruz, 1:8000) for 1 hour at room temperature.Immunoblots were then washed and probed with an ECL Plus chemiluminescence reagent kit (Amersham Biosciences, Arlington Heights, IL, USA) for 5min and detected by exposure to Kodak X-ray film.The optical density of the bands was quantified by Image J software.β-Actin monoclonal antibody (goat anti-β-Actin, Santa Cruz, 1:5000) was also probed, and used as a sample loading control for normalization, and the data were expressed as the ratio to β-Actin.
Data Analysis
The analysis of the data was performed using SigmaPlot software.Data are expressed as the mean ± standard error of the mean and analyzed with one-way analysis of variance (ANOVA) followed by the Student-Newman-Keuls method.The neurobehavior scores were analyzed with Kruskal-Wallis one-way ANOVA followed by multiple comparison procedures by the Dunn method.Data are expressed as the median ± 25th-75th percentiles.Statistical significance was defined as P< 0.05.
GM1 Does Not Significantly Affect Mortality or Physiological Variables of MCAO Rats
To investigate the potential for GM1 to protect against ischemic brain injury and whether this protection involves the autophagy pathway, we employed a rat MCAO model.Additionally, no significant changes of physiological variables (heart rate, blood glucose levels, blood gases analysis [pH, pCO 2 , pO 2 , Na + , Ca 2+ , Cl -, BUN, Glu, Hct, HB, BE, tCO 2 , BB,
GM1 (50 mg/kg) Improves Neurobehavioral Functions and Reduces Cerebral Infarction
To evaluate the sensorimotor deficits after MCAO, a modified Garcia test was performed at 72 hours following MCAO (Fig 1B).No deficits were observed in sham animals.In contrast, a significant decline of neurological scores was detected in the MCAO +saline group (P<0.05 vs. sham).After both the medium-dose (50 mg/kg) and high-dose (100 mg/kg) GM1 treatment, there was a statistically significant improvement in neurobehavioral function when compared to MCAO +saline group (P<0.05).
The cerebral infarction at 72 hours after MCAO was also evaluated using TTC staining.Representative samples of TTC-stained brain sections with corresponding infarction volumes are shown in Fig 1C .The infarction volume revealed a marked increase in the MCAO group compared to the sham animals (26.3±3.6% vs. 0.2±0.2%,P<0.05).However, the infarct ratios were decreased from 26.3% to 19.5% in rats with the medium-dose GM1 (50 mg/kg) treatment (P<0.05 vs. MCAO).Although the brain infarction showed a trend toward reduction following low-dose (25 mg/kg) and high-dose (100 mg/kg) administration, there was no statistical significance reached (25 mg/kg, 24.7±3.0%, 100 mg/kg 24.6±4.8% vs. MCAO 26.3±3.6%,P>0.05).Overall, our results demonstrate that injection of GM1 (50 mg/kg) improves neurobehavioral functions and reduces infarct volume after ischemic insult despite having no effect on the mortality rate in the MCAO model.Based on these findings, a dose of 50 mg/kg of GM1 was used in subsequent studies.
GM1 Decreases LC3-II and Beclin-1 Levels and Increases P62 Levels Following MCAO Injury
MAP1LC3B/LC3 and SQSTM1/p62 are most frequently used as markers to measure autophagy flux.LC3, the microtubule-associated protein light chain 3, exists in a cytosolic form (LC3-I) and an autophagosome-associated membrane-bound form (LC3-II).The ratio of conversion from LC3-I to LC3-II is closely correlated with the extent of autophagosome formation [26].P62, a selective autophagy substrate, was originally discovered as a scaffold in signaling pathways regulating cell growth and proliferation; however, it was also determined to bind to several autophagy substrates, such as ubiquitinated proteins, damaged mitochondria and signaling molecules, promoting their autophagic clearance [27].Beclin-1, another key factor in the autophagic process, is essential for the recruitment of other autophagic proteins during the expansion of the pre-autophagosomal membrane [28,29].To determine the effect of GM1 on neuronal autophagic activities after MCAO, we first measured the LC3-II and Beclin-1 expression levels in brain tissue by immunostaining 72 hours after brain ischemia and quantified both the LC3-II and the Beclin-1 positive cells in the penumbra area ( To verify these findings, we performed Western Blotting analysis of LC3, Beclin-1 and P62 (Fig 3).MCAO promoted a significant increase in LC3-II and Beclin-1 levels (P<0.05 vs. sham), which was obviously suppressed by subsequent GM1 injection (P<0.05 vs. MCAO), though there was no significant difference in LC3-I levels between sham and MCAO or GM1 rats (Fig 3B and 3D).Conversely, the expression of P62 was significantly decreased following MCAO and markedly increased with GM1 treatment compared to MCAO rats (P<0.05,Fig 3C).These findings confirm the role of GM1 in attenuating the effects of MCAO in rats.
Tat-Beclin-1 Enhances Autophagy Levels and Abolishes the Neuroprotective Effects of GM1
To test whether autophagy is involved in the neuroprotective function of GM1, Tat-Beclin-1 was used to enhance autophagy post-MCAO.As expected, Beclin-1 levels were significantly increased in the MCAO+Tat-Beclin-1 rats as compared to the MCAO rats.Furthermore, Tat-Beclin-1 application resulted in an enhanced conversion of LC-1 into LC3-II and enhanced P62 degradation, confirming its role as an autophagy enhancer.Conversely, the administration of both Beclin-1 and GM1 led to a reduction in the levels of Beclin-1 expression, LC3 conversion, and P62 degradation as compared to Tat-Beclin-1 alone, suggesting that GM1 can reverse the effects of Beclin-1 on autophagy (P<0.05,Fig 4A -4F).Tat-Beclin-1 did not clearly increase the infarct volume nor worsen the neurobehavioral functions compared to MCAO rats; however, the effects of GM1 on cerebral infarction reduction and neurobehavioral improvement were abolished by Tat-Beclin-1(P<0.05,Fig 4A and 4B).These findings further suggest that modulation of the levels of autophagy by GM1 contributes to its ability to attenuate the effects of MCAO.
Discussion
Nearly 15 million people worldwide suffer from stroke each year, and this number continues to grow as a result of the aging population.The ensuing symptoms are a burden to the victims, their families, and the surrounding community (WHO The Atlas of heart disease and stroke section 15: Global burden of stroke).Despite the continued efforts to develop new pharmacological strategies, there are no effective treatment options as of yet.In the present study, to explore the potential activity and mechanism of GM1 in providing protection against ischemic brain injury, we investigated the effects of GM1 on MCAO-induced brain injury.
GM1 has been a focus of research investigations due to its function as a neurotrophic drug and its ability to penetrate the blood-brain barrier [30].Although the development of the acute inflammatory polyneuropathy Guillain-Barre´syndrome (GBS) following intravenous ganglioside treatment resulted in the withdrawal of GM1 from European market [31], this adverse effect was shown to be rare, and the relationship between exogenous gangliosides and GBS remains controversial [32,33].These drugs are still available and have been extensively prescribed in other markets including China, where a multitude of neurological maladies have been treated with gangliosides in the absence of resultant GBS or other severe adverse events [34][35][36][37].In a trail conducted by Schneider et al., a 5-year clinical course confirmed the longterm safety of GM1 therapy and suggested favorable efficacy for Parkinson's disease patients [35].Furthermore, the safety and efficacy of GM1in treating ischemic stroke has also been suggested in numerous clinical trials [38][39][40].To understand mechanisms of GM1 that contributes to its therapeutic potential, the present study sought to evaluate its effects in a stroke animal model and its dose-dependent effects on neurological improvement.We did not observe any statistically significant differences in mortality between groups.However, the medium dose (50 mg/kg) GM1 treatment significantly improved neurological performance and alleviated cerebral infarction, while the low dose (25 mg/kg) and high dose (100 mg/kg) treatments did not.The precise mechanisms responsible for the neuroprotective effects of GM1 remain uncertain, but, as demonstrated by our study, GM1 may act in part through the regulation of neuronal autophagic activity.Autophagy (self-eating) is generally viewed as a cell survival mechanism in response to various stress conditions, and occurs without typical hallmarks of apoptosis.On the other hand, enhanced autophagy can also mediate cell death in cerebral ischemia [41].Brain neuronal death following neonatal hypoxia/ischemia injury is largely prevented by Atg7 deficiency, which is essential for autophagy [42], whereas, promotion of autophagy with rapamycin augments cell death in insulin-deficient mice in adult hippocampal neural stem cells [43].In support of a role for GM1 in regulating autophagy, Batten disease (juvenile neuronal ceroid lipofuscinosis) is associated with reduced autophagy, as well as enhanced level of GM1 [44].In the present study, we demonstrated that MCAO resulted in a significant increase in LC3-II and Beclin-1 levels, which were markedly reduced by early initiation of GM1 (50 mg/kg) at 72 hours after reperfusion.Additionally, immunofluorescence staining showed a dramatic increase in the number of LC3-II and Beclin-1 positive cells post-MCAO, both of which were significantly reduced in the penumbra area after GM1 administration when compared with saline-treated rats.These findings suggest that autophagy might be involved in the neuroprotective function of GM1.
To further test whether the neuroprotective effects of GM1 might be explained in part by its ability to modulate autophagy, we tested whether Tat-Beclin-1, a specific cell-permeable autophagy-inducing peptide, can attenuate the effects of GM1.Tat-Beclin-1 was recently identified by Shoji-Kawata and colleagues by domain mapping of the autophagy protein Beclin-1 [45].It has been shown to efficiently induce autophagy in vivo and in vitro, decrease the small polyglutamine expansion protein aggregates, and reduce mortality in chikungunya or West Nile virus-infected mice by interacting with HIV-1 Nef [45,46].Consistent with these studies, in the present study, injection with Tat-Beclin-1 resulted in enhanced conversion of LC3-1 into LC3-II and P62 degradation.Furthermore, GM1 mediated cerebral infarction reduction was blocked by Tat-Beclin-1.Thus, these findings support the hypothesis that GM1 may exert its neuroprotective effects in ischemic stroke, in part, by regulating neuronal autophagic activity.
Our results show that the neuroprotective effects of GM1 only can be observed within a narrow window of concentration, and that GM1 has no statistical neuroprotective ability when both MCAO and Tat-Beclin-1 are applied, which is suggestive of a threshold of autophagy above which neurotrophic activity cannot be observed within our model.Furthermore, it is likely that other activities, in addition to autophagy, can contribute to the function of GM1.The current study has the following limitations: 1) only one autophagy inducer was included and no autophagy inhibitor was studied; 2) only short-term (3d post-stroke) effects were observed, whereas longer term (14d, 21d post-stroke) effects were not studied; 3) while the Garcia test was performed to assess neurobehavioral activity, tests for learning, memory, and social interaction were not performed.For this reason, further experimentation, including additional dosing experiments, specific autophagy inhibitor interventions, long-term studies and additional behavior studies, will be critical for determining the potential of GM1 in clinical ischemia.Nevertheless, the present results suggest that the early use of GM1 is safe and effective in rats after ischemic brain injury, and describe a new mechanism that could account for the neuroprotective effects of GM1.
Fig 1 .
Fig 1. GM1 (50 mg/kg) treatment improves neurological scores and reduces infarct volume at 72 hours after MCAO.GM1 was administrated soon after reperfusion and then once daily for 2 days at three doses (25, 50, and 100 mg/kg) by intraperitoneal injection.MCAO rats were injected with the same volume of saline as a control.(A) Experimental design and animal treatment.(B) Neurological scores for the modified Garcia test in sham, MCAO+saline and GM1 (25, 50, 100 mg/kg) treatment groups.n = 10-12 rats per group.(C) Representative images of TTC stained sections and infarct volume in sham, MCAO+saline and GM1 treatment groups.n = 6 per group.*p < 0.05 vs sham; #p < 0.05 vs MCAO+saline.doi:10.1371/journal.pone.0144219.g001 | 4,732 | 2016-01-11T00:00:00.000 | [
"Biology",
"Medicine"
] |
Technical Note: Influence of surface roughness and local turbulence on coated-wall flow tube experiments for gas uptake and kinetic studies
Guo Li, Hang Su, Uwe Kuhn, Hannah Meusel, Markus Ammann, Min Shao, Ulrich Pöschl, Yafang Cheng 5 1 Multiphase Chemistry Department, Max Planck Institute for Chemistry, Mainz, Germany 2 Institute for Environmental and Climate Research, Jinan University, Guangzhou, China 3 Laboratory of Environmental Chemistry, Paul Scherrer Institute, 5232 Villigen, Switzerland 4 College of Environmental Sciences and Engineering, Peking University, Beijing, China
Motivation
Coated-wall flow tube reactors have been extensively employed for investigations of uptake and reaction kinetics of gases with reactive solid/semi-solid surfaces (Kolb et al., 2010). To simulate various heterogeneous or multiphase reactions relevant to atmospheric chemistry, these coated reactive surfaces can span a broad scale including inorganic salts (Davies and Cox, 1998;Chu et al., 2002;Qiu et al., 2011), organic acids and sugars (Shiraiwa et al., 2012;Steimer et al., 2015), proteins (Shiraiwa et 5 al., 2011), soot (McCabe andAbbatt, 2009;Khalizov et al., 2010;Monge et al., 2010), mineral dust (El Zein and Bedjanian, 2012;Bedjanian et al., 2013), ice (Hynes et al., 2001;Hynes et al., 2002;Bartels-Rausch et al., 2005;Fernandez et al., 2005;McNeill et al., 2006;Petitjean et al., 2009;Symington et al., 2012) and soils (Stemmler et al., 2006;Wang et al., 2012;Donaldson et al., 2014a;Donaldson et al., 2014b;VandenBoer et al., 2015;Li et al., 2016). Due to uptake or chemical reactions of gases at coated tube walls, radial concentration gradients are established within the tube and radial diffusion can be 10 significant. It is therefore necessary to account for this gas-diffusion effect on gas-surface interactions. The most commonly utilized methods for evaluation and correction for gas-diffusion in flow tubes include the Brown method (Brown, 1978), CKD method (Murphy and Fahey, 1987) and a more recently developed simple KPS method (Knopf et al., 2015). All of these methods are derived based on the assumption that gas flow in flow tubes should be well-developed laminar to ensure that the flow velocity profile is parabolic and that the radial transport of gas reactants is solely caused by molecular diffusion.
15
It is well known that the flow conditions in a tube depend on the Reynolds number, Re (Eqn. 1), where ρ is density of the fluid passing through the tube, V avg is the average velocity of the fluid (i.e., the volumetric flow rate divided by the cross sectional area of the tube), d is diameter of the tube, µ and ν are dynamic viscosity and kinematic viscosity 20 of the fluid, respectively. A laminar flow can be expected when Re is less than ~ 2000 (Murphy and Fahey, 1987;Knopf et al., 2015). Here, the expression of Re quantifies the nature of the fluid itself (i.e., ρ, V avg , µ and ν) and the tube geometry (i.e., d), but it does not account for the effects of surface roughness.
Surface roughness effects on flow conditions were firstly discussed by Nikuradse (1950). Based on his work, the Moody diagram 25 has been extensively used in industry to predict the effects of surface roughness (roughness height ε or relative roughness ε/d) on flow characteristics (in terms of friction factor). According to the Moody chart, when the surface roughness is small enough (i.e., ε/d ≤ 5%), the roughness effects within low Reynolds number regime (Re < 2000, characteristic of laminar flow) is negligible.
Recent experimental and theoretical studies, however, have found significant effects of surface roughness on laminar flow characteristics (e.g., fraction factor, pressure drop, critical Reynolds number and heat transfer etc.) in micro-channels and pipes 30 even under conditions of ε/d ≤ 5% (Herwig et al., 2008;Gloss and Herwig, 2010;Zhang et al., 2010;Zhou and Yao, 2011). This is because not only the ratio of ε and d but also other factors, such as shape of roughness elements (Herwig et al., 2008;Zhang et al., 2010) and spacing between different roughness elements , may determine the influence of surface roughness on the flow conditions.
5
Although the surface roughness effects can be potentially important, there has been a long-lasting debate on whether the coating surface roughness could disturb the fully developed laminar flow in flow tube kinetic experiments (Taylor et al., 2006;Herwig et al., 2008) and its effects were usually not well-quantified in most of the previous gas uptake or/and kinetic studies (Davies and Cox, 1998;Chu et al., 2002;McCabe and Abbatt, 2009;Khalizov et al., 2010;El Zein and Bedjanian, 2012;Shiraiwa et al., 10 2012;Wang et al., 2012;Bedjanian et al., 2013;Donaldson et al., 2014a;Donaldson et al., 2014b;VandenBoer et al., 2015). It is, however, conceivable that as the roughness of the coating surfaces increases it would eventually distort the steady laminar regime near tube walls and small-scale eddies would evolve from roughness elements giving rise to local turbulence, and hence corrupt the application of Brown/CKD/KPS methods for the derivation of uptake coefficient. The extent of these effects may depend on the coated film thickness and its surface roughness. It means that the roughness effects on flow conditions to a great 15 extent rely on the various coating techniques applied by different operators, leading to disagreement of the experimental results.
In the present study, the surface roughness effects on laminar flow are quantitatively examined. In view of the special laminar
30
According to the proverbial boundary layer theory proposed by Prandtl (1904), when a fluid (normally a gas mixture, a gas reactant mixed with a carrier gas, in uptake kinetic studies) enters the inlet of a flow tube with a uniform velocity, a laminar boundary layer (i.e., velocity boundary layer) will form very close to the tube wall (Fig. 2). This buildup of laminar boundary layer is because of the non-slip condition of the tube wall and the viscosity of the fluid, that is, viscous shearing forces between fluid layers are felt and dominant within the laminar boundary layer (Mauri, 2015). The thickness of laminar boundary layer δ 35 will continuously increase in the flow direction (axial direction in Fig. 2) until at a distance (from the tube entrance) where the boundary layers merge. Beyond this distance the tube flow is entirely viscous, and the axial velocity adjusts slightly further until no changes of velocity along the axial direction. Then, a fully developed velocity profile is formed and this velocity profile is Atmos. Chem. Phys. Discuss., doi: 10.5194/acp-2017-232, 2017 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 28 April 2017 c Author(s) 2017. CC-BY 3.0 License.
parabolic, which is characteristic of well-developed laminar flow (Mohanty and Asthana, 1979;White, 1998). The development and formation of this velocity profile is illustrated in Fig. 2. Normally, for coated-wall flow tube experiments a pre-tube is employed to function as developing a well-developed laminar flow before it entering into the coated tube section.
As demonstrated in previous studies using micro-channels and pipes (Herwig et al., 2008;Gloss and Herwig, 2010;Zhang et al., 5 2010;Zhou and Yao, 2011), the roughness elements on flow tube coatings can have non-ignorable effects on laminar flow conditions even if these coatings are entirely submerged into the laminar boundary layer. In other words, the disturbance on welldeveloped laminar flow patterns can be achieved artificially by roughness elements of the tube coating. However, there is a critical height δ c within which the roughness effects can become ignorable (Achdou et al., 1998).
10 Figure 3 shows a schematic of the structure of the δ c and its related flow conditions in a coated-wall flow tube. When a roughness height ε (here in Fig.3, the roughness height ε equates to the coating thickness ε max , see Sect. 3.1 for explanation) is larger than the critical height δ c , local eddies may occur in the spaces between the neighboring roughness elements (i.e., case 1 in Fig. 3A). Local turbulence induced by these roughness elements will enhance local transport of air masses within the scales of the roughness heights, which invalidates the assumption of solely molecular diffusion of gaseous reactants and the application of 15 diffusion correction methods for the determination of γ (Brown, 1978;Murphy and Fahey, 1987;Knopf et al., 2015).
Nevertheless, when a roughness height comes into the critical height δ c where viscous effects overwhelmingly dominate, the flow very near the rough wall will tend to be Stokes-like or creeping, shown as Case 2 in Fig. 3B. This Stokes-like flow adjacent to the rough surfaces can eventually avoid local turbulence between the roughness elements and guarantee perfect laminar flow regime (i.e., only molecular diffusional transport of gas reactants to rough reactive coatings at the flow tube wall) formation 20 throughout the whole flow tube volume. Thus Case 2 satisfies the prerequisite for the diffusion correction methods used for flow tube experiments, i.e., ε/δ c < 1. In the next section, we will show how to derive δ c . Achdou et al., (1998) proposed effective boundary conditions for a laminar flow over a rough wall with periodic roughness elements, and observed that when ε/L c < Re -1/2 (ε: roughness height; L c : characteristic length, for a tube the characteristic length 25 L c = d) the roughness elements could be contained in the boundary layer. This means that, for their case, the boundary layer thickness is in the order of L c Re -1/2 . Within the boundary layer, they found that local turbulence could occur between the roughness elements until ε/L c < Re -3/4 , where the viscous effects became dominated in roughness elements and then the flow near the rough wall tended to be creeping. This result coincides with Kolmogorov's theory (Kolmogorov, 1991), in which the critical length ratios between small scale and large scale eddies is also in the order of Re -3/4 , even though this theory only applies to 30 turbulent flow with large Reynolds numbers. Here, we adopt this criterion to judge if local eddies could occur in the spaces between neighboring roughness elements. Thus, the critical height δ c can be expressed as:
δ c derivation
where d is the diameter of the flow tube, Re is the Reynolds number, V avg and v are the average velocity and the kinematic viscosity of the fluid, respectively.
With Eqn.
(2), for a specified experiment configuration (i.e., flow tube diameter, flow velocity and fluid properties etc.) the critical height δ c can be determined, and therefore the effects of coating roughness on laminar flow can be estimated provided the roughness height ε is known.
Error estimation with modified CKD method
The potential effects of coating roughness on laminar flow are described and classified into two cases in Fig. 3 can be applied to obtain accurate γ from flow tube experiments. Case 1, however, can be quantitatively simulated because local turbulence is constrained into the scale of the roughness height ε (see Fig. 3A) and the turbulence effects could be quantified by assuming a proper turbulent diffusion coefficient D t within ε.
10
Hence, for Case 1, in order to estimate the potential error of the effective uptake coefficient (γ eff ) derived from moleculardiffusion-correction using the conventional methods aforementioned (here we adopt the CKD-B method proposed in our previous study Li et al., 2016), we further developed a modified CKD method (M-CKD, illustrated in Fig. 4) to account for local turbulence and therefore derive the real uptake coefficient (γ). In the M-CKD method, the molecular diffusion coefficient D of the gas reactant of interest is used in the main free-stream region (above the rough coating thickness ε max ) while a turbulent 15 diffusion coefficient D t is used in the roughness region to account for local turbulence between coating roughness elements (Fig. 4). The assumption of a whole turbulence layer (no laminar layer) in the roughness region represent an upper limit for the influence of turbulence, which corresponds to a largest uncertainty introduced to the calculation of γ eff . More details about derivation of γ eff and γ by CKD-B and M-CKD can be found in Appendix A.1 and A.2.
20
The turbulent diffusion coefficient D t can be approximately estimated by the following equation (Taylor, 1922;Roberts and Webster, 2002): where V is the flow velocity at the top edge of ε max (blue dashed line in Fig. 4). V is calculated according to the parabolic velocity profile (in tube radial direction) of the main laminar flow in the flow tube. Here the rough coating thickness ε max reflects the 25 largest scale to which a local eddy can develop, because a roughness height ε has the range of 0 ≤ ε ≤ ε max . Half of ε max is used as a characteristic diffusion distance in the turbulence-occurred region (Fig. 4).
Design of coated-wall flow tube experiments
The introduction of the critical height δ c , into the field of gas uptake or reaction kinetic studies using coated-wall flow tubes, and atomic force microscope etc.) are available for surface roughness examination (Poon and Bhushan, 1995). To simplify the discussion, here, we take the thickness of the coating film ε max as a maximum of its surface roughness (sometimes this case can happen), and use the comparison between ε max and δ c as a reference for the design of flow tube coating thickness. Such treatment is more suitable for practical applications because determination of coating film thicknesses can be simply achieved either by calculating the coating film volume (coated mass divided by density) or by means of scanning electron microscope technique, 5 and the condition of ε max /δ c < 1 can definitely ensure the case of ε/δ c < 1.
Larger δ c would allow a wider range of coating thickness ε max without surface roughness effects. Based on Eqn.
(2), larger δ c can be achieved either by increasing the tube diameter d or by decreasing the fluid average velocity V avg . In most cases, constant residence time of gaseous reactants inside flow tubes is needed to allow for enough uptake/reactions within the coated flow tube 10 volume. This requirement can also be fulfilled by adjusting the coated tube length L, that is, to achieve larger δ c the influence of decreasing V avg on residence time can be balanced by reducing L. Fig. 5C). These coatings may have a potential influence on laminar flow pattern and local turbulence may occur within the roughness-constructed spaces (see Fig. 4).
30
For most cases of flow tube experiments design, a coating layer cannot be thin enough due to requirements of reaction kinetics (bulk diffusion and surface reactions can both play important roles) and the thickness of a coating layer had been found to have an influence on gases uptake until a critical threshold was reached (Donaldson et al., 2014a;Li et al., 2016). This means that there is a need to comprehensively consider all the parameters (e.g., coating thickness, tube diameter, tube length, flow velocity etc.) and a compromise of each parameter for the others is necessary to finally ensure both the unaffected laminar flow conditions and 35 the application requirements for diffusion correction methods. For example, to ensure ε max /δ c < 1 for a thick coating (large ε max ), we can increase δ c by increasing the tube diameter or decreasing the flow velocity as shown in Fig. 5.
The conditions of ε max < δ c (constraining a coating thickness within the critical height of δ c , Case 2 in Fig. 3) design, however, some exceptional circumstances can still be foreseen (as in Case 1 in Fig. 3). For example, one may encounter the conditions in Case 1 (Fig. 3A) due to the limit of coating techniques or other specific considerations. Then it is critical to make a priori evaluation of potential error of γ due to coating roughness effects in the design of flow tube experiments. Figure 6 shows the deviation of calculated uptake coefficient γ eff against the real uptake coefficient γ for Case 1. There, three 5 different cases of ε max /R 0 are presented with all the rest experimental configurations being kept the same (see Table A.1). For higher ε max /R 0, the deviation of γ eff is also larger, indicating that a thick coating will result in larger error of the calculated γ eff .
Meanwhile, this error is also closely related to the magnitude of γ: at γ < 10 -5 there is almost no difference between γ eff and γ, but at γ beyond 10 -5 the error is apparent and considerably increases. In the case where local turbulence cannot be avoided (Case 1), Fig. 6 can be used to estimate the error of calculated γ eff . With this error range in mind, the selection of coating techniques or/and 10 parametrization of other experimental conditions can be better constrained, for example, if γ can be assumed to be smaller than 10 -5 the coating roughness effects become negligible.
Wall-roughness-induced error of γ eff in Case 1: for previous flow tube studies
Local turbulence caused by rough coatings had not been well-quantified in previous studies. Nevertheless, it may indeed happen and therefore introduce errors in the calculated uptake coefficient derived from the Brown/CKD/KPS methods (i.e., effective 15 uptake coefficient γ eff ). It is therefore meaningful too, for previous flow tube designers, to have an estimation of the potential error of the measured γ eff if their experiment conditions match Case 1. We show here an example illuminating how this estimation can be accomplished, by means of simulation under the pre-defined experimental configurations identical with those adopted in the above section. 20 Figure 7 shows the maximum relative errors that can be expected for a series of γ eff under the three different cases of ε max /R 0 shown in Fig. 6. Similar to the dependence of γ eff /γ on γ (Fig. 6), the increase of γ eff and ε max /R 0 are also accompanied by the increase of γ eff /γ in Fig. 7. The experiment example with coating thickness matching Case 1 (soil coating in Fig. 5B) is examined and the estimated maximum error (i.e., γ eff /γ, γ eff is 5.5 × 10 -5 , an average value from the measurements by Li et al., 2016) is ~ 1.5 (red solid circle), implying that the coating roughness has a small effect on laminar flow in their case. But for larger γ eff , γ eff /γ 25 versus a wide range of γ eff (red dotted line in Fig. 7) also gives high values, for example, when γ eff > 10 -3 . This simulation further highlights the need to inspect the possible errors of previously measured high γ eff (e.g., > 10 -3 ) if their coating roughness is accord with Case 1.
Conclusions
In this study, a new criterion is proposed to eliminate/minimize the potential effects of coating surface roughness on laminar flow maximum relative error of the effective uptake coefficient γ eff . The error estimation demonstrates that a smaller positive bias of γ eff can be expected for experimental configurations employing gas reactants with lower uptake efficiency or/and smaller relative ratio of ε max /R 0 .
Data availability
The underlying research data and Matlab code for the M-CKD and CKD-B methods can be accessed upon contact with Yafang 5 Cheng (yafang.cheng@mpic.de), Hang Su<EMAIL_ADDRESS>or Guo Li (guo.li@mpic.de).
A.1 Evaluation of the modified CKD method
To have an intuitive feeling of the change of concentration profiles due to local-turbulence-induced enhancement of the uptake coefficient, the CKD-B and the M-CKD have been applied to the experiment configuration of the exampled Case 1 in Fig. 7 (see 10 HCHO in Table A.1), as shown in Fig. A.2. Comparison between Fig. A.2 (A) and (B) shows that local turbulence within the roughness thickness can enhance radial transport of the gas reactant and thus increase the effective uptake coefficient γ eff .
A.2 Derivation procedure of γ eff /γ versus γ or γ eff
Derivation of γ eff /γ versus γ or γ eff is based on a combination of a modified CKD method (M-CKD), assuming that roughnessinduced local turbulence occurs within the domain of 0.5ε max (simulation of Case 1), and the CKD-B method (a CKD-based 15 method using Matlab) which was described in our previous study (Li et al., 2016). The respective derivation procedures are shown in Fig. A.3. For one specific experiment configuration, both CKD-B and M-CKD can generate a correlation table with its first column being concentration transmittance (C/C 0 ) and the second column the corresponding uptake coefficient (γ), and their one-to-one correspondence is indicated by the same subscripts (e.g., k, n, j, m etc.), as shown in Fig
30
Due to the different algorithms employed, the CKD method (Murphy and Fahey, 1987) and the KPS method (Knopf et al., 2015) could derive contrasting γ when turbulence occurs (see Fig. A.4). As shown in Fig. A.4, with ideal laminar flow (without any local turbulence, Case 2) the KPS (with diffusion correction) and CKD show perfect agreement for the derived γ in the C/C 0 range of 0.548 to 1 (shaded area). If C/C 0 is smaller than the critical transmittance value (< 0.548), e.g., because of enhanced Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2017-232, 2017 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 28 April 2017 c Author(s) 2017. CC-BY 3.0 License. mass transport towards the coated-wall due to local turbulence in laminar flow, the KPS results in negative γ (for details, see Knopf et al., 2015) while the CKD has no solutions. For C/C 0 larger than 1, both methods derive negative γ implying emissions of gas reactants from the coating.
Acknowledgments
This study was supported by the Max Planck Society (MPG) and National Natural Science Foundation of China (41330635). (Wang et al., 2012); circle (Li et al., 2016) and star in (B) (Steimer et al., 2015); blue triangle (McNeill et al., 2006) and light red triangle (Petitjean et al., 2009) in (C). 10 15 Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2017-232, 2017 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 28 April 2017 c Author(s) 2017. CC-BY 3.0 License. Figure 6. Maximum error of the effective uptake coefficient (γ eff ) relative to the real uptake coefficient (γ) versus γ, for three cases with different ratio of the coating thickness to tube radius (ε max /R 0 ). The choices of ε max /R 0 cover the general ratio range in previous studies. The curves cannot be further extended due to reaching the limits of diffusion correction methods (see Appendix A.3). | 5,245.8 | 2017-04-28T00:00:00.000 | [
"Engineering"
] |
Time-efficient combined morphologic and quantitative joint MRI: an in situ study of standardized knee cartilage defects in human cadaveric specimens
Background Quantitative techniques such as T2 and T1ρ mapping allow evaluating the cartilage and meniscus. We evaluated multi-interleaved X-prepared turbo-spin echo with intuitive relaxometry (MIXTURE) sequences with turbo spin-echo (TSE) contrast and additional parameter maps versus reference TSE sequences in an in situ model of human cartilage defects. Methods Standardized cartilage defects of 8, 5, and 3 mm in diameter were created in the lateral femora of ten human cadaveric knee specimens (81 ± 10 years old; nine males, one female). MIXTURE sequences providing proton density-weighted fat-saturated images and T2 maps or T1-weighted images and T1ρ maps as well as the corresponding two- and three-dimensional TSE reference sequences were acquired before and after defect creation (3-T scanner; knee coil). Defect delineability, bone texture, and cartilage relaxation times were quantified. Appropriate parametric or non-parametric tests were used. Results Overall, defect delineability and texture features were not significantly different between the MIXTURE and reference sequences (p ≤ 0.47). After defect creation, relaxation times significantly increased in the central femur (T2pre = 51 ± 4 ms [mean ± standard deviation] versus T2post = 56 ± 4 ms; p = 0.002) and all regions combined (T1ρpre = 40 ± 4 ms versus T1ρpost = 43 ± 4 ms; p = 0.004). Conclusions MIXTURE permitted time-efficient simultaneous morphologic and quantitative joint assessment based on clinical image contrasts. While providing T2 or T1ρ maps in clinically feasible scan time, morphologic image features, i.e., cartilage defects and bone texture, were comparable between MIXTURE and reference sequences. Relevance statement Equally time-efficient and versatile, the MIXTURE sequence platform combines morphologic imaging using familiar contrasts, excellent image correspondence versus corresponding reference sequences and quantitative mapping information, thereby increasing the diagnostic value beyond mere morphology. Key points • Combined morphologic and quantitative MIXTURE sequences are based on three-dimensional TSE contrasts. • MIXTURE sequences were studied in an in situ human cartilage defect model. • Morphologic image features, i.e., defect delineabilty and bone texture, were investigated. • Morphologic image features were similar between MIXTURE and reference sequences. • MIXTURE allowed time-efficient simultaneous morphologic and quantitative knee joint assessment. Graphical Abstract Supplementary Information The online version contains supplementary material available at 10.1186/s41747-024-00462-0.
Background
Osteoarthritis is a chronic joint disease with increasing prevalence due to aging and obesity [1].Magnetic resonance imaging (MRI) is clinically well-suited for diagnosing cartilage degeneration as the hallmark change of osteoarthritis.Traditional MRI sequences such as proton density (PD)-weighted fat-saturated (FS) sequences focus on cartilage morphology, i.e., surface integrity and intratissue signal [2].They are insensitive to early degenerative changes of tissue morphology, such as partialthickness defects and fibrillation [3,4].
Quantitative MRI techniques such as T2 or T1ρ mapping may be beneficial in detecting such changes at a potentially reversible stage [5,6].Consensus prevails that adding T2 maps to the routine knee protocol improved sensitivity in detecting cartilage lesions significantly [3,6].The literature is less clear on T1ρ: even though the association between cartilage degeneration and prolongation of T1ρ relaxation is well established [7,8], its potential diagnostic benefits remain to be ascertained.Both mapping techniques have in common that their more widespread adoption is hampered by prohibitively long scan times and other challenges [9,10].
Combined morphologic and quantitative sequences, such as quantitative double-echo in steady-state, provide morphologic images and T2 maps in clinically feasible scan times [11][12][13].Quantitative double-echo in steady-state sequences is diagnostically equivalent to conventional clinical MRI protocols; therefore, they are theorized to (partially) substitute the routine knee protocol while providing additional T2 maps [14].Nevertheless, the morphologic images lack the clinically familiar contrasts of state-of-the-art turbo spin-echo (TSE) sequences.The "multi-interleaved x-prepared turbo spinecho with intuitive relaxometry" (MIXTURE) sequence provides an alternative platform for combined imaging [15].These sequences are designed to acquire at least two morphologic images of different contrast weightings using variable prepulses.Because prepulses, echo times, and spin-lock durations are freely adjustable, the sequences provide quantitative T2 or T1ρ maps as a "byproduct" of the morphologic images.In contrast to quantitative double-echo in steady-state, MIXTURE is based on a three-dimensional (3D) TSE acquisition.Preliminary studies have explored its principal clinical applicability, but systematic comparisons with reference sequences and standardized pathologies are lacking [16][17][18][19][20].
This study aimed to evaluate MIXTURE sequences in two principal configurations for their clinical usage and against corresponding two-dimensional (2D) and 3D TSE reference sequences in an in situ model of standardized cartilage defects.We hypothesized that the morphologic MIXTURE images are diagnostically on par with their reference sequence counterparts while additionally providing T2 and T1ρ maps in clinically feasible time frames.
Study design
The local Institutional Review Board approved this prospective in situ imaging study on human cadaveric knee joint specimens (Ethical Committee, RWTH Aachen University, EK180/16) conducted in 2022 and 2023.Fresh-frozen and nonfixated knee joint specimens from body donors who had given written informed consent prior to study initiation were provided by the local Institute of Anatomy (RWTH Aachen University, Germany).Moderate-to-severe cartilage degeneration of the lateral compartment, such as substantial tissue loss or focal lesions, was screened for during standard clinical scanning (using 2D TSE PD-weighted FS imaging) and defined as an exclusion criterion.Based on a preliminary analysis of the first three specimens, a minimum sample size of 8 was calculated using a statistical power of 80%, a significance level of 0.01, and an effect size (i.e., Cohen's d [21]) of 1.24.Hence, ten knee joint specimens were included.
Workflow
Specimens were left to thaw at room temperature for 24 h.MRI was performed before and after creating standardized cartilage defects.On day 1, predefect MRI was performed.The specimens were kept at 5 °C overnight.On day 2, cartilage defects were created, and postdefect MRI was performed immediately afterward.
Cartilage defects
We present in Fig. 1 the standardized step-wise creation of the cartilage defects.N.P. (pregraduation medical student, 2 years of experience) created the defects.First, the knee joint was accessed through a median longitudinal skin incision and a medial peri-patellar approach.Once the joint was flexed, the patella was everted laterally to fully expose the joint.Second, the weight-bearing region of the lateral femoral condyle was identified.Three defects of 3, 5, and 8 mm in diameter were created in the lateral femoral condyle perpendicular to the condyle's bone contour.The cartilage tissue was removed using skin biopsy punches of corresponding diameters and surgical scalpels.Particular care was taken to maintain the integrity of the subchondral lamella.Third, the joint was thoroughly and continuously irrigated with 0.9% saline solution to remove surgical debris and excess air.Fourth, the joint was sutured layer-wise.
MRI acquisition
All scans were performed on a 3-T MRI scanner (Elition X, Philips, Best, The Netherlands) using an eightchannel transmit-receive knee coil.Specimens were positioned feet-first, supine, and in approximately 30° of flexion in line with clinical positioning.Two MIXTURE sequences were acquired: PD-weighted FS images with T2 maps (scan time 4:59 min:s) and T1-weighted images with T1ρ maps (scan time 6:30 min:s), followed by the respective reference sequences (2D TSE and 3D TSE).The first MIXTURE sequence was a combination of PD-weighted FS (using a spectral attenuated inversion-recovery [SPAIR] prepulse) and T2-weighted morphologic (using a T2-preparation module of 50 ms) acquisitions, in an interleaved manner.Voxel-wise T2 relaxation times were determined based on the two images by monoexponential fitting, and T2 maps were subsequently reconstructed on the scanner workstation using prototype software.The second MIXTURE sequence was a combination of a T1-weighted (without preparation) and two spin lockprepared T1ρ-weighted FS morphologic acquisitions (using T1ρ-preparation modules of 25 ms and 50 ms, respectively, and SPAIR prepulses).Voxel-wise, T1ρ relaxation times were determined based on the three images by mono-exponential fitting, and T1ρ maps were reconstructed accordingly.Further details on the MIXTURE sequences can be found in the literature [15].2D TSE reference PD-weighted FS and T1-weighted sequences were included per our clinical knee protocol, while 3D TSE reference PD-weighted FS and T1-weighted sequences were obtained from the vendor and included.Table 1 summarizes the sequence parameters.
Notably, as 3D TSE acquisitions, the reference 3D TSE and MIXTURE sequences can, in principle, be acquired at isotropic resolution.In this study, however, we aimed to match the 3D TSE sequences to the 2D reference TSE sequences, i.e., the clinical reference standard, for voxel-to-voxel comparisons.Consequently, the
Table 1 MRI sequence parameters
Two MIXTURE sequences were acquired, combining morphologic imaging with quantitative mapping, and 2D TSE and 3D TSE reference sequences of the same weighting.PD-weighted FS images were combined with quantitative T2 maps ("MIX 1") and T1-weighted images with T1ρ maps ("MIX 2").Note that for 3D TSE sequences, TE eff and TE equiv denote the effective and equivalent TE, respectively, as mediated by the choice of the refocusing pattern.In contrast, the 2D TSE sequence uses a constant refocusing flip angle that a single TE can describe.During the TSE readout, different refocusing patterns with variable order and magnitude of the flip angles are employed as designated by the manufacturer.T2-prep TE and SL-prep TSL refer to the duration of the preparation modules that MIXTURE employs to generate the respective contrast weightings 3D TSE sequences were acquired analogously to the 2D TSE sequence, i.e., using thicker slices and higher inplane resolution than achievable with isotropic image acquisitions.
Image analysis
Quantitative analyses were performed in Python version 3.9.9[22].
Defect delineability
Cartilage defect delineability was assessed on the PDweighted FS sequences using line profiles manually annotated in ITK-SNAP version 3.8 [23,24] by N.P.
(pregraduation medical student, 2 years of experience in medical imaging) and visually verified by S.N. (boardcertified musculoskeletal radiologist, 10 years of experience).Line profiles were placed through the defect and adjacent cartilage on the sagittal postdefect PD-weighted FS image that centrally bisected the defects (Fig. 2a).As projections of the signal intensity (SI) along their course, SI line profiles were extracted from the 2D TSE, 3D TSE, and MIXTURE PD-weighted FS images and normalized to the maximum of 1 (Fig. 2b).For every SI line profile, full width at half maximum (FWHM, Fig. 2c) and edge width (EW, Fig. 2d) were evaluated as surrogates of defect delineability.More specifically, a parallel line was defined at half maximum between the background signal level of cartilage and the maximum SI along the line profile.The horizontal distance between the intersections of this line with the SI line profile was determined as the defect's FWHM.Similarly, two vertical lines per defect shoulder defined the 10% and 90% maximum SI.The horizontal distance between these two lines was determined as the respective defect shoulder width, and EW was calculated as the mean of both defect shoulder widths.
Bone texture features
The bone texture on the T1-weighted sequences was quantified using radiomic features (Fig. 3).In ITK-SNAP, circular regions of interest (ROI) with a diameter of 40 pixels were defined directly adjacent to the 5-mm defect (Fig. 3a) on the same sagittal postdefect slice as above.
Before computing texture features, the stacks of the 2D TSE, 3D TSE, and MIXTURE T1-weighted images were normalized between the SI values 0 and 1 (Fig. 3b).
Guided by earlier studies [25], we focused on variance, (joint) energy, (joint) entropy, and inverse difference (synonymous with "homogeneity1" Quantitative parameter maps N.P. segmented the femoral and tibial cartilage plates on the MIXTURE PD-weighted FS images using ITK-SNAP.The central bisecting slice through the defects (postdefect) and the corresponding original slice (predefect) were segmented.The femoral cartilage was divided into an anterior ("aF"), central ("cF"), and posterior region ("pF") based on the outer contours of the lateral meniscus' anterior and posterior horns.The tibial cartilage ("T") was segmented as one region.All segmentation outlines were reviewed and adjusted by T.N. and S.N.T2 and T1ρ values were computed (predefect and postdefect) and provided as mean ± standard deviation for each region and the entire lateral femorotibial compartment.To investigate a potential interplay of cartilage morphology and relaxation times, the digital caliper of the in-house picture archiving and communication system was used to determine cartilage thickness adjacent to the 8-mm defects (postdefect images; PD-weighted FS and T1-weighted images side by side) and at the corresponding location (predefect images).The caliper's step size was one pixel, and the caliper resolution was limited by the image resolution of 0.27 × 0.27 mm.
Statistical analysis
Data are given as mean ± standard deviation unless differently specified.Statistical analysis was performed by N.P., T.N., and S.N. using Graph Pad Prism (v9.5.1, San Diego, CA, USA).Intersequence comparisons of FWHM, EW, and radiomic texture features were performed using repeated measures analysis of variance (ANOVA) Fig. 3 Inter-sequence comparison of bone texture.On the T1-weighted images, radiomic features were analyzed by defining a standardized circular region of interest adjacent to the subchondral lamella underneath the 5-mm defect (a).Image preprocessing included normalization of the signal intensity to the range of 0 to 1 (b).As texture features, variance as a first-order feature and joint energy, joint entropy, and homogeneity (i.e., ID) as gray-level co-occurrence matrix (GLCM) features were extracted (c).For the computation of the GLCM, the normalized images were quantized into 200 evenly spaced bins.GLCM Gray-level co-occurrence matrix, ID Inverse difference followed by the Tukey-Kramer post hoc test.Predefect and postdefect T2 and T1ρ relaxation times were comparatively evaluated per region and overall using Wilcoxon matched-pairs signed-rank tests.Predefect and postdefect cartilage thickness was compared by means of a two-tailed paired t-test.To reduce the number of statistically significant but clinically likely irrelevant findings, the family-wise significance level was set to ∝ = 0.01.Multiplicity-adjusted p-values are provided.
Standardized cartilage defects were successfully created in all specimens.
Qualitative evaluation
In PD-weighted FS images, the cartilage defects were clearly discernable, and the cartilage tissue had the characteristic layer-wise configuration and intermediate SI in all sequences.Menisci and bone marrow appeared homogeneously dark, i.e., suppressed, while intraarticular fluid was homogenously bright (Fig. 4).In the T1-weighted images, the macro-and microstructural bone texture appeared slightly less blurry in the 2D TSE sequence, particularly compared to the MIXTURE image (Fig. 5).Contrast and noise levels appeared largely similar.Regarding artifacts, we noted artificial signal hyperintensities in the
Quantitative evaluation of defect delineability
Table 2 presents the metrics of defect delineability, i.e., FWHM and EW values.FWHM values were substantially lower than the nominal defect diameters but overall largely similar between the sequences.For the 5-mm defects, however, the 2D TSE sequence yielded significantly higher FWHM values than the MIXTURE sequence (p = 0.005).On average, EW values were lower for the 2D TSE than for the 3D TSE and MIXTURE sequences, even though statistically not significant.The latter two sequences exhibited largely similar EW values.
Quantitative evaluation of bone texture
Voxel SIs contained in the ROI were spread out along 43 ± 10 (2D TSE), 49 ± 9 (3D TSE), and 42 ± 9 (MIXTURE) bins, indicating a comparable spread of voxel SI distributions.The radiomic feature analysis indicated comparable bone texture feature values between the sequences (Fig. 6).When comparatively evaluating the individual features, significant differences were only found between the 3D TSE and MIXTURE sequences with significantly higher energy and homogeneity values (and significantly lower entropy values) determined for MIXTURE versus 3D TSE.
Quantitative parameter maps
We observed increased T2 and T1ρ relaxation times after defect creation for all studied regions (Table 3).These increases were mainly non-significant except for T2 in the central femur, where the defects were located
Discussion
Our study evaluated the image quality of MIXTURE PD-weighted FS and T1-weighted sequences relative to corresponding 2D and 3D TSE reference sequences.Focusing on the delineability of cartilage defects and quantitative bone texture features, we found that MIX-TURE sequences were largely equivalent regarding image contrast, morphologic correspondence and coherence, and quantitative features.Simultaneously, MIXTURE sequences provided quantitative T2 or T1ρ maps with little additional scan time.Thereby, MIXTURE sequences could increase the diagnostic information of routine scan protocols beyond mere morphology and may complement (or in parts even replace) current knee MRI protocols.The primary advantage of MIXTURE sequences is their TSE-derived image contrast.Since their introduction to the clinic in the early 1990s, TSE sequences have been considered the standard for knee MRI; thus, radiologists are used to these images, and the American College of Radiology even formally recommends their usage [28].MIXTURE sequences obviate the need for radiologists to familiarize themselves with other contrasts.Additionally, the sequence architecture is flexible and may be adjusted to other TSE-based weightings with or without fat saturation.A broad spectrum of sequence combinations can thus be efficiently acquired at each institution's discretion.
Specifically, we evaluated a PD-weighted FS sequence with T2 maps and a T1-weighted sequence with T1ρ maps acquired with 43 slices across the joint in 5 and 6.5 min, respectively.Previously, Kijowski et al. [3] Table 2 Quantification of cartilage defect delineability The full widths at half maximum (FWHM) and edge widths (EW) were extracted from the line profiles of the PD-w FS images and used as surrogate parameters of defect delineability.EW was averaged over both defect shoulders.Data are presented as mean ± standard deviation [mm].The statistical analysis was performed using repeated measures ANOVA.p-values are given as a function of sequence, delineability parameter (i.e., FWHM and EW), and nominal defect diameter (i.e., 3 mm, 5 mm, and 8 mm).highlighted the clinical potential of adding T2 maps to the routine protocol.Even though the diagnostic benefit of T1ρ maps remains unclear, adding more quantitative images to the morphologic standard images seems welljustified.MIXTURE needs at least two morphologic images, which require more acquisition time than a single image.
The MIXTURE PD-weighted FS sequence depicted the cartilage defects with a level of contrast and sharpness similar to the reference sequences.By trend, EW and FWHM values of the 2D TSE sequences were lower and closer to the nominal defect diameters, respectively, than those of the corresponding MIXTURE and 3D TSE sequences.This finding indicates slightly less clear defect delineability of the latter sequences and may be due to the higher echo train lengths [29,30] or the choice of the refocusing pattern that, besides affecting image contrast and SNR, also influences image blurring [31].
Increased blurring, likely secondary to the choice of the refocusing pattern, was observed for the MIXTURE T1-weighted images and confirmed by the radiomic analysis of bone texture.Bone texture was significantly more homogeneous in the MIXTURE T1-weighted sequence, which may translate into a loss of microtextural detail with as yet unknown clinical relevance.
When designing the study, we aimed to compare cartilage and bone texture voxel-wise.To this end, we matched the image resolutions of all sequences, both in plane and through plane.Yet, this approach precluded the possibility of performing multiplanar reconstructions, an inherent feature of isotropic 3D sequences, which is a prerequisite for precise tissue segmentation (of cartilage and meniscus) for the analysis of morphometry and relaxivity [32,33].
Quantitative analyses indicated increased postdefect relaxation times.Alongside increased T2 and T1 ρ relaxation times, we observed cartilage thickness increases when comparing predefect and postdefect images.Even though the limited number of specimens and the caliper-related inaccuracies in thickness measurements need to be acknowledged, the changes in cartilage morphology and relaxivity are likely due to tissue swelling secondary to surgical handling, extended exposure to unphysiological conditions, and potentially altered tonicity.
While the exact compositional and structural correlates of prolonged T1ρ and T2 relaxation times remain unknown, literature evidence suggests that cartilage hydration is likely dominant [34,35].Surprisingly, we observed higher T2 than T1ρ relaxation times in the cartilage.In biological tissues, T1ρ relaxation times should be longer than T2 relaxation times because the spin-lock pulse forces the spins to precess about a direction different from the main magnetic field B 0 , thereby slowing T2 relaxation [36].Shorter repetition times (as present in the MIXTURE sequence) may have led to T1ρ underestimation [37]: if the repetition time is too short, it may not allow for complete T1ρ relaxation and decrease T1ρ relaxation times.Other factors worth considering are the applied radiofrequency pulse for the T1ρ preparation, the B 1 inhomogeneity, and the magic angle effect [38].Future phantom studies are needed to assess the accuracy and validity of MIXTURE-based relaxivity measurements versus reference measurements, e.g., multi-echo spin-echo sequences (for T2 quantification) and gradient-echo sequences (for T1ρ quantification) [39].
Our study has limitations.First, the in situ defect model using human cadaveric knee joints only approximates the actual in vivo situation.However, the model effectively excludes intersequence motion (and other artifacts such as arterial pulsations) and helps realize reproducible and standardized experimental conditions for voxel-wise comparisons.Regarding clinical translation, this model is inherently limited.Second, Table 3 Quantification of cartilage composition and ultrastructure T2 and T1ρ relaxation times (mean ± standard deviation [ms]) of the segmented cartilage of the central lateral femorotibial compartment before and after defect creation in ten knee joint specimens.The regional assessment included three femoral and one tibial region.Predefect and postdefect relaxation times were compared using the Wilcoxon matched-pairs signed rank test, and multiplicity-adjusted p-values were determined the number of specimens was small, and the study provided, by design, a focused proof of concept.Further diagnostic aspects relating to particular knee joint conditions require larger sample sizes and, ideally, assessment in the clinical routine.
In conclusion, combined morphologic and quantitative MRI sequences, such as the versatile MIXTURE platform, increase scanning efficiency and diagnostic utility by providing familiar contrasts and delivering additional quantitative information about joint cartilage defects in human cadaveric specimens.In a basic research context, MIXTURE sequences demonstrated excellent delineability of cartilage defects and visualization of bone texture on par with the corresponding reference sequences.Once corroborated by larger clinical studies, MIXTURE may be a promising sequence platform for comprehensive and time-efficient joint imaging.
Fig. 1
Fig. 1 Standardized cartilage defect model.a Intact knee joint.b Complete surgical exposure of the knee joint specimen through the longitudinal arthrotomy, medial peripatellar incision, and lateral eversion of the patella.c By use of biopsy punches, cartilage defects of variable diameters, i.e., 3 mm (top), 5 mm (center), and 8 mm (bottom), were aligned anteroposteriorly.d The wound was closed by layer-wise suturing under continuous irrigation
Fig. 2
Fig.2Inter-sequence comparison of cartilage defect delineability.a For this representative defect of 5 mm in diameter, a line was manually annotated to transect the defect and adjacent cartilage at mid-substance (red line, sagittal PD-weighted FS image).b For each sequence, i.e., 2D TSE, 3D TSE, and MIXTURE, the line profiles (corresponding to the pixel-wise signal intensity along the red line) were extracted, normalized to the maximum signal intensity of 1 (blue circles), and used to calculate the FWHM (c) and the EW (d) as surrogate measures of defect delineability.c FWHM was determined by determining the half maximum (dashed green line) between the cartilage background signal intensity (dashed red line) and the maximum signal intensity and by measuring the horizontal distance between the intersecting points of the half maximum with the signal intensity profile (green dots).d Analogously, EW was determined by defining the horizontal distances between the 10 and 90% maximum intensity levels (green dots, dashed green lines) on both defect shoulders.EW Edge width, FS Fat-saturated, FWHM Full width at half maximum, MIXTURE Multi-interleaved x-prepared turbo spin-echo with intuitive relaxometry, PD Proton density, TSE Turbo spin-echo [PyRadiomics]) to quantify the spatial distribution of SI values, characterize the underlying bone structure, and capture what the radiologist assesses on the microstructural level (Fig. 3c).The texture features were determined using PyRadiomics [26].Variance is a first-order feature that measures SI value spread within the ROI; high variance indicates high heterogeneity and large differences from their mean SI.Entropy, energy, and inverse difference are gray-level cooccurrence matrix features.The gray level co-occurrence matrix quantifies how often different neighboring voxel value pairs are present within the ROI.Entropy measures disorder or complexity; high entropy indicates bone tissue with a complex texture characterized by diversely varying neighboring SI values.Energy measures textural uniformity; high energy indicates many repetitions of the same neighboring SI values.Homogeneity measures local image uniformity; high homogeneity indicates more uniform gray levels.High entropy, energy, and inverse difference values indicate more randomness, homogeneous patterns, and local homogeneity [26, 27].
Fig. 4
Fig. 4 Representative PD-weighted fat-saturated images and MIXTURE T2 maps.Sagittal images before (orange frame (a)) and after (blue frame (b)) the creation of standardized cartilage defects.The slice that centrally bisected the three defects and the corresponding slice of the intact joint was selected.Cartilage defects of 3 mm, 5 mm, and 8 mm diameter (from left [anterior] to right [posterior]) are displayed.Zoomed images (indicated by the inset boxes in the leftmost images) are from left to right: the 2D TSE sequence, the 3D TSE sequence, and the MIXTURE sequence.Corresponding MIXTURE-based T2 maps before (c) and after (d) defect creation.The scale bar on the right extends from 0 to 100 ms.MIXTURE Multi-interleaved x-prepared turbo spin-echo with intuitive relaxometry, PD Proton density, TSE Turbo spin-echo
Fig. 5
Fig. 5 Representative T1-weighted images and MIXTURE T1ρ maps.Morphologic images (a, b) and corresponding MIXTURE-based T1ρ maps (c, d) are visualized before and after defect creation.Figure organization as in Fig. 4. MIXTURE Multi-interleaved x-prepared turbo spin-echo with intuitive relaxometry
Fig. 6
Fig.6 Analysis of bone texture features.Based on the radiomic feature analysis, a representative region of interest of the subchondral bone was defined and compared between the T1-weighted sequences, i.e., the 2D TSE, 3D TSE, and MIXTURE sequences.Variance (a), energy (b), entropy (c), and homogeneity (d) were quantified and analyzed as measures of bone texture.Levels of statistical significance were stratified as "ns, " "*, " "**, " and "***" to indicate p > 0.05, 0.01 < p ≤ 0.05, 0.001 < p ≤ 0.01, and p ≤ 0.001, respectively.MIXTURE Multi-interleaved x-prepared turbo spin-echo with intuitive relaxometry, TSE Turbo spin-echo Significant differences are indicated in bold typeANOVA Analysis of variance, EW Edge width, FS Fat-saturated, FWHM Full width at half maximum, MIXTURE Multi-interleaved x-prepared turbo spin-echo with intuitive relaxometry, PD Proton density, TSE Turbo spin-echo * The post hoc details (Tukey's test) regarding multiplicity-adjusted p-values for pairwise sequence comparisons were p = 0.030 for 2D TSE versus 3D TSE, p = 0.005 for 2D TSE versus MIXTURE, and p = 0.910 for 3D TSE versus MIXTURE Post After defect creation, Pre Before defect creation * Significant difference | 5,968.2 | 2024-06-05T00:00:00.000 | [
"Medicine",
"Engineering"
] |
The Dardenne Brothers and the Invisible Ethical Drama : Faith without Faith
The cinema of the Dardenne brothers represents a new kind of cinema, one that challenges a number of our conventional ways of thinking about the distinction between religion and secularism, belief and unbelief. Their films explore the intricacies of spiritual and ethical transformations as they are experienced within embodied, material life. These features of their cinema will be examined primarily through the lens of Emmanuel Levinas’s philosophy of the imbrication of the drama of existence and the ethical intrigue of self and Other. The work of the Dardenne brothers can be understood as an attempt to express what I describe as a “faith without faith”—a recognition of the absolute centrality of belief for the development of a responsible subject but in the absence of a traditional faith in a personal deity.
Introduction
The films of Jean-Pierre and Luc Dardenne represent a new kind of cinema that forcefully captures our current existential and spiritual state of affairs 1 .I contend that the power of their films derives from an awareness of the necessity for transcendence in a time when traditional, institutional religion appears incapable of addressing that very need for significant numbers of people in the western world.Their films can be conceived as attempting to perform a task formerly carried out by means of explicitly religious rituals and practices.
How we characterize what is "new" in their cinematic philosophy, however, is critical.One needs the right words to describe what is at play in their work.Some critics and commentators have been justified in recognizing the persistent presence of religious themes in their films-beyond the very secular surface that we first encounter on the screen.Attentive viewers will undoubtedly notice the many allusions to biblical sources in those films 2 .Without careful qualification, however, the term "religious" is simply misleading.We are fortunate to have at our disposal a number of illuminating interviews that one or both of the brothers have given over the years.In addition, we now have ample written material, in particular, from Luc Dardenne about the significance of their art 3 .His written reflections reveal a genuine preoccupation with religion, but from the standpoint of someone who grapples with the question of the death of God.Indeed, that theme is repeated so many times in his most recent work that a superficial reading of that text might lead one to think that we are dealing with a straightforward atheistic view of life.He himself recently admits of having no faith in a personal God.However, the simple epithet of "atheist" is as problematic as "religious" in this context.When asked about his own relation to faith, his response puts us on notice that this is no simple atheism: "that does not mean that I cannot speak about God, or that I cannot feel a relationship with such a Being that does not exist" 4 The religious impulse goes to the heart of what it means to be human, for him, and, presumably, his brother 5 -even if, at the end of the day, he, admittedly, lacks faith in the personal God that is attested to in the monotheistic traditions.The Dardenne brothers' films, I maintain, take religious experience seriously but beyond the traditional and conventional ways that we conceive of the dichotomies of religion and secularism, belief and unbelief.
In an effort to better understand what is at stake in their remarkable work, I will make special use of the philosophy of Emmanuel Levinas.Much has already been said about the obvious links between Levinas and the Dardenne brothers 6 .My own distinct approach frames that relationship in terms of Levinas's depiction of the imbrication of two dramas that are at play in human life.That framing, I hope, can shed further light on the brothers' highly original cinematic style-one whose phenomenological reduction (it is not a stretch to call it that) has the capacity to reveal what is most hidden in the drama of existence and the ethical relation with the Other.Finally, Levinas also provides us with a useful way of positioning their aesthetic production with respect to the question of faith.
Levinas and the Invisible Drama of Self and Other
One would be hard pressed to find a more felicitous coupling of film and philosophy then the relationship that conjoins the cinema of the Dardenne brothers and Levinas's thought.What makes this particular pairing between philosopher and artist so fruitful is the fact that the filmmakers in question are intimately familiar with Levinas's ideas.Luc Dardenne, who pursued graduate level work in philosophy, studied with Levinas when he was a visiting professor at the University of Louvain.The filmmaker explicitly acknowledges the brothers' indebtedness to Levinas's philosophy in various interviews and in his books.However, before discussing their work, I want to briefly sketch some of the basic attributes of what, for Levinas, constitutes the human drama and how art either neglects or distorts these features.
The human drama, for Levinas, is comprised of two separate but related dramas 7 .The first drama concerns our relationship to being or existence.Levinas refers to this relationship as the "drama of being" ( [10], p. 40).This drama is primarily characterized by the self's precarious place within being.For Levinas, the self is an ontological point or moment that temporarily arrests the flux or indeterminacy of being, what he calls the il y a or "there is."The il y a, for Levinas, represents being in its most diffuse state, existence shorn of all determinations or thingness.Its biblical corollary is the tohu wa bohu of Genesis 1:2; the formless and empty darkness out of which a world eventually emerges.Despite its accomplishment, the fact that it has staked out a place within the impersonal murmur of being, the self is nevertheless constantly threatened by the possibility of being undone by the diffusiveness of the il y a. Levinas describes the il y a as threat, but also as a site of fascination for the self.The fascination has in large part to do with some of the extreme states that the il y a can give rise to.When in the grip of the il y a, the boundaries of selfhood begin to dissolve giving rise to intense affective paroxysms.In 1947, Levinas observes that when the ego is under the spell of the il y a it is "submerged by the night, invaded, depersonalized, stifled by it" ( [11], p. 58).These 4 "Interview de Luc Dardenne" [5].An English translation of this interview will appear in Accursed Films: Postsecular Cinema between the Tree of Life and Melancholia.Edited by John Caruana and Mark Cauchi [6].
Despite the fact of frequently speaking with a common voice in many of their interviews, I cannot, of course, say with certainty that they share similar views on the matters discussed in my article.And, so, while I leave open the real possibility that Jean-Pierre may not always concur with his brother, Luc, on some of the details discussed here, I will, nevertheless, for the sake of my exposition, assume that their views coincide.6 See, in particular, Sarah Cooper [7] and Philip Mosley [8].
7
I develop this notion of the dramatic dimension of Levinas's philosophy in greater detail in "The Drama of Being: Levinas and the History of Philosophy." [9].liminal states create the impression of a kind of transcendence, relief from the burden of existence.For Levinas, however, this impression is nothing more than a powerful illusion.For the impersonal and anonymous experience of the il y a only serves to confirm the ubiquity of being, its vice-grip hold on the self.Far from offering freedom from being, the il y a represents for Levinas the "no exit" par excellence.Put otherwise: there is so proper solution to the problem of human existence from within the strict parameters of being itself.
The existential desert of the il y a represents, for him, the antithesis of human encounter and solidarity.The drama of being unfolds within an anonymous space devoid of an accountable "I" who can genuinely express "here I am" 8 in response to the ethical demands of a "You" who stands before it.As far as Levinas is concerned, the only genuine site of transcendence is to be found in another drama, one that his later work refers to as the "ethical intrigue" ( [12], p. 200).With this second drama, another figure appears on the scene: the Other-and with, the Other, a hint or trace of the divine also makes itself known.Levinas's mature work attempts to describe the nature of the invisible drama that draws together self, Other, and an ever-receding Wholly Other who we are constantly in danger of ossifying or idolizing as a result of the inevitably self-limiting nature of names and concepts.As a way to avoid this danger, Judaism has a long-standing tradition of proliferating the names for the absolute Other.Levinas repeats this gesture in his own writing, employing, as he does, a multitude of names for this third party within the ethical drama of human existence: the traditional designation of "God", the philosophical term "Infinite", and even neologisms like "Illeity".It is critical for Levinas that we avoid focusing on this third at the expense of the other parties that make up the ethical intrigue.The encounter with the human Other gives rise to what Levinas calls the "idea of God that comes to mind"-a jarring awareness that something greater, Infinite, breaks through in that encounter, but which simultaneously removes itself from my reach.As such, for Levinas, talk of "God" is legitimate only within the context of the asymmetrical relationship between myself and the flesh-and-blood Other who, from an ethical perspective, always has the upper hand over me."Ethics is not the corollary of the vision of God", the Jewish philosopher states, "it is the very vision.Ethics is an optic, such that everything I know of God and everything I can hear of His word and reasonably say to Him must find an ethical expression" ( [13], p. 17).The Other impinges, breaks through with her ethical demands, before I have had an opportunity to respond or make sense of her.It is in that relationship that I am left with a trace, a hint, of a force that we can legitimately-but always with the utmost caution-name "God".
Ethical substitution-one of the central concepts of his later philosophy-is meant to capture the peculiar relation that is established between myself and the others in my midst.For Levinas, the other human being has the remarkable capacity to morally interrogate my existence.The Other accuses me, challenges me, and makes seemingly infinite demands on me.These actions on the part of the Other are not initiated by the self.That is, left to ourselves, we would gladly pursue our own selfish desires.However, the Other will not have it that way.She questions my very attempt to establish myself as the center of my reality.There is a kind of displacement that transpires when we encounter others.Our entire focus shifts-from self-concern and self-preservation-to a responsiveness that is reoriented outside of the self.This displacement is so radical at times that the self finds itself seemingly occupying the Other's place.I can feel so utterly obligated towards the Other as to literally feel that I have been taken hostage by him or her.This is what Levinas means by "substitution" [14].For Levinas, ethical substitution is at the heart of our humanity.If the Other lacked this capacity to alter our way of being in the world, we would be condemned to the meaningless vagaries of the il y a as well as to a state of perpetual war, that is, an endless, futile battle of ego against ego.As a result of the ethical drama, the 8 In the Jewish scriptures, the "here I am" [hineni] expresses the most vigilant form of readiness before the otherness of God, as attested mostly notably by Abraham and Moses.See, for e.g., Gen. 22:1 and 22 :11.threat and fascination of the il y a within the tragic drama of being provisionally loses its grip on the human imagination, to be replaced by the promise of hope, frankness, and responsibility.
The nature of ethical substitution is such that it powerfully suggests, for Levinas, a connection between the other human being and the divine or Infinite.That the Other can make the self morally question its place under the sun and can do so by short-circuiting the cogito's representational and rationalizing capacities suggests to Levinas-adapting a famous expression of Plato's-the trace of the good beyond being.This insight receives, for Levinas, its first major articulation with the advent of Judaism.The Jewish scriptures identifies the essence of the divine with an obligation to look after the "widow", "orphan", and "stranger", that is, the most vulnerable in society (for example, Deut.10:18).The New Testament extends this view, for example, in the admonition that if religion has any authentic significance, it does so, not as a set of badges that one wears to publicly express one's supposed religiosity, but rather in one's ethical comportment, and specifically-repeating almost word for word the earlier Mosaic demand-in one's responsibility "to look after orphans and widows in their distress" (James 1:26-27).For Levinas, the ethical commandment that comes from the Other highlights the revelatory nature of ethics.By emphasizing the exteriority of the call, ethics for Levinas is itself a kind of religious encounter, albeit invisible.For, as Levinas sees it, the trace of the divine always already recedes from the ethical relation even as it underwrites the relation as such.
Given Levinas's emphatic position that what is essential to the human drama remains invisible to the phenomenological gaze, it would seem that art and its various methods of representation cannot do justice to the religious or spiritual background of the ethical intrigue.An initial reading of Levinas's most well-known statement on aesthetics, his 1948 essay "Reality and its Shadow," would strongly suggest that not only can art not do justice to the human intrigue, but that it will inevitably distort and displace what is truly at stake in that intrigue.In this essay and later in Totality and Infinity, Levinas maintains that art substitutes façade for the face, illusory spectacle for the provocations of the ethical intrigue."Reality and its Shadow" decries the image as that which "marks a hold over us rather than our initiative, [it is] a fundamental passivity" ( [15], p. 132).The image, Levinas goes on to say, is a form of "incantation", a primal rhythm that possesses the subject, rather than a mediated concept that maintains a critical distance from the object to which it refers.What seems most vexing for Levinas is the capacity for the image to disengage us from reality, to obscure the exigencies of material existence.Art, for Levinas, belongs to a realm prior to creation, where indeterminacy rules and subjectivity has yet to emerge or has been eclipsed.The plasticity of art, in other words, transports us to the inhuman realm of the il y a.Its lifeless images point to other formless images and away from the ethical intrigue.In this way, art shares the same nature as the hallucinations that possess the self when it finds itself in the grip of the il y a.In short, the aesthetic experience could not be further removed both in practice and in the content that it divulges from the ethical encounter of self and Other.Levinas does not limit these critical remarks to the plastic arts, but extends them to include "music, literature, theatre", and, not surprisingly, "cinema" as well ( [15], p. 139).
But is this the entire story?Is Levinas's final view then that art-cinema included-is fundamentally irreconcilable, even in conflict, with the spiritual and ethical truths of human existence?To get beyond the impasse that such an interpretation leads to, it might be useful to keep in mind a couple of points concerning Levinas's critique that have already been foregrounded by commentators like Richard Kearney [16], but which bear repeating in this present context.First, Levinas, much like Plato before him, and with whom he is often linked on this issue, does not condemn art tout court as much as challenge some of the more grandiose claims that are made on its behalf.In this respect, art, Levinas states categorically in "Reality and its Shadow", is "not the supreme value of civilization" as some would have us believe ( [15], p. 142).The second point is that Levinas's privileging of criticism over art does not rule out the possibility that the artist and the critic can be one in the same-a claim that is touched on towards the end of "Reality and its Shadow", and developed further in his subsequent essays on aesthetics.In later essays on writers like Celan and Proust [17], Levinas expresses his praise for a form of literature that calls into question the very illusory trappings of art.
Such art is sometimes deployed in the service of drawing our attention to both the ethical demands of the Other and the ways that we evade them.Of course, even if we acknowledge the possibility of an ethically oriented literature in Levinas's thought, the question still remains as to whether or not the same can be said of cinema.There are so few references to cinema in Levinas's work as to make it impossible to definitively answer this question on the basis of textual evidence alone.However, it might not be so difficult for us to imagine what such a redeemed cinema might resemble.Certainly, the history of cinema criticism provides us with some possible examples, including two from the list of great luminaries of film theory: Siegfried Kracauer and André Bazin.When Kracauer declares that the technical features of film must be deployed in the service of the camera's capacity to record reality [18], it is impossible to miss the ethical imperative behind his claim.For Kracauer, film should be a form of truth-telling.The ethical and spiritual force behind Bazin's theory of realism is likewise evident.Bazin describes the realist filmmaker's vocation as one of love for the reality he or she captures with the camera.
There is no way of completely understanding the art of Flaherty, Renoir, Vigo, and especially Chaplin unless we try to discover beforehand what particular kind of tenderness, of sensual or sentimental affection, they reflect.In my opinion, the cinema more than any other art is particularly bound up with love ( [19], p. 72).
The Dardenne brothers undoubtedly could be added to this list of filmmakers who demonstrate a genuine compassion for their characters or subjects.I also suspect that Levinas would regard the creative efforts of his former student and his brother as a redemptive form of cinema, one that grapples with the essentials of the human drama.
As deeply engaged and responsible filmmakers, the Dardenne brothers consciously employ the tools of cinematography not to dazzle their audience, but rather to capture what defines us as both individuals and partners in the human intrigue.In this regard, the brothers are very aware of how images when treated for their own sake can, and most certainly do, traffic in the inhuman nature of indeterminate being.Thus Luc Dardenne writes in his diary: Many hours spent in front of a television screen: broadcasting a neutral, muted flux, a continuum that envelops, an indeterminate presence that numbs.A state of torpor that closely resembles what Levinas writes in relation to the il y a, the murmur of being.No word, no sound, no image can bring this state to an end . . .To drown in the flux, in the density of the void . . .that is the deep desire of the telespectator . . .( [2], p. 59).
To this use of imagery, which is of course not limited to the moving images of television but is equally present on the screens of movie theatres, Luc Dardenne counterpoises another type of image, one that speaks to the ethical intrigue rather than to the seductive deluge of vapid, amorphous images which have the power to seize us and pull us back into the void of the existential wasteland of the tohu wa bohu.
Faith without Faith: A Postsecular Cinema
The Dardenne brothers, I maintain, offer us an excellent example of an ethical cinema-one that tracks the key contours of the ethical intrigue and the perils of the drama of being wherein the ego takes flight into the seductive allure of anonymous existence.Like Levinas, they too share an apprehension of the image 9 , of its propensity for the inhuman.The brothers charge themselves with the demanding task of representing the ineffable features of our humanity.The desire to represent the deepest aesthetic, moral, and spiritual dimensions of the human being seems further complicated, even insurmountable, 9 For an extended discussion of the Jewish apprehension with the image as it applies to Levinas's thought and cinema, see Libby Saxton's "Blinding Visions: Levinas, Ethics, Faciality."[20].
in the case of artists whose vocation is cinema.If the prohibition of the graven image means that the static image is minimally met with initial reservation, then it is doubly so with the moving image.Cinema's capacity to reflect back both the surface and movement of reality makes it an even more powerful tool of enchantment and sorcery.The consternation with the image is intensified in the case of film, because cinema's formal features (i.e., special effects, montage, etc.) make it possible to graphically render the contents of the world as well as the human imagination in spectacular and mesmerizing ways.Although, as Bazin, amongst others, points out, cinema is at the same time remarkably well suited to give testimony to experiences that often elude other media.
Rejecting the idea of creating images for their own sake or for the purpose of enchanting an audience, the Dardenne brothers have set for themselves the alternative task of offering testimony to the human condition and in particular to witnessing the ethical intrigue.Over the past two decades, they have created films that attempt to represent the transformative possibilities of ethical substitution.Their distinct camera work can be conceived as a meticulous spiritual exercise, one which permits them access to phenomena and experiences that otherwise fail to register in our minds.The use of the term "spiritual" in this context is not contrived.The Dardenne brothers are not averse to using that language themselves to talk about their work.As they note in one interview "perhaps filming gestures and very specific, material things is what allows the viewer to sense everything that is spiritual, unseen, and not a part of materiality" ( [21], p. 132).The movement of their camera is painstakingly attentive to the materiality of the world-in the service of showing what otherwise might go unseen: "[w]e tend to think that the closer one gets to the cup, to the hand, to the mouth whose lips are drinking, the more one will be able to feel something invisible-a dimension we want to follow and which would otherwise be less present in the film" ( [21], p. 132).Developing this observation in greater detail, Luc Dardenne, in one of his most perceptive film journal entries, notes: The movements of our camera are rendered necessary by our desire to be in things, inside the relations between glances and bodies, bodies and scenery.If the camera films a body in profile, immobile, with a wall behind it, and if this body begins to walk along the wall, the camera will go there, passing in front of the body, slipping between the wall and the body making a movement that will frame the body in profile and the wall behind it, and then the wall and the body . . .( [2], p. 138).
We immediately recognize in these words a succinct description of one of the hallmarks of their cinema.Their distinctive way of tracking the human body is evident from La Promesse (1996) [22] onwards.It is especially pronounced in Le Fils (2002) [23] in which most of the narrative telling is communicated through the presence of the body rather than the manipulation of narrative temporality, that is, montage.A brief analysis of the opening sequence reveals the intricacies of their cinematic style.The body is the precise starting point of the film just as the title credit appears.After a few seconds of the black backdrop of the title credit that first appears we suddenly realize that what we are actually seeing is a darkened, extreme close-up of the protagonist's back.From that moment onwards, the camera meticulously trails every movement of Olivier's restless body.This often produces the impression that the camera is always just catching up to his corporeal shifts and turns.This approach to filming the human body is established in the opening shot of the film.It is maintained for the duration of the film.In the opening seconds of the film, as the camera slowly moves away from his back, it gradually shifts upwards to shoulder level.We now see the back of Olivier's head and neck (Figure 1).From there, it pans leftward, revealing in the process a middle-aged woman (Figure 2).The camera then pans rightward, returning to the previous vantage point directly behind his head and neck.It then moves slowly down following the contours of his left arm and stops to show us that Olivier is holding a document (Figure 3)-presumably something that was just handed to him by the woman.We are always just an arm's reach away from the objects and people in Olivier's immediate environment.A minute later-after a brief interruption involving a malfunctioning circular saw-we are made to appreciate, again through his bodily gestures, that the document communicates something jarring (Figure 4).The body registers an apprehension.The camera moves in even closer.Olivier's troubled face fills the screen in profile.And while he continues to move the camera suddenly stops.The face of the woman who just a moment or two earlier stood before him now occupies the frame (Figure 5).The expression on her face mirrors Olivier's consternation.Like this sequence, the rest of the film intimately follows the motility of Olivier's body as it contends with its ever-dynamic environment.
Religions 2016, 7, 43 7 Olivier's consternation.Like this sequence, the rest of the film intimately follows the motility of Olivier's body as it contends with its ever-dynamic environment.Olivier's consternation.Like this sequence, the rest of the film intimately follows the motility of Olivier's body as it contends with its ever-dynamic environment.Olivier's consternation.Like this sequence, the rest of the film intimately follows the motility of Olivier's body as it contends with its ever-dynamic environment.What motivates this "desire to be in things"?Luc Dardenne poses this very question: "Why this desire that my brother and I share absolutely?Why don't we keep our distance from bodies…?" Reflecting as he does on the intensity of their camera's attentiveness in relation to the phenomena at hand, the Belgian cineaste, directs his self-questioning specifically to the human subjects that their films track with unrelenting dedication."Why these solitary, uprooted, nervous, bodies…?"And, furthermore, why, he inquires, do the brothers not film these bodies at a distance?To which he confesses that he and his brother, of course, "would like to…but something in us resists that", and compels them, instead, to do otherwise.And, then, finally, an admission of what is to be found "inside" those things and bodies."Perhaps it's because we find there", he continues, "close to things, between bodies, a presence of the human reality, a fire, a warmth that irradiates, that burns and insulates us from a sadness that reigns in the void, the very great void in life.It is our way of not despairing, to have faith again" ([2], pp.138-39).Rather than adopting the safety of distance which their lens afford them, their camera, instead, consciously plunges into the very being of their subject matter, inhabiting it in much the same way that meditation requires one to be present to intrusive or even painful thoughts and sensations, rather than to flee them, as we are all want to do.
The materiality that the brothers have in mind is not the crude sort that is often peddled by reductive materialism.There is an irreducible depth in material existence.Disrupting the drama of being-at the material level of vulnerable, exposed bodies encountering one another-is the ethical intrigue.Again, Levinas seems to be a major influence here.In the final phase of his philosophy, Levinas prefers the term proximity to his earlier invocation of the face.The face he had insisted on earlier is not simply what we see, its plasticity.There is "more" to the face, a surplus beyond what is observed by our senses.As a result of repeated misunderstanding, Levinas began to employ "proximity" as a way to suggest a form of irreducible presence to our embodied life-that marks a paradoxical experience that is at once near and intimate but simultaneously far (in the sense that it is outside my capacity to control) and strange-that yet exceeds my capacity to master it.No matter how much I try, I cannot assimilate the excessive surplus of the Other's proximity.The Other's What motivates this "desire to be in things"?Luc Dardenne poses this very question: "Why this desire that my brother and I share absolutely?Why don't we keep our distance from bodies…?" Reflecting as he does on the intensity of their camera's attentiveness in relation to the phenomena at hand, the Belgian cineaste, directs his self-questioning specifically to the human subjects that their films track with unrelenting dedication."Why these solitary, uprooted, nervous, bodies…?"And, furthermore, why, he inquires, do the brothers not film these bodies at a distance?To which he confesses that he and his brother, of course, "would like to…but something in us resists that", and compels them, instead, to do otherwise.And, then, finally, an admission of what is to be found "inside" those things and bodies."Perhaps it's because we find there", he continues, "close to things, between bodies, a presence of the human reality, a fire, a warmth that irradiates, that burns and insulates us from a sadness that reigns in the void, the very great void in life.It is our way of not despairing, to have faith again" ([2], pp.138-39).Rather than adopting the safety of distance which their lens afford them, their camera, instead, consciously plunges into the very being of their subject matter, inhabiting it in much the same way that meditation requires one to be present to intrusive or even painful thoughts and sensations, rather than to flee them, as we are all want to do.
The materiality that the brothers have in mind is not the crude sort that is often peddled by reductive materialism.There is an irreducible depth in material existence.Disrupting the drama of being-at the material level of vulnerable, exposed bodies encountering one another-is the ethical intrigue.Again, Levinas seems to be a major influence here.In the final phase of his philosophy, Levinas prefers the term proximity to his earlier invocation of the face.The face he had insisted on earlier is not simply what we see, its plasticity.There is "more" to the face, a surplus beyond what is observed by our senses.As a result of repeated misunderstanding, Levinas began to employ "proximity" as a way to suggest a form of irreducible presence to our embodied life-that marks a paradoxical experience that is at once near and intimate but simultaneously far (in the sense that it is outside my capacity to control) and strange-that yet exceeds my capacity to master it.No matter how much I try, I cannot assimilate the excessive surplus of the Other's proximity.The Other's What motivates this "desire to be in things"?Luc Dardenne poses this very question: "Why this desire that my brother and I share absolutely?Why don't we keep our distance from bodies . . .?" Reflecting as he does on the intensity of their camera's attentiveness in relation to the phenomena at hand, the Belgian cineaste, directs his self-questioning specifically to the human subjects that their films track with unrelenting dedication."Why these solitary, uprooted, nervous, bodies . . .?" And, furthermore, why, he inquires, do the brothers not film these bodies at a distance?To which he confesses that he and his brother, of course, "would like to . . .but something in us resists that", and compels them, instead, to do otherwise.And, then, finally, an admission of what is to be found "inside" those things and bodies."Perhaps it's because we find there", he continues, "close to things, between bodies, a presence of the human reality, a fire, a warmth that irradiates, that burns and insulates us from a sadness that reigns in the void, the very great void in life.It is our way of not despairing, to have faith again" ([2], pp.138-39).Rather than adopting the safety of distance which their lens afford them, their camera, instead, consciously plunges into the very being of their subject matter, inhabiting it in much the same way that meditation requires one to be present to intrusive or even painful thoughts and sensations, rather than to flee them, as we are all want to do.
The materiality that the brothers have in mind is not the crude sort that is often peddled by reductive materialism.There is an irreducible depth in material existence.Disrupting the drama of being-at the material level of vulnerable, exposed bodies encountering one another-is the ethical intrigue.Again, Levinas seems to be a major influence here.In the final phase of his philosophy, Levinas prefers the term proximity to his earlier invocation of the face.The face he had insisted on earlier is not simply what we see, its plasticity.There is "more" to the face, a surplus beyond what is observed by our senses.As a result of repeated misunderstanding, Levinas began to employ "proximity" as a way to suggest a form of irreducible presence to our embodied life-that marks a paradoxical experience that is at once near and intimate but simultaneously far (in the sense that it is outside my capacity to control) and strange-that yet exceeds my capacity to master it.No matter how much I try, I cannot assimilate the excessive surplus of the Other's proximity.The Other's irreducible strangeness unsettles me, pulls me out of myself, denudes me, exposes me, turns me inside out.
The debt owed to Levinas can be seen in a number of their films, and, perhaps, most explicitly in L'Enfant (2005) [24], which received the prestigious Palme d'Or award at Cannes in 2005.This film, like their previous three major features (La Promesse, Rosetta, and Le Fils), as well as subsequent films since then (like Deux jours, une nuit), illustrates the redemptive force of responsibility, specifically, how the self is transformed in its response to the Other's unrelenting interrogations and demands.L'Enfant tracks the movement from the arid, narcissistic, often violent, universe that the ego inhabits to the ethically meaningful world that is made possible by the Other's provocation to respond.In other words, this film, like their other work, narrates the tensions that exist between the drama of being and the ethical intrigue of self and Other.It is important to note, however, that these dramas do not play themselves out in two distinct realities, one material, the other, immaterial.Rather, we experience ethical and spiritual transcendence in the very heart of embodied, material existence.In the first installment of Au Dos de nos images, Luc Dardenne singles out the following line from one of Levinas's commentaries on rabbinical texts dealing with the messianic: "spiritual life is essentially a moral life and its home is the economic sphere" ([13], p. 62; trans.modified).To which the filmmaker immediately adds: "this view of Levinas is also that of our cinema" ([2], p. 71).
The film features a freewheeling homeless youth by the name of Bruno and his girlfriend Sonia.Bruno lives Sonia's welfare checks and his petty theft activity.We learn at the beginning of the film that Sonia, his girlfriend has just had their baby.While in the hospital, without telling her that he has done so, he sublets her apartment in order to pocket the extra money.When she finally tracks him down on the street to introduce him to his son, the hurried Bruno can barely spare a moment for his own prodigy.When Sonia lifts Jimmy, their son, so that Bruno might embrace him, his awkwardness barely conceals his disinterest.Indeed, he seems perversely relieved when a second or two later he notices that the would-be-victim he has been staking across the street is about to get away and he uses this as an opportunity to call his accomplice on his cell phone.That he reaches for his phone rather than the creature in Sonia's extended arms establishes for the viewer the extent of his self-immersion.Bruno's self-regard reaches its high-point when tired of having to wait in line for Sonia's social assistance cheque he decides to take up the offer of one of his shady associates who claims to know someone who traffics in babies for the adoption black-market.Some time later that afternoon, Bruno, unbeknownst to Sonia, exchanges his own flesh and blood for a hefty wad of cash.
For someone who has not seen the film, the temptation at this point in my description of it might be to imagine Bruno as evil incarnate.Yet, the film discourages this view.The fact is that for much of the film, Bruno comes across as a likeable person.His playful demeanor and his easy-going attitude make him no different than any typical twenty-something male in our midst.As seen through the camera of the Dardenne brothers, Bruno is no monster.If he is capable of carrying out as objectionable a deed as selling his own son, we see that this has more to do with his all-too-human egoism and indifference than it does to some ostensible evil inclination operating within him.If we must speak of evil in this context then we must qualify it as Hannah Arendt did as utterly banal [25].From his own selfish perspective, Bruno's decisions appear to him not only beneficial but justified as well.Selling Jimmy for a few thousand Euros appears to him eminently reasonable.His son, after all, will be in better hands and he and Sonia desperately need the money.Thus when the alarmed Sonia asks him where Jimmy is, after returning to her with an empty baby stroller, Bruno responds matter-of-factly, "I sold him".To which he then adds, as if to reassure her, "don't worry we can always make another one".
Bruno's phenomenology resembles that of the natural ego that Levinas so vividly describes in Totality and Infinity [26].His own immediate needs lead him in general to overlook the fact that the world is populated by distinct others.Bruno carries himself in such a way as to make the most of what he can get from others at the same time that he shuns their accusations or demands.The viewer cannot help but be struck by the fact that for virtually the entire film, Bruno avoids his son's gaze.Similarly, throughout much of the film, even while in the midst of others, Bruno is shown to be engrossed with his cell-phone, as if to suggest that he prefers the distancing that such technologies afford him over the directness of human proximity.However, the power to stand back from the world-viewing the world as a spectacle for one's own benefit-concomitantly reveals the self's fettered state.Bruno's universe is the utilitarian world of maximizing pleasure and avoiding pain.However, that way of comporting oneself, as Levinas shows, belongs to the impersonal drama of being.That is, the universe as it exists prior to the ethical orientation revealed by the Other, a universe comprised exclusively of things and images in flux and not creatures assuming the responsibility that proximity to one another demands.The flipside of the ego's desire for anonymity and sovereignty is existential and spiritual imprisonment.
The theme of self-imprisonment and isolation is central for the Dardenne brothers as it is in the cinema of one of their most important influences, that is, the work of Robert Bresson.The older French director is worth noting for our purposes because he too sets for himself the goal of representing the hidden spiritual drama of the human being.Admittedly, Levinas and Bresson are working on different philosophical and religious registers.For Levinas, the inaccessibility of the divine acts as a foil that is meant to return us towards-what he calls the supreme "detour" 10 -the only legitimate access to the divine, namely, the inter-human drama.By contrast, for the Catholic Bresson, the distance from the divine serves to underscore man's fallenness.For Bresson, if the Hidden God reveals Himself he does so primarily as the result of the unmerited grace that God shows for His creatures.Despite this difference-which ought not to be exaggerated, because frequently for Bresson the catalyst for spiritual conversion involves other human beings-there are shared concerns between Bresson and Levinas that are worth noting.These similarities, as we will see, can be gleaned from the work of the Dardenne brothers who are indebted to both figures.
Like Levinas, Bresson has an eye for the deep ambiguity that riddles the self.The ego for both of them simultaneously represents sovereignty and imprisonment.On the one hand, the fact that the ego emerges against the background of an infinitely vast universe is, to use Pascal's language, a virtual miracle.The ego's emergence is a major accomplishment in the face of the anonymous indeterminacy of being.On the other hand, the ego faces its own existence as a crushing burden from which it cannot free itself.Bresson, especially in his early and middle films, made use of the prison theme as a metaphor for the human condition.The Dardenne brothers similarly convey the insularity of Bruno by frequently showing him against a wall, in a solitary space, contained by a fence, looking through bars (Figures 6-8).A similar device was used in Rosetta to heighten the principal character's isolation from the world (Figure 9).
Religions 2016, 7, 43 10 afford him over the directness of human proximity.However, the power to stand back from the world-viewing the world as a spectacle for one's own benefit-concomitantly reveals the self's fettered state.Bruno's universe is the utilitarian world of maximizing pleasure and avoiding pain.However, that way of comporting oneself, as Levinas shows, belongs to the impersonal drama of being.That is, the universe as it exists prior to the ethical orientation revealed by the Other, a universe comprised exclusively of things and images in flux and not creatures assuming the responsibility that proximity to one another demands.The flipside of the ego's desire for anonymity and sovereignty is existential and spiritual imprisonment.
The theme of self-imprisonment and isolation is central for the Dardenne brothers as it is in the cinema of one of their most important influences, that is, the work of Robert Bresson.The older French director is worth noting for our purposes because he too sets for himself the goal of representing the hidden spiritual drama of the human being.Admittedly, Levinas and Bresson are working on different philosophical and religious registers.For Levinas, the inaccessibility of the divine acts as a foil that is meant to return us towards-what he calls the supreme "detour" 10 -the only legitimate access to the divine, namely, the inter-human drama.By contrast, for the Catholic Bresson, the distance from the divine serves to underscore man's fallenness.For Bresson, if the Hidden God reveals Himself he does so primarily as the result of the unmerited grace that God shows for His creatures.Despite this difference-which ought not to be exaggerated, because frequently for Bresson the catalyst for spiritual conversion involves other human beings-there are shared concerns between Bresson and Levinas that are worth noting.These similarities, as we will see, can be gleaned from the work of the Dardenne brothers who are indebted to both figures.
Like Levinas, Bresson has an eye for the deep ambiguity that riddles the self.The ego for both of them simultaneously represents sovereignty and imprisonment.On the one hand, the fact that the ego emerges against the background of an infinitely vast universe is, to use Pascal's language, a virtual miracle.The ego's emergence is a major accomplishment in the face of the anonymous indeterminacy of being.On the other hand, the ego faces its own existence as a crushing burden from which it cannot free itself.Bresson, especially in his early and middle films, made use of the prison theme as a metaphor for the human condition.The Dardenne brothers similarly convey the insularity of Bruno by frequently showing him against a wall, in a solitary space, contained by a fence, looking through bars (Figures 6-8).A similar device was used in Rosetta to heighten the principal character's isolation from the world (Figure 9).What finally breaks the illusion of absolute sovereignty is the Other's accusation.This provocation, and the possibility to respond ethically on the part of the self, offers the only genuine way to transcend the ego's limited purview.In response to the Other's provocation, the self undergoes a kind of self-hollowing or self-evisceration.The Other erodes the self's psychic armor, so to speak, and, in the process, makes the self receptive to the demands of others.Only the Other's face-or, proximity, to use Levinas's later preferred term-has the power to address the ego, to call it out of its tendency to hide or dissolve itself into its surroundings.In the absence of the Other, that is, where immanence prevails, "everything is absorbed, sunken into, walled in being..." ( [14], p. 182).What finally breaks the illusion of absolute sovereignty is the Other's accusation.This provocation, and the possibility to respond ethically on the part of the self, offers the only genuine way to transcend the ego's limited purview.In response to the Other's provocation, the self undergoes a kind of self-hollowing or self-evisceration.The Other erodes the self's psychic armor, so to speak, and, in the process, makes the self receptive to the demands others.Only the Other's face-or, proximity, to use Levinas's later preferred term-has the power to address the ego, to call it out of its tendency to hide or dissolve itself into its surroundings.In the absence of the Other, that is, where immanence prevails, "everything is absorbed, sunken into, walled in being..." ( [14], p. 182).What finally breaks the illusion of absolute sovereignty is the Other's accusation.This provocation, and the possibility to respond ethically on the part of the self, offers the only genuine way to transcend the ego's limited purview.In response to the Other's provocation, the self undergoes a kind of self-hollowing or self-evisceration.The Other erodes the self's psychic armor, so to speak, and, in the process, makes the self receptive to the demands of others.Only the Other's face-or, proximity, to use Levinas's later preferred term-has the power to address the ego, to call it out of its tendency to hide or dissolve itself into its surroundings.In the absence of the Other, that is, where immanence prevails, "everything is absorbed, sunken into, walled in being..." ( [14], p. 182).What finally breaks the illusion of absolute sovereignty is the Other's accusation.This provocation, and the possibility to respond ethically on the part of the self, offers the only genuine way to transcend the ego's limited purview.In response to the Other's provocation, the self undergoes a kind of self-hollowing or self-evisceration.The Other erodes the self's psychic armor, so to speak, and, in the process, makes the self receptive to the demands of others.Only the Other's face-or, proximity, to use Levinas's later preferred term-has the power to address the ego, to call it out of its tendency to hide or dissolve itself into its surroundings.In the absence of the Other, that is, where immanence prevails, "everything is absorbed, sunken into, walled in being..." ([14], p. 182).The temptation of the ego-its The ending of L'Enfant is reminiscent of the moral conversion undergone by some of Dostoevsky's own male protagonists in the presence of female companions who they have otherwise mistreated.One thinks here of Liza in Notes from Underground and, of course, Sonia's namesake in Crime and Punishment.Dostoevsky we know consciously conceived of these female characters as Mary Magdalene figures, and more importantly, as Mary Magdalene transfigured by the divine presence of Christ.These female characters resemble Christ in two important ways: like him, their essential goodness makes them targets of abuse, and like him they demonstrate an extraordinary capacity for love and compassion even in the face of the bitter hatred that is personally directed at them (as a side-note, the original title for L'Enfant was supposed to be "The Force of Love").This idea is equally central for Bresson who we know was profoundly influenced by Dostoevsky's literature.At the end of Bresson's Pickpocket (Figure 12), the male protagonist finally breaks down before the infinite patience of Jeanne who he had previously rebuffed and neglected.The final words of Pickpocket-"Oh Jeanne, to reach you, what a strange path I had to take"-reminds me of Levinas's recourse in Otherwise than Being to an old Portuguese adage which he notes succinctly summarizes the ethical drama that implicates self, Other, and God.The proverb in question is "God writes straight with crooked lines" ([14], p. 147).The crooked lines here refer to the torturous and strained relations between self and Other in the human intrigue.The redemptive possibilities that open up as a result of these "crooked lines" gesture towards another more radical and invisible source of the human intrigue, namely, the divine.The ending of L'Enfant is reminiscent of the moral conversion undergone by some of Dostoevsky's own male protagonists in the presence of female companions who they have otherwise mistreated.One thinks here of Liza in Notes from Underground and, of course, Sonia's namesake in Crime and Punishment.Dostoevsky we know consciously conceived of these female characters as Mary Magdalene figures, and more importantly, as Mary Magdalene transfigured by the divine presence of Christ.These female characters resemble Christ in two important ways: like him, their essential goodness makes them targets of abuse, and also like him they demonstrate an extraordinary capacity for love and compassion even in the face of the bitter hatred that is personally directed at them (as a side-note, the original title for L'Enfant was supposed to be "The Force of Love").This idea is equally central for Bresson who we know was profoundly influenced by Dostoevsky's literature.At the end of Bresson's Pickpocket (Figure 12), the male protagonist finally breaks down before the infinite patience of Jeanne who he had previously rebuffed and neglected.The final words of Pickpocket-"Oh Jeanne, to reach you, what a strange path I had to take"-reminds me of Levinas's recourse in Otherwise than Being to an old Portuguese adage which he notes succinctly summarizes the ethical drama that implicates self, Other, and God.The proverb in question is "God writes straight with crooked lines" ([14], p. 147).The crooked lines here refer to the torturous and strained relations between self and Other in the human intrigue.The redemptive possibilities that open up as a result of these "crooked lines" gesture towards another more radical and invisible source of the human intrigue, namely, the divine.The ending of L'Enfant is reminiscent of the moral conversion undergone by some of Dostoevsky's own male protagonists in the presence of female companions who they have otherwise mistreated.One thinks here of Liza in Notes from Underground and, of course, Sonia's namesake in Crime and Punishment.Dostoevsky we know consciously conceived of these female characters as Mary Magdalene figures, and more importantly, as Mary Magdalene transfigured by the divine presence of Christ.These female characters resemble Christ in two important ways: like him, their essential goodness makes them targets of abuse, and also like him they demonstrate an extraordinary capacity for love and compassion even in the face of the bitter hatred that is personally directed at them (as a side-note, the original title for L'Enfant was supposed to be "The Force of Love").This idea is equally central for Bresson who we know was profoundly influenced by Dostoevsky's literature.At the end of Bresson's Pickpocket (Figure 12), the male protagonist finally breaks down before the infinite patience of Jeanne who he had previously rebuffed and neglected.The final words of Pickpocket-"Oh Jeanne, to reach you, what a strange path I had to take"-reminds me of Levinas's recourse in Otherwise than Being to an old Portuguese adage which he notes succinctly summarizes the ethical drama that implicates self, Other, and God.The proverb in question is "God writes straight with crooked lines" ([14], p. 147).The crooked lines here refer to the torturous and strained relations between self and Other in the human intrigue.The redemptive possibilities that open up as a result of these "crooked lines" gesture towards another more radical and invisible source of the human intrigue, namely, the divine.Luc Dardenne recognizes that religion traces the contours that delineate the dramas as described by Levinas.In this regard, he has on more than one occasion acknowledged the legacy of the monotheistic traditions in shaping our spiritual interiority and ethical sensibility.In an interview with the Belgian review, Toudi, he remarks: The biblical texts say what life should be according to God, life according to the Law, according to Love, according to Justice.If one sees something religious in our films, it's no doubt because the experiences of our characters refer to this life according to God which even without God today continues for us the life that is the most humanly possible, not for all situations in life, but at least for some [6].
Despite this admission, his own lack of faith leads him to speculate that the name of "God" is perhaps an echo of the mother's unremitting love for her child 11 .In this way, he departs from both Bresson and Levinas, for whom the divine or the Infinite cannot be traced back to any particular person or thing in the world.That is Luc Dardenne's wager.
Nevertheless, it is important to note that the "atheism" that is expressed in his understanding of his cinema challenges a dominant account of secularism that has until recently defined the modern age-but which we are only now seeing signs of a creative renewal.Today, philosophers like Jürgen Habermas [27] speak of "postsecularism".The postsecular is not an abandonment of secularism.It represents an effort to free Enlightenment secularity from the false dichotomies that it sets up between belief and unbelief, faith and reason.Most significantly, the postsecular drops the pretence that secular reason is fully rational and transparent while religion ostensibly is driven by an irrational faith.Instead, the postsecular position concedes-as Luc Dardenne is prepared to do-that faith in something outside of oneself is a necessary precondition for subjectivity in general.This is a faith in the Other's capacity to save me from the despair of impersonal existence, though not a faith in a personal deity: faith without faith.For these reasons, I think it best to conceive of the cinema of the Dardenne brothers as an exemplary expression of the postsecular.In her own exploration of their cinema, Sarah Cooper makes a similar claim: "This is post-secular filmmaking that is comfortable with the place of religion in an industrial world in which faith and unbelief sit side by side, and through which the remnants of Judeo-Christian commands and counsel live on in secular morals and conduct" [28].The postsecular cinema of the Dardenne brothers bears witness to the absolute necessity for transcendence from the binding limitations of our egoism, as the sine qua non of human solidarity.
Conclusions
The postsecular cinema of the Dardenne brothers poses a unique challenge to both a traditional understanding of faith and the modern concept of secularism.Some of Luc Dardenne's recent writing might be misinterpreted as a form of atheism.But this is no conventional atheism.He fully recognizes and respects that religion-and in particular, the monotheistic faiths-pays careful attention to a dynamic of faith that is so essential for the well being of the self and the solidarity of a community.Nevertheless, in his most recent work, he conceives of our current task as learning to live without recourse to a personal God.His reflections on these matters suggest that the brothers' cinema is an appreciation of the critical role that belief in an infinitely loving Other plays in the emergence of human subjectivity.Without such a belief, the self would collapse back on itself.And, at the same, time, their cinema can be understood as an attempt to live that faith without necessarily invoking in name or in practice the institutions or discourses of traditional religion.Consequently, I think it might be best to describe the cinema of the Dardenne brothers as an expression of faith without faith.
Figure 3 .
Figure 3. Opening Sequence of Le Fils.Figure 3. Opening Sequence of Le Fils.
Figure 3 .
Figure 3. Opening Sequence of Le Fils.Figure 3. Opening Sequence of Le Fils. | 13,038.2 | 2016-04-26T00:00:00.000 | [
"Philosophy",
"Art"
] |
An Artificial Management Platform Based on Deep Learning Using Cloud Computing for Smart Cities
In today's world; smart city management uses sensors, cameras and mobile devices which is internet-connected. These devices, called Internet of things (IoTs) generate large amounts of data. A big data-based approach is needed to store these data appropriately and to provide real-time access to the data. In smart city management applications that use many different sources of information, traditional machine learning methods for classifying large data and generating meaning can be inadequate. Deep learning approach is widely used today to solve similar problems. In this study, a cloud computing based architectural approach is proposed that enables data mining using deep learning on big data produced by IOTs.
Introduction
Smart city is a concept based on the principle of using technological infrastructure in city management for the purposes of management, planning, analysis, improving the quality of service new services [1].In this sense, smart cities include smart management, smart transportation, smart technologies, smart economy, smart health and so on, as in "Fig.1".In today's world, many countries are producing new projects for smart cities in metropolitan areas and allocating huge budgets for this purpose.This is the biggest disadvantage of smart cities as well as their advantages [2].In smart city applications, it is necessary to collect, evaluate and analyze the data to be obtained from many sources of information and sensors and to interpret the results, and the actions need to be achieved [3].The action to be achieved can be cost reduction, taking new managerial decisions, improving the quality of service and so on.For instance, smart traffic application gives real-time road information for the drivers by means of mobile applications.This information may indicate real-time situations such as instant traffic density on the route, if any, maintenance work and accident.The input data of the system to be created for this purpose consist of traffic lights, city surveillance and cameras.The system has outputs such as informing, dynamic operation of traffic lights and directions [4].Smart city applications generally require algorithms with different types of data inputs and outputs that will perform realtime learning on this big data.Traditional machine learning methods are not sufficient for smart city applications in terms of processing power and memory consumption.
Deep Learning (DP) is successfully applied in image recognition, object tracking, analysis and interpretation applications on big and multilayer data [5,6].Fig. 1.Smart city applications [5] Smart city applications, as in this example, generally require algorithms with different types of data inputs and outputs that will perform real-time learning on this big data.Traditional machine learning methods are not sufficient for smart city applications in terms of processing power and memory consumption.Deep Learning (DP) is successfully applied in image recognition, object tracking, analysis and interpretation applications on big and multilayer data [5,6].Deep Belief Networks, Convolutional Neural Networks (CNNs) and Deep Boltzmann Machines are widely used for DP applications in the literature.However, CNNs are more commonly used along with their easier training, need for fewer parameters, ease of implementation and success rates compared to other methods.
The biggest reasons for the popularity of deep learning networks after 2011 are the fact that learning process is simpler and that the costs of graphics process unit (GPU) are at very reasonable levels compared to other machine learning algorithms.Deep learning algorithms working on GPUs provide speed increase by tens of times compared to traditional Central Programming Units (CPUs) depending on GPU properties [8,9].Apart from PC and mobile devices, various sensors and electronic appliances have structures connected to internet via wifi or mobile networks.Nowadays, the number of these devices, that are called internet of things (IoT), has exceeded 1.2 billion [10].IoTs collect information such as moisture, heat, carbon dioxide ratio, motion, speed, etc. through the sensors they have.Thus, IoTs are widely used in remote and real-time applications such as smart building management, smart parking and smart monitoring.Even world's giant companies like Google are producing development cards and cloud-based software solutions for IoT, and IoTs expected to have a greater importance in our lives in the near future [11].When the number of IoTs used in distributed architecture for different purposes, the amount of data they produce and the rate of data generation are taken into account, the resulting data type is considered as big data [12].It is necessary to store these data by modeling with a suitable architecture for processing, to clear inconsistent and unnecessary data and to perform dimension reduction before processing the data.A Big data which is appropriate for this structure is usually not regular and consistent, therefore it cannot be stored as sql-based.For this purpose, it should mainly be stored in a cloud-based and distributed structure with a nosql -based database [13].Big data is mentioned by the concepts called 5V in the literature.These 5V concepts can be associated with smart cities and IoTs as the following.Variety: It refers to variety of data.Data is not hierarchical and uniform when IoTs that include image, video, automation sources and many sensors as input data in smart city applications are taken into account.Velocity and Volume: It refers to the growth rate and size of data.More than 1.2 billion IoT's worldwide, more than 200 million cameras that are connected to internet are recording 1.4 trillion hours video per year [8].Verification: The data obtained from many different sources do not have a hierarchical and homogeneous structure.It is necessary to filter out the data that are unnecessary or disturb the hierarchical structure for learning from data.Value: It refers to the value created by the data.It is the value created by the output which is obtained as a result of data analysis.A new smart city application, ensuring cost in the management of city resources and improving the quality of services offered to the public can be as examples.When the literature on smart city applications is examined, two major technologies, IoTs and Big Data, come to the forefront.Paganelli et al. defined a web architecture for accessing remote IoTs to be used in smart cities.In the study, each IoT was represented by a unique identity.Along with the Restful-based web architecture, the data from these IoTs were accessed via json apis [13].Dlodlo et al. defined the cloud-based data storage architecture by giving information about IoT development environments such as Aurdino and Raspery Pi for smart applications to be used in smart cities [15].
Shah and Misra implemented a mobile application that performs remote monitoring by reading the environmental variables such as moisture, temperature and CO2 in the environment through IoTs for smart environmental monitoring application with the aim of reducing air pollution [16].Sakhardande et al. proposed a structure that uses more than one IOT network with power supply and wifi network in order to perform monitoring even in disastrous situations in smart cities [17].Costa and Santos defined architecture for the use of Big Data in smart cities.In the study called Basis, they presented an architecture that uses hdfs in which Big Data is stored with cloudbased hadop technology [14].Horban proposed an architecture that performs data mining on Big Data for smart energy management by detailing the relationship between the concept of Big Data and smart cities [18].Alshawish et al. proposed an architecture for the use of Big Data in smart city applications.The prominent suggestion in the study is the 6-step, reusable Big Data pipeline architecture for smart management [19].
IoTs
Restfull framework for IOTs to be used in Smart City [13]
IoTs
Research on Iots platforms that can be used in Smart cities [15] IoTs IoTs-based approach for smart city's condition monitoring [16] IoTs IoTs-based framework for smart city disaster management [17] Big Data Big Data structure for smart cities [14] Big Data Data mining analysis on Big Data for smart energy use [18] Big Data Big Data applications in smart cities [19] Two major technologies, Big Data and IoT, come to the forefront in smart city applications in the literature examples summarized in Table-1.However, when today's needs for smart city applications are taken into account, it is thought that Deep Learning will be the most appropriate machine learning on Big Data from IoTs for the processes of real-time action decisions, analysis and acquisition of valuable information.
In addition to these two major technologies, the proposed system needs to use a cloud-based distributed architecture to handle large data and produce results.In this study, an architecture using Deep Learning was proposed for valuable knowledge mining in smart city applications.
Proposed Approach
In this study, an approach based on Deep Learning that uses CNN on Big Data coming from all IoTs was proposed for smart city management.The study involves the steps given in "Fig.2".A unique id that will replace the identity is given for each IoT.This is necessary to obtain the type, location, sensors and values of the relevant IoT [13] in a restful architecture as in the study.
The collected data should be stored in a distributed architecture in accordance with the Big Data architecture on a nosql-based cloud server.In Big Data architecture, data should be stored on distributed server clusters and be subjected to map-reduce process without processing [20].The outputs obtained after the mapreduce processing will create the entries for the Deep Learning training process.
Internet of Things (IoTs)
Electronic devices such as cameras, sensors and measuring devices are regarded as Things provided that they are connected to at least one sensor, a unique identity and internet.A device that fulfills these three conditions is accessible from all over the world, so it is manageable.Nowadays, IoT platform is provided with many corporate development cards and hardware and software supports.IoT developer kit for Google cloud-supported platform is seen in "Fig.3" [21].The developer kit has sensors such as wifi, acclerometer, temperature, light, rotary and distance and is ready for use via google cloud platform.
Big Data
When the fact that there are 1.2 billion IoT around the world is taken into account, Big Data is the type of data that occurs in terms of the variety, rate, size of the produced data and the value created by it.The basic properties that should be known about Big Data are explained below in accordance with the Big Data life cycle is seen in "Fig.4".
Hadoop Distributed File System (HDFS):
It is the file system consisting of distributed server clusters for Big Data.Map: Map-Reduce processing was announced by Google in 2004 [23].In the map phase, the data received from the host node is divided into smaller segments and distributed to the child nodes.Reduce: In this phase, concrete analysis results are obtained on the data obtained in the map phase.
Hadoop: It is a project that performs Map-Reduce process on Distributed File System [24].
Deep Learning
Deep Learning, which is a machine learning method and is a special form of Artifical Neural Networks (ANN) is successfully applied in applications such as information retrieval [25], image recognition, object tracking and language processing [26,27].CNNs are preferred for the reasons such as the simplicity of its training process, fast running in test phase and ease of application among Deep Learning Algorithms.In general, CNNs consist of 5 stages.
Input Layer: It creates the input data of the system.In the proposed method, the outputs of Big Data Map-Reduce process will create the CNN entry for smart city management.The use of outputs after Map-Reduce process in Deep Learning training will ensure that the training data does not contain any inconsistencies and that a model that characterizes the learning model very well will emerge.
Convolutional Layer: It is also known as subsampling.It is the process of subsampling by taking a core matrix of smaller size on input data as it is given in "1".
Pooling Layer: It is the feature extraction step.Feature selection is performed on the entries multiplexed in the previous step.A kernel matrix is also used while performing feature selection.For instance, In the event that 2x2 kernel matrix is used, max, min or mean value is selected as feature from a total of 4 values.Thus, a feature vector is obtained as given in "2,3".
Conclusions
Smart cities is a whole of concepts in which smart applications are used for the purposes of the management of city resources, improving the quality of service, obtaining valuable information that will affect managerial decisions and reducing costs.It consists of sub-elements such as smart traffic, smart health, smart building, smart monitoring and smart infrastructure.
IoTs are the source of information in smart city applications.The fact that they are manageable due to the sensors they have and their internet connection provide unmanned source of information in smart city applications.However, both the number of IoT, the amount of data produced and the rate of data generation require the use of a Big Data-based structure for IoTs.
Traditional machine learning algorithms will be inadequate in processing this data in terms of workload and working speeds in order to perform data mining, supervised/unsupervised learning in such applications where there is a large number of data input and the data type is Big Data.Especially after 2011, Deep Learning has been successfully applied in many applications such as object recognition, scene interpretation, language processing and driverless tools for the reasons of reduced GPU costs, working speeds, ease of implementation.
In this study, a deep learning based approach was proposed with the purpose of performing artifical management in smart cities where IoTs constitute information resources.
The proposed approach has four steps.It was projected that the raw data received from IoTs are stored in a distributed architecture in accordance with the Big Data architecture in the first step, that a consistent and well-modeled data set to be used in the training process is obtained by performing Map and Reduce process on Big Data in the second step, and that a CNNbased Deep Learning training is performed in the third step.The final step is that the whole system works in a distributed cloud architecture.Incoming requests are met by the load balancer and the most appropriate node is forwarded and processed.This approach provides for expandable, modular and parallel operation.Thus, in this study, an approach using deep learning was proposed with the aim of obtaining valuable information by performing data mining in a smart city management.
Fig. 2 .
Fig.2.Block diagram of the proposed approach
Fig. 4 .
Fig.4.Big data map-reduce process [22] 2.3.Cloud Computing All client service providers for today's big data-processing web applications.It uses cloud based technology to meet both scenarios such as disaster recovery and simultaneous service with all clients.The major advantages of cloud-based technology, shown in "Fig.5", are listed below [17,21,24]. Diseaster Recovery: One or more of the cloud services provide continuous protection, even if it is inoperable [17]. Load Balancing: Provides load balancing for all clients (monitoring) or IoTs interacting with the system. Distributed Computing: Data distributes the load on the appropriate services.The incoming service meets the most appropriate service. Backup: It provides affordable solution for backup and big data. Performance: Adding and removing new nodes to the network can be done easily depending on the performance required.
Table 1 .
Literature Review | 3,480.4 | 2017-08-21T00:00:00.000 | [
"Computer Science",
"Engineering",
"Environmental Science"
] |
Photon Emission Rate Engineering using Graphene Nanodisc Cavities
In this work, we present a systematic study of the plasmon modes in a system of vertically stacked pair of graphene discs. Quasistatic approximation is used to model the eigenmodes of the system. Eigen-response theory is employed to explain the spatial dependence of the coupling between the plasmon modes and a quantum emitter. These results show a good match between the semi-analytical calculation and full-wave simulations. Secondly, we have shown that it is possible to engineer the decay rates of a quantum emitter placed inside and near this cavity, using Fermi level tuning, via gate voltages and variation of emitter location and polarization. We highlighted that by coupling to the bright plasmon mode, the radiative efficiency of the emitter can be enhanced compared to the single graphene disc case, whereas the dark plasmon mode suppresses the radiative efficiency.
Introduction
Technological advances in the field of nanofabrication have provided a powerful tool to tailor light-matter interaction. Metallic nanoparticles support surface plasmon resonances where collective oscillations of electron and photons can result in a localization of electromagnetic fields into subwavelength scales [1]. Apart from localizing the incident plane waves, these plasmon modes strongly modify spontaneous emission properties of quantum emitters [2], such as quantum dots, placed close to them. In particular, radiative decay rates of such fluorescent particles can be tuned, depending on whether the emitter couples to a radiative or non-radiative plasmon mode [3]. This approach of tunable fluorescence quenching and enhancement finds its uses in applications such as molecular imaging [4]. A number of geometries have been explored for such decay rate engineering, for instance metallic planar surface [5], photonic crystals [6] and various collections of metal nanoparticles [3]. In particular, collections of nanoparticles, such as dimers provide an additional parameter, namely, separation and orientation of the individual particles with respect to each other to tune the local electromagnetic density of states. In certain dimer systems with inversion symmetry, symmetric or antisymmetric dipole modes can be excited, even using a plane wave, in accordance with the plasmon hybridization model. The antisymmetric or dark mode has a cancellation of the two induced electric dipole moments hence suppresses far field radiation. The converse is true for the symmetric or bright mode where the electric dipole moments add up constructively. Fig. 1. Geometry for studying decay rate engineering: A) In this geometry, the emitter can excite only one of the two modes, depending on its location and polarization B) In this geometry, the emitter can excite both the dark as well as the bright modes. Hence it is suited to studying comparatively, the effect of these modes on the the decay rate of the quantum emitter.
The numerical values of the disc separation D and the radius R, which were used for the BEM simulation are shown here.
Metals, as plasmonic materials, however suffer from a number of drawbacks, such as high losses [7] and limited tunability of electronic carrier concentration. In recent years, graphene [8] has emerged as a very efficient plasmonic material in the far infrared and terahertz range [9]. Because of its unique bandstructure, electrons in graphene behave as Dirac fermions. A consequence of this is that backscattering of electrons from impurities is forbidden [10], which results in graphene plasmons being much less lossy in the far infrared compared to metals in the visible range. In addition to chemical doping, the carrier density or equivalently, the Fermi level in graphene can be tuned via electrostatic gating [11]. Thus graphene is an excellent candidate for tuning light-matter interaction in this wavelength range.
In this work, we discuss how decay rate can be engineered via plasmon modes in a dimer of vertically stacked graphene nanodiscs. Plasmons in a single Graphene disc have been shown to provide very high total decay rate enhancements [12]. A dimer system of nanodiscs, while still having these advantages, provides a route to engineer radiative decay rates via excitation of dark and bright modes. Such bright dipolar modes have recently been experimentally observed in the case of graphene micro-disc dimers [13].
We will firstly solve for the eigenmodes of the systems using a quasistatic approximation. Secondly, a general recipe using the eigenresponse theory will be provided which can be used to model the spatial and polarization dependence of the local density of states. Some of these results for the lowest dipolar mode will be compared with full-wave boundary element simulations.
Finally we use the example of the lowest dipolar mode to show that the fluorescence quantum yield can be tuned by modifying the Fermi level of the graphene nanodiscs and the possibility of obtaining vacuum Rabi splitting in the cavity.
Methods
In this section, we briefly mention the various analytical and numerical techniques that were employed to arrive at the results in this paper.
The first term in Eq. (1) represents intraband contribution and the remaining terms are contributions of the interband transitions to the total graphene conductivity. Here τ is the electron relaxation time. The relaxation time typically has contributions from 1) scattering from impurities in infinite graphene, 2) coupling to acoustic and optical phonons (hω OPh = 0.2 eV) in graphene and phonon modes of polar substrates and 3) edge scattering in the case of finite nanostructures, such as the one discussed in the present paper( [15] and references therein). In literature, relaxation times as high as 1000 fs have been reported [16,17]. For frequencies larger than the optical phonon frequency of graphene, τ ∼ 50 fs [9]. For the results of this paper, we use the conservative lower end of this range. For all the results concerning excitation of the disc modes via local emitters, we use τ = 50 fs and T = 300 K. For the plane-wave excitation result, a larger τ of 100 fs is used since for smaller values, the extinction peak for the dark mode is too broad to be separated out from the background, dominated by the bright mode. However, both these values we used for τ are on the very conservative side of the range of experimentally measured values.
Simulation of graphene plasmon modes
All the simulations of this paper were performed using a Boundary Element Method (BEM) code, SCUFF-EM suite [18], a free, open-source software implementation of the boundaryelement method that implements specialized algorithms for efficient computation of scattered and absorbed power in scattering problems [19]. For simulations we consider a very small thickness "effective graphene" [20]. This now becomes a 3D structure, whose conductivity is given by dividing the 2D conductivity by the thickness of the effective graphene. This allows us to define a permittivity for effective graphene using Maxwells Equations: where ∆ is the thickness of the effective graphene. Convergence tests were performed with ∆ as a parameter and a value of ∆ = 0.25 nm was chosen as the appropriate thickness for the specific range of frequencies and lengthscales of our problem.
Calculation of decay rates
In the second half of the paper, we will discuss the decay rates of emitters placed close to or inside the nanodisc cavity. An ideal dipole emitter can get rid of its energy through two pathways: 1) radiatively into free space propagating modes or 2) non-radiatively into material absorption. The decay rate into the plasmon mode is mostly dominated by absorption. We calculate these decay rates as follows. The total decay rate is calculated using the scattered electric field at the location of the emitter [21]: where Γ 0 is the decay rate of the emitter, if it were in free space. The radiative decay rate is given by Γ rad /Γ 0 = P rad /P 0 = 1 + P sca /P 0 , where P rad is the power radiated to the far field, P sca is the total scattered power and P 0 is the power radiated by the emitter when placed in free-space. The non-radiative decay rate, which is the dominant contribution from the decay into the plasmon mode is given by Γ abs /Γ 0 = P abs /P 0 ≈ Γ plasmon /Γ 0 , where P abs is the power absorbed in the graphene nanostructure.
Vacuum Rabi splitting calculation: The most common way for describing atom-cavity interaction quantum mechanically is through the Jaynes-Cummings Model (JCM) Hamiltonian. The JCM Hamiltonian, in the rotating wave approximation (RWA) is given by: where ω 0 is resonant frequency of the quantum emitter and ω is the frequency of the plasmon mode. σ + (σ − ) are the atomic raising (lowering) operators and a † (a) are the creation (annihilation) operators for a cavity photon.
For the present problem we use an open quantum system approach, in order to incorporate absorption and radiative decay. Thus the evolution of the density matrix is given by [23]: where κ is the rate of decay of the plasmon mode. κ contains both the radiation as well and absorption mechanisms for broadening the plasmon resonance [24]. However, it is usually dominated by absorption. Γ ′ is the decay rate of the quantum emitter into free space modes, modified by geometrical effects. However, in this paper we assume the Γ ′ to be equal to Γ 0 , the spontaneous emission rate in free space, in accordance with Wigner-Weisskopf theory. The system density operator evolves according to Eq. (5). In the single excitation manifold, only the states {|g |1 , |e |0 , |g |0 } need to be retained. Here |g and |e are the ground and excited states of the atom and |0 and |1 denote the number of photons in the cavity mode. It can then be shown [25] that for Rabi oscillations to exist, on resonance (∆ = 0), one needs the condition |g/(κ − Γ ′ )| > 1/2.
In the JCM, the coupling strength g is determined by the details of the cavity field mode and the atomic dipole matrix element. For our purpose, we determine g classically, using the limit of a low finesse cavity. In this limit, g satisfies the following equation: where ∆ is the detuning between the resonant frequency of the plasmon mode and that of the quantum emitter. For the present work, the typical spontaneous emission rate of the emitter is much smaller than the cavity line-width. Thus, Eq. (6) suggests that on resonance, Γ tot /κ = (g/κ) 2 . Hence the g factor can be determined. This expression also points out that Rabi oscillations should exist when Γ tot /κ > 1/4.
Calculation of the eigen-modes in the quasistatic limit
Firstly, we discuss the mathematical formulation of the eigenvalue equation in terms of an electrostatic potential. This is essentially a solution of the Laplace equation for the disc geometry.
This section is divided into two subsections. In the first subsection, we repeat the derivation for the case of single disc, which had been worked out by Fetter [26] in 1986. In the second subsection, we formulate the eigenvalue problem for the case of a stack of two discs.
In the second section, we discuss a numerical framework to solve the eigenvalue equations we obtained in the previous section. In the third section, we summarize the results of the calculation, by providing details of the eigen-mode plots and a comparison to full wave boundary element simulations.
Mathematical framework for a stacked dimer of discs
The general strategy is to solve the Poisson equation for the electrostatic potential Φ(r), with the surface charge boundary condition due to the graphene discs. The surface charge density itself can be related to the electrostatic potential via the continuity equation and the surface conductivity of graphene. This leads to an eigenvalue equation with Φ on both sides. Subsequently, numerical techniques are used to solve this eigenvalue problem to get the resonant frequencies as well as the potential profile. This potential can then be used to calculate various other quantities of interest such as surface polarization and surface current density.
The calculation for the single disc case, was carried out in the Fetter's paper [26] on magnetoplasmons in disk geometries of 2DEG. For the sake of comparison the notation from a recent paper [27] on edge plasmons in a single graphene disc, has been used here.
It should also be noted that for a single disc in the quasistatic regime, closed form solution is possible [28]. However, we use a numerical approach here which can be easily extended to stacks consisting of arbitrary number of discs, where closed form solution becomes very cumbersome.
The geometry consists of two identical discs each of radius R, stacked vertically with a separation D in between (see Fig. 1). The location of the discs in our chosen coordinate system is z = ±D/2. The approach that will be presented here can easily incorporate the case where the two discs are non-identical. However, for the sake of clarity for our specific case, we will only consider identical discs for now.
First note that because of circular symmetry, the potential can be expressed as Φ(r) = Φ(r, z)e iLφ , in cylindrical coordinates.
We follow a two step procedure to get to the eigenvalue equation: • Express the surface potential Φ(r, z = ±D/2) in terms of the surface charge density σ b , using the Laplace equation and the normal electric field boundary condition • Express surface charge density σ b in terms of the surface potential Φ(r, z = ±D/2), using the continuity equation and the current-field relation Expressing Φ in terms of σ b : The Poisson equation in this case is given by: where σ b is the surface charge density (and not the surface conductivity, which is represented by σ ). One way of solving such problems is to write the general form of the solution in the regions on either side of the the boundaries and then match the boundary conditions. We will use this approach.
Thus we write the solution for the Laplace equation for z = 0 and then use the boundary condition for the normal electric field. To be specific, these equations are given below: where ε m is the relative permittivity of the medium in between the discs. Note that the e iLφ dependence was suppressed in the boundary condition equation. (Note that there is another boundary condition which is the continuity of the potential across z = ±D/2.) Now let us express Φ in terms of its Hankel transform component: where the Hankel transform is only taken in the radial coordinate of the cylindrical system. Now we substitute Eq. (11) into Eq. (8). After some manipulation and using Bessel's differential equation, we obtain the following simplified form: For z = ±D/2, Eq. (8) holds for the potential Φ(r) in real coordinates. Equivalently, for z = ±D/2, Eq. (12) holds for the Hankel transformed potential. We can write down the form of the solution in the three different regions as follows: There are four unknowns We also have four equations, two for the continuity of the potential across the discs and the other two for the normal electric field boundary condition.
It is quite straightforward to solve for the general case of different relative permittivities. In the following, we choose ε u = ε m = ε d = ε for simplicity. Solving the above linear system of equations, we get the solution for the Hankel-transformed potential on the discs: Now we go to real space, by taking the inverse Hankel transform on each side of the above equation. For brevity, we denote the Hankel transform operator asĤ(p; r ′ ) = ∞ 0 dr r ′ J L (pr ′ ) and its inverse operator asĤ −1 (r; p) = ∞ 0 d p pJ L (pr). With this notation, we can express the real space solution as: Expressing σ b in terms of Φ: There are two equations that we need to express σ b in terms of Φ. One is the continuity equation for surface current density and the other is the relation between surface current density and the electric field. These equations are given below: Now using the relation E || = −∇ || (Φ(r, z = 0)e ıLφ ), we arrive at the relation: As the last step, we move to normalized coordinates x → r/R and express the final equation in operator form: where η = σ (ω)/ıωε 0 εR. The reader is reminded that hereÎ uu L (x; x ′ ) andÎ dd L (x; x ′ ) are associated with the on-site term for the upper and lower disc respectively, whereas the offdiagonal terms represent the interaction between the two discs. In terms of normalized coordinates, Φ(r, z = ±D/2) → φ u,d (x)e ıLφ . Eq. (20) is an eigenvalue problem in the parameter η = σ (ω)/(iωε 0 εR), which can be related to the resonant frequencies of the modes. Note that in the normalized coordinates, the exponential term in the off-diagonal kernel depends on the ratio D/R instead of just D.
The solution to Eq. (20), will give us the resonant frequencies and the mode-profiles of the plasmons in stacked dimer of graphene discs.
It should be noted that this kind of approach is easily extensible to more than two discs or discs with different radii or surface conductivities.
We solve the eigenvalue problems for the single disc and the stacked dimer of discs case using the standard method of polynomial expansion. The results for the eigen-frequencies are shown in Fig. 2.
Comparison with full wave simulation
We compared the resonant frequencies obtained using the quasi-static solution to those obtained using a full-wave boundary element simulation (BEM). For this comparison we only choose the L = 1, n = 1 mode since that is the mode that we will be concerned with in the rest of the paper, when talking about photon emission rate engineering. It should also be noted that for the simulations, we use a realistic absorption in the graphene conductivity. The comparison is presented in Fig.2. Figure 2 suggests that there is a good overall match between the resonant frequencies found from the BEM and the quasistatic result. The resonant frequencies in the simulation were obtained from the LDOS spectrum.
Two features in Fig. 2 are worth highlighting. Firstly, the resonant frequencies of both the modes increase with E F . This is due to the fact that increasing E F results in an increase in the carrier density, which in turn causes an increase in the restoring force. This explanation is similar to how the plasma frequency in noble metals increases with carrier concentration. Secondly, the frequency splitting ω B − ω D increases with E F . This is due to the fact that at higher E F , the plasmon modes of individual discs are more leaky. This results in the interaction between the two discs being even stronger, resulting in a larger splitting.
Eigen-response theoretic framework and calculation of overlap
To understand dependence of the spatial and polarization dependence of the LDOS, we resort to an eigen-response theory [29]. In the present case of a system with a symmetric and an antisymmetric mode, the polarization density can be expressed as: where α A,L and α S,L are the eigen-polarizabilities of the antisymmetric and symmetric modes and P A,L and P S,L are the eigenmodes. The excitability of the modes is related to the overlap terms, which are the inner products of the mode profile and the excitation.
Relation between overlap and LDOS: The spatial dependence of the LDOS is contained in the overlap term, since the eigen-polarizability is usually only frequency dependent. In general the projected LDOS in Eq. (3) can be written as: where |i is a quantity proportional to the surface polarization, for each mode. In the following section we present the calculation of the surface polarization, which will help us calculate these overlap terms. Calculation of surface polarization: From our quasistatic approach, we determined the surface potential on both the discs. Using the surface potential, it is straightforward to obtain the surface polarization. The surface polarization P s can be related to the potential on the discs in the following way. In the absence of magnetization, the surface current J s can be related to P s as follows: J s can also related to the electrostatic potential using the relation: Thus we have the relation between P s and Φ: The potential Φ obtained by solving the eigenvalue equation as mentioned in an earlier note, can be plugged in Eq. (25) to calculate P s and subsequently the overlap terms. We will consider the source dipole generating a field which excite various infinitesimal dipoles on the disc surface. For this purpose, we will need the Green's tensor which is defined as [21]:Ĝ In order to keep the mathematical framework completely general, we consider the geometry shown in Fig. 1. For this geometry, r 0 = X dx + Y dŷ + Z dẑ and the location of the infinitesimal dipoles r = xx + yŷ + zẑ. The overlap terms for the case of a single disc are shown in Fig. 3.
Decay rate engineering
Single graphene nanodiscs, have been shown to provide very high Purcell factors [12]. Such high enhancement factors are possible due to the very small plasmon mode volume, which is a general characteristic of graphene films and nanostructures in the far infrared and terahertz range [22]. Using a vertical nanodisc dimer cavity, should provide an additional degree of freedom for engineering the decay rates. For instance, other than applying gate voltage or changing the radius of the discs, there is now an additional parameter, which is the separation of the discs, that can be used to tune the resonances [30]. The dark dipolar plasmon mode only weakly couples to plane waves hence its excitation using plane wave is not very efficient. However, if we use local emitters, such as quantum dots, it is indeed possible to excite the dark mode very efficiently [30]. The coupling of quantum emitters to such plasmon modes is reflected in the modification of the decay rates of the former. Depending on the nature of the plasmon mode, radiative decay of an emitter can be suppressed (quenching) or enhanced. In the case of graphene nanodisc dimer cavities, one can easily modify the radiative or non-radiative processes, by using gate voltage or disc separation to tune the dark or the bright mode to be resonant to the quantum dot. In this context, location and polarization of the emitter is another variable [31], which will be discussed in this work.
We consider two situations (Fig. 1) for studying the coupling between the emitter and the plasmon modes. Firstly, we study the case when the emitter placed inside the cavity. If the Fig. 3. Semi-analytically calculated overlap terms for a Single Disc: The emitter is located at a vertical distance of 15nm above the disc and moves along Y d = 0. The colors correspond to different polarizations of the emitter:x(-),ŷ(-) andẑ (-) emitters are located at the inversion plane then it allows us to excite either the dark or bright plasmon mode, depending on the emitter polarization and position. We will use the eigenresponse theory to calculate the overlap terms as a function of the position and polarization of the emitter. We will also present a comparison with LDOS for the lowest dipolar mode calculated via BEM simulation and show that the shape of the LDOS spectrum can be well explained by the calculated overlap terms. Secondly, we place the emitter outside the cavity, resulting in it being able to excite both the modes, for the same location and orientation of the emitter's dipole moment vector. This enables us to directly compare the bright and dark modes, in terms of their efficiency in suppressing or enhancing the decay pathways and quantum efficiency of the emitter, under similar excitation conditions. To avoid repetition, in this section we will only present simulation results for the radiative decay rates.
Case A: Emitters inside the stacked disc cavity
In principle, one can calculate the decay rates for all emitter positions inside the cavity. A randomly oriented emitter should then couple to both the dark as well as a bright mode. However, if the quantum dots are placed inside the cavity, on the inversion symmetry plane parallel to the plane of the disks, then depending of the alignment of the dipole matrix element, the quantum dot will only be able to couple to either the dark or bright mode, but not both. Hence in order to understand the physics of the problem, this is a convenient choice of emitter location.
Position and polarization dependence of the LDOS: Having worked out the matrix elements for the case of the single disc, we now move on to consider the cases of the dimer of discs. Because of mirror symmetry we can predict that there are even and odd modes. In literature, these are often called the bright and dark modes respectively. If we label the upper disc by U and the lower disc by L, then the following relations hold: • Bright mode: dp U = dp L , z U = −z L = d/2 • Dark mode: dp U = −dp L , z U = −z L = d/2 Now to calculate the total overlap for the two modes, we add the contributions from the two discs, taking into account the appropriate sign changes as mentioned in the relations given above: Bright Mode: Since the sign of the infinitesimal dipole moment does not change, we only need to consider the sign change in the z coordinate of the two discs. This results in the total sum of the overlap term for the z = ±d/2 giving a zero for the z−polarization of the emitter. The other two terms for the x and y polarizations survive and are basically twice of the corresponding value for the single disc case. Dark Mode: In this case, the sign of the infinitesimal dipole moment does change, and so does the sign change in the z coordinate of the two discs. This results in the total sum of the overlap term for the z = ±d/2 giving a zero for the x− and y−polarizations of the emitter. The only nonzero term is the one for the z−polarizations and is just twice of the corresponding value for the single disc case. The overlap terms are calculated by evaluating integrals of the form dp * mode · E exc , as discussed earlier. The calculated overlap terms are presented in Fig. 4 for the bright mode and Fig. 5 for the dark mode.
To verify our approach, we perform BEM simulations for a dimer of graphene discs, each 100 nm in diameter and separated vertically by 30 nm. The frequency range for the simulation is chosen to highlight the contribution of the lowest dipolar mode. The semi-analytical calculations using the eigen-response theory gives the dips in the LDOS spectrum ( Fig. 4 for the Similarly for the bright mode, there is a contribution from the z-polarized emitter. This effect is due contributions from neighbouring resonances. in fact this effect can be easily taken in to account in the eigen-response theory, if we include losses in the eigen-polarizabilities. These calculations will be published elsewhere. In the next section, we point out how radiative decay rates can be engineered using these dark and bright modes. In order to compare quantitatively, the effect of the dark and bright modes on the quantum emitter decay rate, we now study a geometry in which the emitter can excite both modes. Any location of the emitter other than the inversion plane is a valid choice. However, for simplicity, we chose to consider the emitter located outside the cavity as shown in Fig. 1. This might also be a convenient scheme, in as far as experimental realization is concerned. We can study all three orientations of the emitter dipole moment as done in the previous section. However, since in this section our main aim is to highlight the engineering of radiative decay rates, we will consider only one one polarization of the emitter, namely the x-polarization. To gain some qualitative insight into the excitability of the modes, we analytically calculate the overlap terms. Based on this calculation (not shown), we find that for an emitter close to the center X d = 0 nm, both the modes are excitable, with highest probability. Radiative decay rate engineering: We then performed simulations to calculate the radiative, non-radiative and total decay rates, when the quantum emitter, located at (X d , 0, Z d ) couples to either of the modes at their respective resonant frequencies. An example spectrum, for X d = 0 nm is shown in Fig. 7(a). In this case, the total decay rate enhancement, which is close to the non-radiative part, is almost the same for the two modes. This is consistent with our analytical calculation of the overlap terms which show that at X d = 0 nm both the modes are equally excitable. However, there is a difference between the radiative decay rate enhancement. This situation results because of partial cancellation of the induced moments on two discs, for the dark mode.
Further, we demonstrate the tunability of the radiative efficiency as a function of carrier concentration. Our calculations in Fig. 7(b) show that the contrast in radiative efficiency increases with increasing E F . The dark mode becomes less, and the bright mode, more radiative, as the carrier concentration increases. As mentioned before, this is because at higher carrier concentrations, the modes of the two discs can interact more strongly.
We also study the dependence of the radiative efficiency, as a function of position X d of the emitter, in Fig. 8(a). Clearly, the bright mode has a higher radiative efficiency compared to the . Z d is fixed such that the emitter is located 15 nm above the closest disc. The bright mode is found to enhance radiative decay rate, but the dark mode does not. b) Vacuum Rabi splitting versus Fermi level: Normalized vacuum Rabi splitting for the dark and bright modes when an x-polarized emitter is located at (0, 0, Z d ) outside the cavity. Z d is fixed such that the emitter is located 15 nm above the closest disc. The dashed line represents g/κ = 1/2 below which splitting will not be observed. (κ ≈10 meV).
dark mode, for various locations of the emitter. We find that the radiative efficiency drops from a maximum at X d = 0 nm to a local minimum, as the emitter approaches a certain horizontal distance ( X d = 40 nm in the specific case of Fig. 8(a)) and then rises again. This behaviour can be qualitatively explained by looking at the overlap term (not shown), which predicts that the mode excitability being maximum at X d = 0 nm, drops to zero at a certain X d to subsequently rise again. Similar pattern in observed for the dark mode.
In the analysis provided here, one must note that the dominant decay pathway for the emitter in this geometry is still non-radiative. Hence the overall quantum efficiency when the emitter couples to either the dark or the bright mode, is rather small for both cases.
Strong coupling regime: Strong coupling regime in a system composed of a single graphene disc and a quantum emitter was predicted in [12]. Here we briefly mention that the same can be obtained in the dimer system as well. However, so far we have not carried out a complete study in this direction.
In our classical electrodynamic simulations carried out for the graphene nanodisc cavity, we obtain Γ tot ≈ Γ plasmon > κ on resonance. This condition ensures the existence of coherent coupling between the plasmon and the emitter and the possibility of observing vacuum Rabi splitting [23].
Using the approach mentioned in the Methods section, we calculated g factors for different Fermi-level energies, ranging from 0.2 − 0.8 eV, for different positions of the emitter both inside and outside the cavity. Our calculations for the normalized Rabi splitting are shown in Fig. 8(b), for an example case of the emitter placed outside the cavity (same geometr y as considered earlier in this section). We find that for a range of values of the doping level, we do obtain g/κ > 1/2 and we predict Rabi splitting values g of up to 10 meV at room temperature. Here for calculating the absolute value of g, we have used Γ 0 ≈ 5 × 10 7 s −1 [12]. We note that it should possible to increase the vacuum Rabi splitting by using higher quality graphene samples, so that τ is larger. Another important trend suggested by Fig. 8(b) is that the value of g/κ decreases with increasing E F . This can be qualitatively understood as resulting from the increasing leakiness of the plasmon mode as the carrier concentration is increased. This results in a larger mode volume V , which is expected to cause a decrease in the decay rate Γ tot . Since κ ∼ 1/τ is almost independent of Fermi Level E F , therefore g/κ ≈ Γ tot /κ decreases with increasing E F .
Conclusion
We have performed a complete study of the plasmon modes of a vertically stacked dimer of graphene discs. The eigenmodes and resonant frequencies of the modes were calculated semianalytically using a quasistatic approximation. We showed a convincing match between full wave BEM simulations and the quasistatic approach. Secondly we explained the position and polarization dependence of the LDOS using the framework of eigen-response theory. In this case also, results were consistent with simulations.
Subsequently, we focused on the dark and bright plasmon modes formed out of dipolar modes of each constituent disc of vertically stacked graphene disc dimer cavity. Due to the different symmetry of these two modes, completely different behaviour is observed in the far field response as well as decay rates. We have shown that it is possible to engineer the decay rates of a quantum emitter placed inside and near this cavity, using Fermi level tuning, via gate voltages and variation of emitter location and polarization. We highlighted that by coupling to the bright plasmon mode, the radiative efficiency of the emitter can be enhanced compared to the single graphene disc case, whereas the dark plasmon mode suppresses the radiative efficiency. Such a system can offer new degrees of freedom for controlling radiative and non-radiative emission properties of quantum emitters. | 8,099 | 2013-09-09T00:00:00.000 | [
"Physics"
] |
Design of Rigid Pavement by Self Cured Concrete Utilizing Coarse Fly Ash Aggregates and Curing Admixture
In concrete industry, a huge amount of natural aggregates is used in the making of concrete every day. The environment is being exploited by mining for the gain of natural aggregates, resulting in an environmental instability in nature. As a result, an alternate source to substitute natural aggregates in concrete is required. A lot of waste materials have gain attention now a days into the concrete industry as a substitute to natural materials. Fly ash, a waste product of thermal power plants, meets the criterion for being utilised as an aggregate substitute in concrete because of its pozzolanic activity. Coarse fly ash is manufactured using a good manufacturing method and is light in weight. Keeping this into view, the impact of partial replacement of natural coarse aggregates with coarse fly ash aggregates produced using the colds bonded method is explored in this paper. The major focus of this study is on testing for flexural strength of self-cured concrete, as flexural strength is a key criterion for rigid pavement design. In this study, coarse fly ash aggregates are utilised in concrete in different proportions to substitute natural aggregates, and the optimal value for flexural strength is determined using a curing additive. The findings of this experiment indicated that when fly ash aggregates and curing additives were used optimally, the flexure strength improved, which is enough for the construction of rigid pavement as criteria fixed by Indian Standards.
Introduction
The use of concrete in the construction industry has escalated dramatically in recent years due to its long durability. However, its production is hazardous to the environment and depletes natural resources. On the other hand, the conventional concrete curing process is still inadequate for cement hydration. For skyscraper structure and mass concrete construction, curing is very difficult. Wastage of water is there for conventional curing process. To overcome these problems, concrete with self curing and waste utilization will be the appropriate answer [1][2][3]. Self curing or internal curing is a process in which the water is feed by the water reservoir inside the concrete in form of light weight aggregates. Aggregates manufactured by cold bonding technique utilising fly ash are porous and absorb high water quantity, which may be employed for self curing [4,5].
The influence of utilization of fly ash in concrete has been explored by different researchers for the strength characteristic of concrete. From the previous literature study, it different curing admixture have been tested in concrete. The durability of concrete by using different curing admixture of content having high fly ash content have been used. Fifty percent of total cement was replaced with fly ash, and spinacia oleracea, polyethylene glycol was used for internal curing. When 50 percent of the cement is replaced with fly ash, the strength was significantly reduced at 7 and 14 days. However, due to the pozzolanic effect of fly ash [6][7][8], significant strength was acquired after 56 days. Self cured concrete showed greater durability against the action of acids as compared to conventional concrete . It was revealed that by using sintered fly ash aggregates concrete having strength in between 23 MPa to 74MPa can be produced. Sintered fly ash aggregates have a superior interfacial transition zone in terms of thickness and quality, and the pozzolanic reactivity between the aggregates and paste increases as the temperature rises . It was found that the compressive strength of geopolymer concrete based on low calcium fly ash increases by up to 20%, when OPC was used as a fly ash replacement. The inclusion of calcium-based products resulting in the hydration process, which coexisted with the alumina silicate polymer product, enhanced the strength . It was revealed that concrete mix with a water cement ratio of 0.44 using pre wetted vacuum saturated lightweight aggregate, the strength enhanced after 365 day. At a volume replacement of 8.9 to 13.8 percent, lightweight aggregates were employed to replace standard weight aggregates, and according to the findings of study all lightweight aggregates concrete combinations shrunk less than normal conventional concrete . It was observed that in self compacting concrete by using cold bonded fly as substitute fine aggregates in varying percentages increases the workably of concrete and a slump of 25 cm was found. The compressive strength of fly ash aggregates was low at first due to their reduced specific gravity of aggregates, but as time went on, the compressive strength rose due to the pozzolanic activity of fly ash . It was revealed that self-curing concrete mixes have a greater water retention rate, as indicated by weight loss over time [9][10][11]. Under sealed conditions, self-curing concrete experienced less self-desiccation than traditional concrete and water absorption and water permeability values for self-curing concrete declines with age as a consequence of the cement's continuous hydration, implying a reduced proportion of permeable pores . It was found that concrete shrinkage and mechanical qualities are affected by self-curing process done by water-saturated fine aggregates and admixture. Zeolite is a reactive mineral additive having certain ion-exchange ability in concrete and enhanced the cement hydration . It was revealed that adding fly ash (10%) and superplasticizer (0.25-0.35%) to concrete improves workability and marginally enhances compressive strength. Concrete produced was fulfilling the requirements of self compacting concrete [12][13][14][15].
Polyethylene glycol-400 was used as the curing admixture by many researchers. By using three different curing technique water ponding, curing through wax based curing compounds and self curing through PEG-600, it was revealed that at 28 days curing trough wax compressive strength was 91% of that cured in water tank. Strength for self cured cube was 95% of that cured in water tank . It was observed that with increase in percentage of PEG-400 (0%,0.5%, 1%, and 1.5%) the workability also increased, and the strength was also high for self cured concrete . It was found that by incorporating super absorbent polymer at 0.05-0.2%, PEG at 2-4% and light weight aggregate at 15 % dosage in concrete showed the same strength as normal concrete . It was revealed that the self compacting concrete prepared by replacing up to 60% cement with fly had a compressive strength of 40N/mm 2 . With increase in fly ash content in concrete the water absorption of concrete also increased which reduced the durability of concrete . It was revealed that for architectural concrete when 1.5 percent polyethylene glycol and 2 percent steel fibre were added, the compressive strength improved by 4.13%and split tensile strength got improved by 4.49% . It was found that by using steel fibre, polypropylene fibre and glass fibre by 1% volume fraction in pavement concrete, the flexure strength obtained by using steel fibre was almost double as compared to other fibres. By using fibre reinforced concrete, the design thickness can be lowered . It was revealed that the stresses computed in pavement concrete using IRC techniques are greater than those calculated using PCA methods. The stresses were found to be 40 percent to 50 percent greater for 8 tonne loads, and around 22 percent higher for 16 tonne tandem axle loads. From the past study it was revealed that, maximum stresses occur in edge region in case of finite element method, for empirical method at corner region [16,17].
Cement
In this study, Ordinary Portland Cement 43 grade was used confirming to IS: 269-1989 for casting of concrete. The cement was bought from local material market and stored in air tight bag in laboratory. The properties of cement are mentioned in Table 1. Table 1. Physical properties and composition of ordinary Portland cement
Natural Aggregates
Fine aggregate and coarse aggregates procured from local stone crusher are utilised for this work. The fineness module for fine aggregates is 2.90. The size of coarse aggregates used in the study is 10-20mm.
Fly ash aggregates
Fly ash aggregates manufactured by cold bonded process are utilised for research purpose having size 10-20 mm. A fly ash brick made by cold bonded technique was broken into small pieces to procure coarse fly ash aggregates, see Table 2.
Polyethylene glycol
Polyethylene glycol was bought from online shopping market. The atomic weight of polyethylene glycol is 400 (PEG-400) and the density of PEG-400 is 1.26 gm/ml at 4 0 C for and viscosity is 85-105 cm 2 s -1 at 20 o C as supplied by the manufacturer.
Water
Potable water was used for the construction and curing of concrete.
Batching and mixing
Using a weighing machine, the materials were precisely measured. For the ingredients, weight batching was used. Hand mixing was used to ensure that all of the concrete materials were well mixed. The coarse aggregates and fine aggregates were thoroughly mixed first, and then the cement was added. The water was then carefully added so that no water was lost during the mixing process.
Physical properties
Fineness (
Concrete Mix Design
A M40 concrete mix design was done according to IS: 10262-2019. Coarse aggregates were replaced by fly ash aggregates by different proportion (5%, 10%, 15%, 20%, 25%, 30%) by weight. Water cement ratio used was 0.38 and Polycarboxlic ether based superplasticizer was used to attain workability. Beams of size 100mm 100mm 500mm was casted and tested for the flexural strength. The different concrete proportion are shown in Table 3. The concrete mix are designated with a code where "M" represents the standard concrete and "FM" represents the fly ash concrete and the number represents the amount of aggregates replacement. Table 3. Concrete mix design
Curing of samples
Concrete moulds were left for curing for a period of 24 hours so as to gain proper strength and then demoulded. The specimens were properly cleaned and were placed in curing tank for the period of testing. The samples prepared by self curing concrete were covered with plastic sheets in laboratory at room temperature.
Flexural Strength 3.1.1 Effect of fly ash aggregates on flexural strength
The natural coarse aggregates were replaced by fly ash aggregates in varying percentages by weight to check their efficacy on strength characteristics. The flexure strength of normal concrete mix was 4.49 N/mm 2 after the curing period of 28 days, on replacing natural coarse aggregates with fly ash aggregates in varying amount, the flexural strength increased. On replacing natural coarse aggregate with 15% fly ash the compressive strength was 4.72 N/mm 2 which was the highest. Fly ash aggregates concrete exhibited high strength than normal concrete at later ages. Thus 15% replacement with fly ash was considered as optimum replacement for strength. This may be due to the reason that the fly ash has pozzolanic properties as studied by various researcher [16]
Design of rigid pavement
The design of the rigid pavement was done according to IRC: 58-2015 and pavement is designed for self cured concrete with 1% PEG-400 and 15% fly ash aggregates.
Stresses in the Pavement Due to Wheel Load and temperature stresses
Stresses calculated by empirical formulas given by IRC: 58-2015 are tabulated in Table 4. Factor of Safety = =1.12 Therefore, the total stresses in the pavement concrete were less than the flexure strength, and the factor of safety is more than one, so the given thickness 25 cm is safe for the design. Table 5. Thus, the pavement designed can sustain these number of repetitions, when overloaded vehicles with axle loads more than the design load passes through pavement.
Conclusions
The following are the results of the research in which natural coarse aggregates were partially replaced with fly ash aggregates and PEG-400 was introduced to self-curing concrete: The strength of concrete on replacing natural coarse aggregate partially by fly ash aggregates improves on later ages; the maximum improvement occurred for 15% replacement with fly ash aggregate. The flexure strength of concrete on 15% replacement was 4.72 N/mm 2 at 28 days; which is 5.12% higher than normal concrete. The addition of 1% PEG-400 to 15% fly ash aggregates enhanced the flexural strength to 4.94 N/mm 2 ; which is 9.13% higher in comparison to normal concrete. The self-cured concrete of thickness 25 cm, prepared by the addition of 1% PEG-400 and 15% fly ash aggregates is safe for designing rigid pavements against all critical stresses with a design wheel load of 7000 Kg. | 2,860.2 | 2021-11-01T00:00:00.000 | [
"Materials Science",
"Engineering"
] |
Optimization Design and Development of Sensing Coil and Analog Signal Conditioning Electronics for Fluxgate Magnetometer Sensor
The design of fluxgate magnetometers is typically a nonlinear multi-objective optimization problem. Different objectives often conflict with each other, and sometimes an optimal Fluxgate Magnetometer Sensor (FMS) performance is difficult to achieve. The sensitivity of the sensor decreases with an increase of noise level while trying to reduce the sensor dimension. Hence, there is need for a systematic optimization approach for FMS design to find its optimum performance. The combined modified multi-objective Firefly Optimization Algorithm (FOA) and systematic optimization approach is suggested to improve FMS’s design in this research by simultaneously optimizing the sensitivity and noise of a FMS while the sensor core, pick-up coil, and detection circuit are minimized. The developed model allowed improved sensitivity of 86.65%, reduction of noise level by 59.97% while still keeping the sensor size small by 14.29%.
INTRODUCTION
FMS are commonly used magnetic field sensors for measuring DC or low frequency magnetic field vectors (Lu and Huang, 2015). FMS have very high sensitivity spans a wide range from 100 pT to 100 μT (Lv and Liu, 2013), low noise, small size, small power requirements, and high temperature stability (Frydrych et al., 2014). Moreover, the advancements in magnetic materials which form the heart of FMS and characterize their sensitivity, noise level, and linearity range responsible for their popularity among other competitive magnetic field sensors (Can and Topal, 2015). These make them one of the magnetic field sensors that still attract the attention of many researchers because of their wide applications (Todaro et al., 2012) in space research and navigation systems (Indrasari et al., 2012), particularly in Earth's magnetic field exploration surveys (Kim et al., 2013;Lv and Liu, 2014).
In order to optimize the performance of magnetometers, different optimization techniques for their structures and core materials had been developed. For example, the conventional approach was based on Part-by-Part Optimization (PPO) technique, which includes designing the sensor core first, then select the dimension of pick-up coil, and finally develop a low noise detection circuit. However, PPO technique is too slow, time consuming, and expensive (Grosz and Paperno, 2012). Another optimization technique for the magnetometer parameters proposed by Chen et al., (2011) and Grosz et al., (2011) was based on an analytical model, which was numerically solved to obtain improved large set of parameters such as volume and weight of pick-up coil, power consumption, and the noise of the signal conditioning circuit. However, the analytical optimization technique becomes unnecessarily complex when performing large number of numerical calculations to optimize the magnetometer, hence, introducing difficulty in interpreting the results obtained (Grosz and Paperno, 2012). Recently, optimization of Fluxgate Magnetometer Sensors (FMS) had been based on Finite Element Method (FEM) by using simulation software such as ANSYS, FEMM, Flux 2D, and others. On the other hand, most of these tools could not offer the users the ability to fully express their optimization purposes by formulating the objective functions (Kim et al., 2013).
Driving the excitation coil and detecting the pick-up coil signal require careful design of the excitation and detection electronics circuits (Baschirotto et al., 2010;Velasco et al., 2011). In most literatures, the excitation circuits for fluxgate sensors are typically based on a sinusoidal (Zorlu et al., 2010), triangular (Baschirotto et al., 2010) or pulsed excitation (Waheed and Rehman, 2011). The pulsed excitation is easier to generate than sinusoidal or triangular excitation (Ripka, 2001) and represents a trade-off between the sensitivity and power consumption (Cui, 2013). A pulsed excitation reduces power consumption at the expense of low sensitivity (Baschirotto et al., 2010;Cui, 2013).
The block diagram of the entire fluxgate magnetometer is shown in Figure 1. To periodically saturate the ferromagnetic material, the excitation signal is fed into the excitation coil of the fluxgate sensor with a square excitation current waveform with excitation frequency f. The pick-up coil of the FMS detects the signal by the rising and falling edges of core magnetizing current (see Figure 2.6(c)). The pick-up coil is fed into a pre-amp with mild tuning to the second harmonic (2f) of the excitation frequency because the second harmonic fluxgate magnetometers produce the highest sensitivity and the lowest noise (Lu and Huang, 2015). Therefore, it is possible to extract the information on the external magnetic field by a synchronous demodulation (Miles et al., 2013). Demodulation is usually accomplished with a phase sensitive detector, typically, a CMOS analogue switch following the pre-amp (Tumanski, 2013).
Figure 1: Typical Second Harmonic Demodulator Scheme
In this paper, the design of fluxgate magnetometers is assumed to be a nonlinear multi-objective optimization problem. Different objectives often conflict with each other, and sometimes optimal magnetometer performance is not achieved. Metaheuristic algorithms are very powerful in dealing with non-linear multiobjective optimization problem (Yang, 2013). The multi-objective Firefly Optimization Algorithm (FOA) was proposed in this research, because FOA is one of the nature-inspired metaheuristic algorithms, which is capable of handling the design problems in electromagnetics with a large number of design variables and multiple objectives under complex nonlinear constraints (Yang, 2013). For instance, sensitivity and noise of a fluxgate magnetometer can be improved while the sensor core, pick-up coil, and detection circuit are minimized. The combined multi-objective FOA and systematic optimization approach is suggested to improve FMS's design in this work by simultaneously finding the dimensions and geometry of the sensor core, pick-up coil, and detection circuit in order to reduce its noise and increase its sensitivity.
SYSTEM DESIGN 2.1Square Wave Current Generator
Frequency generator circuit was designed using different electronics components such as operational amplifier, transistors, hex inverters, and so on. The IC 4069 is a CMOS logic chip having six independent inverters (Fairchild, 2002). It was used for interfacing and to make simple square wave generators as shown in Figure 2.
Where F is the frequency in Hz, R 1 is the timing resistor in ohms and C 1 is the timing capacitor in Farad. As shown in Figure 2, the circuit used a few components such as two resistors with R 1 used as timing resistor and a capacitor C 1 and consumes less power. The output frequency produced by the square wave oscillator depends on two components R1 and C1. Changing the values of the resistor R1 and capacitor C1 will result in change of output frequency.
Voltage to Current Converter
The frequency divider output produce analog output voltage but the current was weak and cannot be directly used to drive the magnetic core to saturation. The oscillator and the frequency divider can only produce a few tens of milli-amperes at most, while the sensor core requires many amperes. Hence, there is need for current amplification by using power transistors. Therefore, a complementary emitter-follower is commonly used for efficient bipolar current amplification. Figure 3 shows a low-noise class-AB power amplifier using NPN and PNP transistors Q1 and Q2 respectively. The two transistors Q1 and Q2 were configured as a complementary emitter-follower. Class AB amplifier is very similar to class B amplifiers, but their performance is improved by the addition of two diodes that eliminate the crossover region and allows both transistors to be turned on at the same time. The efficiency (around 50%) is not as high as class B because both transistors are turned on simultaneously, but accuracy is improved. It is the most commonly used voltage to current converter amplifier.
Where R4 is a bias resistor and R3 is equal to R4. V cc is the supplied voltage, V be is the emitter-base voltage and I ref is the transistor base bias current. Resistors R1 and R2 are equal in values and set the operating current for the output of the transistors. The values of the emitter biasing resistors R1 and R2 are calculated as: Where R1 is an emitter bias resistor and R1 is equal to R2. V is the supplied voltage, I o is the output current, and I ref is the transistor base bias current. The two diodes serve to bias the transistors and reduce the cross-over distortion that occurs when the input waveform crosses zero. Without the diodes, the oscillator output would have to swing 1.4V to turn one transistor on and then bring the other transistor off. Connecting the excitation circuit to the excitation coil requires coupling capacitor that serves to isolate the AC signal from any DC bias voltages. In order for a coupling capacitor to operate effectively, it must have the right size. Capacitor C1 served to block the DC component of the current source from reaching the excitation resonant circuit. The equation to calculate the value of the coupling capacitor C 1 is: Where C is the capacitance in Farads, and f exc is the excitation signal frequency in Hertz. R coil is the impedance on the load side of the capacitor which in this case is the excitation coil resistance. As shown in Figure 3, when transistor Q1 is turned on, capacitor C3 is charged smoothly as the charging current is limited by inductor L1. The main function of the inductor L1 is to limit the current drawn from the source for the fluxgate excitation current. This was achieved by the use of a high impedance (larger than the Vol.10, No.4, 2019 fluxgate sensor) inductor, which operates in the non-saturated mode over part of the excitation period. In the non-saturated state, the high impedance of the inductor limits the current flowing from the source to the excitation circuit.
Detection Electronics
The pick-up coil of the fluxgate sensor detects the signal induced by the flux collapse (saturation) and flux recovery (de-saturation) of the core magnetizing current (Evans, 2006). This small induced voltage output signal of the pick-up coil was detected and it was compensated by amplifying and filtering the signal. The second harmonic component of the induced voltage across the pick-up coil was conditioned by using the electronics circuit shown in Figure 4.
Figure 4: Schematic Diagram of Detection Electronic
The output voltage from the pick-up coil is usually small compared to the reference voltage of the Analog to Digital Converter (ADC). A low noise operational amplifier is needed to amplify the output voltage of the pick-up coil. As shown in Figure 4, at very low frequencies (within the cut-off frequency or frequency response of the amplifier), the capacitor C2 is an open circuit and the gain of the signal conditioning circuit was high, which made it acted as an amplifier. In order to evaluate the performance of FMS associated with the detection circuit, the frequency response of FMS detection amplifier was computed as (Han et al., 2012;Chen et al., 2015): Where, " M = ! " H E ! " N , R w is the pick-up coil winding resistance, L w is the pick-up coil inductance, C f is the amplifier feedback capacitor, R f is the amplifier feedback resistor.
Power Supply
To provide a stable voltage to the excitation circuit, a voltage regulator circuit was used to maintain a stable supply voltage to excitation circuit. The power supply was used to provide all the voltages necessary for driving the various components along with reference voltages. This adds up to the regulated supply voltages of +5 V, ±10 V, and ground.
EXPERIMENTAL SETUP
Component values were optimized to improve performance and to minimize currents in the fluxgate magnetometer. Then the circuit was tested on a breadboard and soldered on a prototyping board. After an iterative process of design optimization, the excitation frequency selected was 5 kHz square wave. The schematic diagram of the complete excitation circuit is shown in Figure 5. Figure 5, in order to ensure proper saturation of the ferrite magnetic core material the whole circuit was driven by a square wave oscillator at 20 kHz frequency. A duty cycle of 50% on both the 5 kHz and the 10 kHz output waveforms was ensured. A duty cycle different from 50% could compromise the demodulation of the signals produced by the sensing coils and, hence, it has to be avoided. As shown in Figure 5, by using equation (1), the oscillator block was built around resistors R1, R2 and potentiometer VR1 with capacitor C1 and Integrated Circuit (IC), hex-inverter 4069BD (U1A & U1B). The oscillator circuit was tuned to twice the excitation frequency (2f exc ) and variable between 1 kHz and 20 kHz by means of a 100 kΩ potentiometer (VR1).
The completed ring cores with excitation coils wound circumferentially were put inside a pick-up coil bobbin, which were constructed to hold the pick-up coil as shown in Figure 6. Finally, the pick-up coil with 646 turns was wound diametrically on the core with copper wire having 0.2 mm. Figure 8) on the testing board. The tuned second harmonic output voltage responses of the developed FOA fluxgate sensors as a function of the applied external magnetic field were verified. As seen in Figure 9, Tektronix Digital Multimeter (2050 model) (with root mean square (rms) range selected) was connected to the pick-up coil output terminal (point 'c' shown in Figure 8) of the prototype sensors to measure the peak output voltages of the sensors corresponding to the external magnetic field supplied from the Helmholtz coil. Magnetic field strength meter (calibrated magnetometer) was used to evaluate the magnitude of the magnetic field supplied from the Helmholtz coil to the fluxgate sensors. This was done by placing the magnetic field strength meter (sensor) perpendicular to the magnetic field of the Helmholtz coil (Figure 9).
RESULTS
Based on the optimum values of the dimensions and geometric parameters obtained from the FOA design, FMS prototype was fabricated and studied. The sensor was made from MnZn ferrite with a thickness of 2 mm. The FOA was written in Matlab environment and the results were obtained by running the developed program on a 1.50 GHz Intel® core™ Duo CPU Windows 7 Ultimate 32-bit personal computer. The sensors were designed to withstand up to 99 mA current. Copper wire having 0.411 mm was wound as the excitation coil on each sensor with 49 turns. Parameters and the values of the FOA model are contained in Table 1. 1.80 kΩ The measurements of the output response to field variations were made with 15 μT as the minimum magnetic field up to 75 μT maximum with step of 10 μT. The 15 μT resolution of these measurements was limited by output capability of the Helmholtz coils at the minimum driving current. It was noticed that the minimum driving current for the Helmholtz coils was 8 mA. Below this minimum driving current (8 mA), the Helmholtz coils operation became unstable resulting in the fluctuations of the output magnetic field of the calibrated magnetometer used. In addition, the 75 μT maximum magnetic fields of these measurements were limited by the induced excitation field of ±67.99 μT inside the sensor core. This was validated by the principle of fluxgate magnetometer. The smaller the sensitivity the higher the magnetic field ranges of sensor. It was noticed that the magnetic field linearity of these sensors was increasing as the core dimension was decreasing. This was Innovative Systems Design and Engineering www.iiste.org ISSN 2222-1727 (Paper) ISSN 2222-2871 (Online) DOI: 10.7176/ISDE Vol.10, No.4, 2019 23 due to the increasing nature of the excitation current as the core dimension decreases which in turn increases the linearity of the sensor. Table 2 presents the tuned second harmonic output voltage obtained from the experimental measurements (Figure 9) of the pick-up coil output voltage. 86.45 This data shows that the FOA simulation routine provided an accurate geometric dimension of the sensor core and the pick-up coil. During the experimental data measurements, it was observed that the sensitivity increased as the core dimension decreased up to an optimum dimension of the core at which the voltage sensitivity began to decrease.
The plot shown on Figure 10 (obtained from Table 2) shows the response of the pick-up coils of the developed sensors when tuned to second harmonic of the excitation frequency under the imposed external magnetic field from Helmholtz coils. The FMS exhibited voltage sensitivity of 1139V/T at 75μT external magnetic field range and 5 kHz excitation frequency. Figure 11 shows the FMS responses to external magnetic field due to a bar magnet (scaled to 50µs per division on the horizontal axis and 550 mV per division on the vertical axis). These responses reflected both the influences from the applied field of a bar magnet and the remnant magnetic field. It was observed that when positive external magnetic field (North-pole of a bar magnet) was applied to the sensor, the amplitude of the pick-up coil output signal was increased and the values of magnetic field signature, S1 was higher than S2 ( Figure 11). Vol.10, No.4, 2019 S2 were 550 mV and 330 mV, respectively, while the value of the negative peaks in both S1 and S2 was 275 mV (Figure 11(a)). When negative external magnetic field (South-pole of a bar magnet) was applied to the sensor (Figure 11(b)), the amplitude of the pick-up coil output signal was also increased, but in the opposite direction and the values of magnetic field signature, S1 was smaller than S2 (Figure 11(b)). The values of the positive and negative peaks were different too. The value of the positive peaks in both S1 and S2 was 275 mV (Figure 11(a)), while the values of negative peaks in S1 and S2 were 330 mV and 550 mV, respectively. These observations could be due to the simultaneous decreased permeability of the hysteresis loop in both reverse and forward magnetization curves.
5.CONCLUSION
In this research activity, a miniature fluxgate magnetometer with magnetic ring core in a square cross section of 2 mm was realized. The sensor was wire-wound based on traditional technology process. The signal conditioning of the developed modified FOA fluxgate sensor was done by external electronics circuits. A square wave excitation current produced by a frequency oscillator circuit was fed to the excitation coil. The Helmholtz coils produced the external magnetic field to be measured. The second harmonic frequency of the induced voltage across the pick-up coil was measured with detection electronic circuit synchronized with the reference frequency from the square wave oscillator.
The introduction of the combined Firefly Optimization Algorithm (FOA) and the systematic optimization approach to FMS design problem in this research, by simultaneously finding the optimum dimensions and geometry of the sensor core, pick-up coil, and detection circuit, significantly improved the matching of the excitation and detection circuits. The developed optimized sensor for earth's magnetic field exploration showed good sensitivity of 97.09 mV/μT and linearity in the range of about ±49.44 μT. The power consumption of the sensor was 131 mW, the sensor electronics consumed 315 mW, while the whole sensor system consumed 446 mW. Hence, good sensitivity and the possibility of detecting magnetic field along two perpendicular directions make the developed FOA-based sensor suitable for portable compass application. | 4,398.8 | 2018-12-29T00:00:00.000 | [
"Physics"
] |
Does the geographic location influence takeovers?
Abstract The aim of this study is to examine whether there is an impact of geographical proximity, between the acquirer and the target company of a takeover process, on the success of the transaction. In order to do this we analysed a complete database of all the takeover bids between 2000 and 2014 on the Romanian capital market. The evidence reveals that not only is the geographical proximity important for the success of the transaction, but that also takeovers can occur across distant locations if the target firm has a long history of its activity that can be monitored by the acquiring companies. We have also tested if there are NUTS regions characteristics that can influence the decision-making, but the results were inconclusive.
Introduction
It is well known that mergers and acquisitions (M&As) are a way for there to be economic development of a city, region or even a state. They can lead to some important changes in a company's business and also can influence the economic activities in the target area.
In our point of view, Romania is a special case that needed attention. The state was under a communist regime for more than 50 years and the centralised economy followed a policy of industrialisation. Under the socialist beliefs that people must pursue the interest of everybody and everything is under common ownership, many socialist companies were founded regardless of the social necessities, and were equally distributed all over the state. Then, after Romania became a democracy, the state could not hold on to those non-performing companies and faced a massive privatisation. This is the beginning of the M&A activities in Romania.
In this paper, we examine if the geographical distance between the target and the acquiring company can play an important role in the decision-making of an M&A transaction. If two companies are in proximity, then there is a significant chance that they know each other well, what facilities they have, if the human capital is qualified or if the company is highly profitable. In other words an increase in the distance between the acquiring and the potential target reduces the chance of knowing each other and so reduces the likelihood of a merger between companies (Ensign, Lin, Chreim, & Persaud, 2014;Uysal, Kedia, & KEYWORDS takeover bids; geographical proximity; m&a decision; probability of success Panchapagesan, 2008). This finding was first noticed in the domestic M&As in the banking industry (DeLong, 2001;Houston, James, & Ryngaert, 2001), but there are studies that have concluded that cross-country transactions have the same trend (Eckbo & Thorburn, 2000). Many of the M&A transactions involve a company placed in important metropolitan areas that is willing to expand its economic activities to nearby areas. This can bring an improvement in productivity and economic performance due to the stimulation of knowledge spillovers (Böckerman & Lehto, 2006). Nowadays, we live in an information era, when we can find a great deal of information about any company anywhere in the world. Take, for example, the balance sheet and the income statement. They are very useful for knowing about a firm's physical assets, leverage or the value of its total sales and profit, but not as useful when we think about the firm's productivity, the R&D division or other internal issues. The geographical proximity can facilitate the transmission of some soft information that could be very hard to transmit and be interpreted by someone far away (Uysal et al., 2008;Kang & Kim, 2008;Chakrabarti & Mitchell, 2015).
The aim of this study is to investigate the patterns of takeovers in Romania in the last 15 years. The study contributes to the existing literature, first of all, by investigating the influence of the proximity between the target and the acquiring company on the success of the transaction not for a single M&A sector, but for the entire M&A market of the state. Second, to our knowledge, this is the first study on this topic that analyses a post-communist economy, and so the findings could be in the interest of both researchers and business people. In the end, our study's contribution to the existing literature consists of finding that there are some specific industries that are more attractive for acquirers who are placed near target companies, from a geographical point of view. In addition, domestic acquirers are less successful in takeover processes than foreign investors. This paper is structured as follows. In Section 2 we present the main findings of other studies in the field. The next section deals with our main research questions, and Section 4 presents the database and methodology. After this, Section 5 presents the main empirical results and the final section contains the conclusions and discussion on the study findings.
Related studies
There are many studies that analyse domestic and cross-country mergers and acquisitions, but only a few that are aimed to provide findings as to whether nearby regional or state characteristics can influence the decision-making and success of a transaction. So, the importance of the geographical distance in M&As, is a frequent research question, as is the possible advantages that the proximity confers to the companies engaged in M&A transactions (Chakrabarti & Mitchell, 2013;Green, 1990;Green & Cromley, 1984). Studies in this field are limited only to the US, Canadian, German and Finnish domestic M&A markets, but a lot of research has been done in this area recently. Green (1990) investigated the US market in the 1970s and 1980s and concluded that the economic centres of a country are key factors in developing the nearby areas. According to this idea, M&As contributed to the increase in corporate control for the urban agglomerations and also to a reinforcement of the existing company headquarters (Chapman & Edmond, 2000). Therefore, cities are considered to be the locations of power and control and with the help of the M&As their power will increase even more . Geographical proximity can facilitate the success of a transaction through the ease of transmitting information about the future investments (Sorenson & Stuart, 2001). In addition, the information is more accurate when the companies are close to each other and the flow is direct (Ensign et al., 2014;Uysal et al., 2008).
All the related literature shows that geographic distance negatively influences the success of a M&A deal. The increase in distance will lead to a decrease in the chance of the M&A being successful. Some studies positively associate distance with the asymmetric information between the acquirer and the target (Ragozzino & Reuer, 2011). Nearby acquirers are in a better position to appreciate the key resources of the target company (e.g., human capital, key technologies, brands, growth prospects, and relationships with other firms and customers) while distant acquirers are likely to lack of this kind of information.
From a political point of view, the companies try to reduce their competition in the technology or in market products through an M&A. So acquiring nearby rivals can give a company the comfort of not competing with others. In this case, the companies can reduce their R&D spending in short run and make savings in terms of personnel or equipment (Cassiman, Colombo, Garrone, & Veugelers, 2005). This could be another reason why geographic distance is an important factor for M&A decisions (Bertrand, 2009). Kang and Kim (2008) -in a study on the US M&A market (block acquires), this time on a sample between 1990 and 1999 -indicate that acquirers have a strong preference for geographically nearby targets and there is a strong link between geographical proximity and corporate governance. Kang and Kim (2008) conclude that proximate block acquirers are more likely to engage in post-acquisition governance activities with the target companies. For example, if they are located in the same state or within 100 km from their targets it is more likely that their representatives will be on the target's board or management team. The conclusion that geography matters is also present in other studies over the last decade of the US M&A market. Erel, Rose, and Weisbach (2012) states that the odds of acquiring a firm in a nearby state are substantially higher than the odds of acquiring a firm in a distant state. In addition, higher economic development, and better accounting quality increase the chances of the target being acquired.
In Canada, domestic mergers and acquisitions are concentrated in the major cities in the south of the country (Aliberti & Green, 1999). Both in the US and Canada the findings suggest that distance is a key element when it comes to decision-making in the M&A sector (Green, 1990). realise an interesting study on the changes that occur in the German economic activity do to the M&A waves after the 1990. They focused their attention on the importance of the location of different industries and the dynamics of these industries. The results show that there is a massive exchange of capital and knowledge between firms situated in nearby areas. This finding is often recorded in the 'new economy' industries, such as media, services, and IT&C, and they are less obvious in the 'old sectors' , such as heavy industry, the textile industry or vehicle construction.
In Finland, the evidence reveals that geographical closeness directly influences the takeovers in cases of domestic M&As. The findings suggest that the acquiring firm can monitor the target company in the pre-acquisition period if the companies are in close geographic range, but also the acquiring company can do the same thing with companies from distant places if it has a long history and is a permanent presence in the market (Böckerman & Lehto, 2006). In addition, the main takeover flows are within the regions that contain a large number of companies.
Other studies are focused on cross-country mergers and acquisitions and, similar to the domestic transactions, geographical proximity can influence the success of a transaction. Green (1990) analysis this factor among others -such as the market, similarity in language or legal structure -and concluded that there is a dominance of the UK and Canada in the US market of M&As. Rodríguez-Pose and Zademach (2004) found the same thing in Germany. Other studies on cross-border M&As identify that cultural boundaries (language, religion or political regime) (Ahearn et al., 2010) along with the increase in distance (Rose, 2000) can increase the costs of combining two firms and so the likelihood of success of the transaction drops. In the end, even if we have a cross-border M&A, the distance between the acquirer and the target plays an important role in defining the international transactions (Hijzen, Görg, & Manchin, 2008;Portes & Rey, 2005).
Some studies found only a small impact of geographic distance on cross-country M&As (Coeurdacier, De Santis, & Aviat, 2009). A possible explanation is that, as a large number of transactions studied took place from 1995 onwards, and the database only consists of M&As from developed countries, the information is present everywhere and can be accessed by anyone, so the geographical proximity of the acquirer and target company can become statistically insignificant. The same study concludes that this is not the case if we analyse developing countries. Here, distance could make the difference between a successful and an unsuccessful transaction.
In the Romanian M&A market, several studies have been carried out. Their main focus was the estimation of the control premium and the determinant of this premium (Dragotă et al., 2007;Dragotă, Lipară, Ciobanu, 2013). Some of the indicators used in these studies were also used in ours to see whether they can explain the probability of success of an M&A transaction in the Romanian post-communist economy.
Study hypotheses
As Ragozzino andReuer (2011), Cassiman et al. (2005), Ellwanger and Boschma (2014) or Uysal et al. (2008) concluded in their studies, geographical proximity can facilitate the success of a transaction by the ease of transmitting information between the two parties involved. We expect the same result from our study, which takes into account transactions between companies in the Romanian market, that a nearby acquirer will have an advantage over a distant one in a takeover bid. Hypothesis 1. The geographical proximity influence the takeover bids. Böckerman and Lehto (2006), Erel et al. (2012) concluded that the success of an M&A transaction depends on the good knowledge of the target. We expect that the age of the firm can express this feature and so the chance of a successful takeover rises with the age of the target company. We will test whether this is the case in Romania, a former communist country that moved to a market economy not many years ago and where many state companies were privatised through different national programmes.
Hypothesis 2. The age of the target company can increase the probability of success of a takeover.
Database and methodology
The database contains all the transactions that took place on the BSE (Bucharest Stock Exchange) and RASDAQ (the Romanian equivalent for US NASDAQ) during the period 2000-2014. The analysed the takeover bids that resulted in a change in control of a company or could have led to a change of the controlling ownership in case of a successful transaction. A shareholder is considered to have the controlling position in a company if he/she owns 50% or more from the firm's equity. We considered there to be a successful transaction if, at the end of the takeover bid, the company had a change of the controlling shareholder. We did not take into account transactions that not imply the possibility for the bidder to take effective control of the company.
During 2000 and 2014 more than 1600 purchase public offers where recorded on BSE and RASDAQ. From these, only in 490 cases did the buyer want to gain the controlling position, and so could be classified as takeover bids. We excluded from the database the takeover bids where the acquirer was a physical person, because the aim of this study is to analyse the impact of the geographical proximity of two companies. In the end, our database consisted of 320 transactions where both the bidder and the target are legal persons. We considered the location of a company by the city where its headquarters are placed. In order to place all these takeover transaction in country regions, we used the NUTS 2 level proposed by Eurostat. 1 The region analyses can provide interesting information about the geographical development and they were also used in several studies (Škuflić & Botrić, 2009). In Romania, there are eight regions named after their geographical position in the country (see Figure 1).
The methodology proposed for this study consists of two separate analyses. First we realise a statistical overview of the M&A market in Romania with a series of maps highlighting the preferred region for takeovers or the acquirers' regions from where the transaction started. Second we estimate, using regressions, the probability of the success of a takeover judging by the geographical distance between the acquirer and the target, the age of the target company and other characteristics of the firms or region involved.
After the transition to a democratic regime after 1990, Romania faced two M&A waves. The first began in 1995-1996 and ended around 1999 and was characterised by a massive privatisation programme that included many important and profitable Romanian state companies. Both domestic and foreign investors were attracted because, in many cases, the cost of the transaction was reduced, and the possibilities of future profits where high as long as a professional management team was in charge. By the end of this period, no more than 20% of the state company's shares were actually transferred to private owners and less than 10% of the companies were actually privatised (Earle & Telegdy, 1998). A second wave came after 2000 when Romania, like other East European states, faced an increasing politic and economic stability that ended with the accession to the European Union in 2007. In this period, many companies were accepted for trading on the BSE or RASDAQ market, so the transfer of ownership was easier to implement. In this way, many companies were acquired by large domestic and foreign firms with the same industrial profile.
For the first part of our analysis, we want to investigate whether the takeover bids are equally spread around the state or are concentrated in one or more cities or regions. In accordance with the literature (Aliberti & Green, 1999;, many of the takeovers bids were started or targeted firms that are located in the main urban agglomerations. Only in less than 13% do we find that the target company is located in a rural or a small city location. So we directed our investigation to the regions where the headquarters of the companies involved in the takeover offers are located. We generated two sets of maps. In each one there is a map for the target's region and one for the acquirer's region ( Figure 2).
In the acquirers map we only included the domestic acquirers. In almost 20% of the cases, the acquirers were companies with headquarters located outside the Romanian border, so these transactions were excluded from the analysis. Even if these companies are registered in other countries, many of them have Romanian ownership and are located in states like Cyprus for tax payments facilities.
As we can see in Figure 1 (left map), the target companies are located mainly in the Centre and South-Eastern region and Bucharest region. In the communist period, the Centre region was characterised by the clothing industry (factories in Miercurea Ciuc, Târgu Mures or Sibiu), the machinery industry (in Braşov and Sibiu) and the mining and manufactory industry (the largest salt and natural gas resources in Romania). The South-Eastern region also had other types of companies, which are specialised in machinery construction (especially naval and maritime) and services or commerce (due to the proximity of the Black Sea and Danube River, a main route between Western Europe and the Middle East). All these types of companies are also highly profitable at present, so this is why they are of continuous interest to investors. Besides the foreign companies, many of the domestic companies involved in the takeover bids are located in the Centre, Western South-Eastern and Bucharest regions. This confirms the hypothesis that the geographical proximity does matter in M&A activity. Similar to the results of other studies on different countries (Uysal et al., 2008 Böckerman & Lehto, 2006 on Finland), the acquirer was located in the same region or in the proximity of the target's region. If we express the geographical proximity in kilometres we can illustrate the density of the takeover bids by the spatial distance (Figure 3).
We statistical demonstrated that geographical proximity influences takeover bids in Romania. Another question arises. Is this is a common thing, or are there some industries where this is happening in most cases or industries where only for a few transactions are the target and acquirer located in the same region? To analyse this we split our database into 11 industry sectors: Clothing industries, Commerce, Constructions, Real Estate, Food Industry, Chemical Industry, Machinery, Manufactory, Services, Transport and Tourism. We created a chart that illustrates which sector is above the country mean by the number of takeover bids and if, in those cases, it happened that the target and the acquirer are in the same region ( Figure 4). In five sectors the number of regional takeover bids exceeds the country average, but in only two cases does the number of takeover bids per industry exceed the country average per sector.
Furthermore, we carried out a deeper analysis to see what changes appear when we limit our database only on those takeover bids that ended with a change in control ( Figure 5). Again, we can see that there is the same distribution of the takeover bids. Also, if there is a high number on takeovers where the target is located in the Centre, South-Eastern or Bucharest region, the acquirer is located almost always in the same region or the regions nearby. In conclusion, the hypothesis that the geographical proximity is a determinant of the M&As is still confirmed.
In the second part of our study, we analysed which are the determinant variables that can influence the probability of a takeover bid being successful. We used a Probit model and we based it on the assumption that the probability of an event occurring is linearly related to a set of explanatory variables. The coefficient estimated for the linear probability model can be interpreted as the change in the probability for the dependent variable when the independent variable will equal 1. In order to do that, we conducted several regression models to test the hypotheses presented in Section 2 by using variables that characterise the companies involved in the takeover bid and regions where the target's or acquirer's headquarters is located. The variables used are documented in Table 1. Some descriptive statistics for the variables used in the model are presented in Table 2. Overall, the percentage of the target's equity owned before the takeover bid by the acquirer was around 15%. The mean age of the target company is around 44 years with a maximum of 172 years. Almost 56% of the target companies are located in same region and approximately 39% even in the same city. The mean distance between the target and the acquirer is 131 kilometres. Almost 80% of the takeover bids involved domestic transactions.
We have also tested for multi-collinearity. The correlation matrix for this is presented in Table 3. We did not consider in the same regression the variables correlated at a higher level than 0.3. Green and cromley (1984), Green (1990), chakrabarti and mitchell (2008) Region characteristics mEanGDP the level of GDP per capita of the regiondummy variable (1 if the region is above the country's average, 0 if not) mEanR&D the level of R&D spending per company of the region -dummy variable (1 if the region is above the country's average, 0 if not)
Results
To test the hypotheses outlined in Section 2 we used several regression models with the dependent variables presented in the methodology section. In support of our hypotheses is Table 4, which presents the probit regression results performed on our database consisting of all the takeover bids on the Romanian capital market between 2000 and 2014. Our dependent variable is a binary variable that takes the value 1 if the takeover bid is successful and 0 otherwise. We consider a successful takeover bid to be the case that the buyer wanted and succeeded in gaining a controlling position (50% or more of the firm equity) in the target company. For the first set of regressions we estimated the probability for all 320 transactions, where the target companies are located in Romania and the acquirers can be domestic or foreign. We did not include the distance variable because of the lack of data on the exact location of the foreign company. For the second set of regressions, we limited the database only to domestic acquirers. For a better analysis, from the author's point of view, we excluded the recordings where the acquirer and the target company are placed in the same city. The findings from these models are consistent with those in the first part of our analysis. They also confirm the hypothesis from the second section. The geographical proximity influences positively the probability for a successful takeover. This is in accordance with the results of the studies conducted by Erel et al. (2012), Kang and Kim (2008), Bertrand (2009) or Ragozzino and Reuer (2011). If the target and the acquirer are located in the same city the probability for the takeover bid to be successful is higher. In contradiction to this, domestic acquirers tend to have less chance of a successful takeover than foreign investors. Our results are in contradiction with those of Green and Cromley (1984), Green (1990) or Chakrabarti and Mitchell (2008), who concluded that domestic acquirers have more success in a takeover because of the geographical proximity. In our study, a possible explanation for the results can be that domestic acquirers usually do not have the amount of investment capital that foreign acquirers have.
The age of the company is also directly related to the probability of success, which confirms our second hypothesis. Indeed, older firms have more public information available about them and also they are well known to the general public. If the acquirer companies have good monitoring capacity they may be able to overcome the geographical boundaries more easily. The same results are found in Böckerman andLehto's (2006), andErel et al. 's (2012) studies.
The percentage of shares owned before the takeover bid was also significant. If an acquirer already owns some of the target's equity, the probability of the success is very high. Furthermore, it seems that it is easier to succeed in a takeover if there is a concentrated ownership. If an acquirer wants to buy the controlling stake, he only has to convince a limited number of shareholders. According to our findings, if the company is larger, the source: own calculation and results. to estimate the probability of a successful takeover bit we used the probit regression model. the regression uses 320 observations of takeover bids on the Romanian capital market between 2000 and 2014.
We consider a takeover bid if the buyer wanted to gain a controlling position in the target company. We did not consider in the same regression the variables correlated at a higher level than 0.3. t-statistics are in parentheses. the symbols *, **, *** represent significance levels of 10, 5 and 1%. probability of a successful takeover bid is reduced. None of the region characteristics was significant. Surprisingly, the region dummy was also insignificant so we cannot confirm the statistical findings from the first part of our analysis.
Our results show that both the GDP and the R&D spending are not significant. So the success of takeover bids is not influenced by the economic and innovative development of the region where the companies are placed. A possible explanation can that in the communist regime, state companies and large factories were intended to be uniformly spread out in Romania. After the transition to the market economy, some geographic regions felt behind in terms of strategic development, but the companies remained placed there and were targeted for takeover transactions.
According to the literature, most of the studies were conducted only on domestic M&A transactions (Böckerman & Lehto, 2006;Erel et al., 2012;Kang & Kim, 2008). In order to do this, in our study we limited our database only to domestic acquirers and also we excluded the takeover bids where the target and the acquirer company are located in the same city. Again we used the probit regression (Table 5). In these models we included the geographical distance between the dependent variables.
Again, the geographical proximity of the acquirer and the target can influence positively the probability of success of the takeover bid. If the distance is higher, the probability that the takeover will be successful is lower. This confirms again out hypothesis from Section 2. Similar results are found in Erel et al. (2012), Kang and Kim (2008), and Bertrand (2009). Also, in this case, when the database consists only of domestic takeover bids and the acquirer and target company are not placed in the same city, the acquirer has a higher probability of success if it is placed in the same region as the target company.
Another interesting result is the region dummy for GDP per capita. If the region GDP per capita is higher, the probability for a successful transaction is lower. This means that We consider a takeover bid if the buyer wanted to gain a controlling position in the target company. We did not consider in the same regression the variables correlated at a higher level than 0.3. t-statistics are in parentheses. the symbols *, **, *** represent significance levels of 10%, 5% and 1%. domestic acquirers look to invest in a low developed region, where there are possibilities of synergetic gains. This result is interesting because it is complementary with our previous finding when we analysed the entire database. A possible explanation can be that if a domestic investor is willing to invest in another city, he will probably look to a more developed region. The results are different for the R&D expenses being a not significant variable for explaining the success of takeover bids. This can be due to the fact that R&D expenses are quite low in Romania and investors do not see a competitive advantage in being present in one region or another.
Conclusion
The probability of a successful takeover transaction is highly influenced by geographical factors. First of all, if the acquirer is located in the proximity of the target company there is an increase in the probability that the transaction will end with a change of control. We measured the geographical proximity by using either a distance quantitative variable (expressed in kilometres) or dummy variables (for the same city or region of the companies involved in the takeover). Second, the age of the target company is also a significant factor for the probability of a successful takeover. The older the firm is the greater the probability. An older company expresses more confidence to investors and there is a good chance there will be positive synergies from the transaction.
As presented in the study, some results are consistent with those found in the literature, (the geographical proximity, age, capitalisation, ownership, etc.) but we also found some particularities for Romania. For example, foreign acquirers have greater success in takeover bids because, in many cases, they score better with regard to financial resources, management strategy and innovation than domestic acquirers.
This study is important both for practitioners and for those in academia. From the author's point of view, the results are of great interest for practitioners working in the M&A divisions of various consulting agencies. The results can be used by investors and M&A specialists in Romania to see which sectors are more appealing to acquirers and also in which geographical regions it is better to develop a business. For the further development of our study we can make parallel analyses on every geographical region to see why some are more appealing to acquirers then others. In forthcoming studies, I believe that an analysis of the industry sectors and knowledge capital (number of patents, R&D spending) of the target firm is necessary to see whether these characteristics can also influence the possibility of a successful takeover transaction. Note 1. The NUTS classification (Nomenclature of territorial units for statistics) is a hierarchical system for dividing up the economic territory of the EU. This was made for the purpose of the collection, development and harmonisation of EU regional statistics, socio-economic analyses of the regions (NUTS 1: major socio-economic regions; NUTS 2: basic regions for the application of regional policies; NUTS 3: small regions for specific diagnoses) and framing of EU regional policies (regions eligible for aid from the Structural Funds (Objective 1) have been classified at NUTS 2 level, areas eligible under the other priority objectives have mainly been classified at NUTS 3 level). | 7,274.4 | 2016-01-01T00:00:00.000 | [
"Business",
"Economics",
"Geography"
] |
Cellcano: supervised cell type identification for single cell ATAC-seq data
Computational cell type identification is a fundamental step in single-cell omics data analysis. Supervised celltyping methods have gained increasing popularity in single-cell RNA-seq data because of the superior performance and the availability of high-quality reference datasets. Recent technological advances in profiling chromatin accessibility at single-cell resolution (scATAC-seq) have brought new insights to the understanding of epigenetic heterogeneity. With continuous accumulation of scATAC-seq datasets, supervised celltyping method specifically designed for scATAC-seq is in urgent need. Here we develop Cellcano, a computational method based on a two-round supervised learning algorithm to identify cell types from scATAC-seq data. The method alleviates the distributional shift between reference and target data and improves the prediction performance. After systematically benchmarking Cellcano on 50 well-designed celltyping tasks from various datasets, we show that Cellcano is accurate, robust, and computationally efficient. Cellcano is well-documented and freely available at https://marvinquiet.github.io/Cellcano/.
Supplementary Figures
Supplementary Figure 1 Supplementary Figure 12 A diagram shows our procedure of data preprocessing and data analysis.
Download scATAC-seq raw data (fragment files or bam files) Genome liftOver (hg19 for human PBMCs datasets; mm10 for mouse brain datasets) Generate genome-wide bins, peaks and gene scores by ArchR Curate cell types in human PBMCs and mouse brain datasets
Supplementary Note 1: Details on designing celltyping tasks
In total, we designed 50 celltyping tasks involving different individuals as reference and target datasets from six datasets (four human PBMCs datasets and two mouse brain datasets). We design the celltyping tasks to mimic the following prediction scenarios: -Intra-dataset individual prediction: users have one confidently annotated scATAC-seq profile from one individual and want to use it to annotate all other individuals from the same study. -Inter-dataset individual prediction: users have one confidently annotated scATAC-seq profile from one individual and want to use it to annotate other individuals from different studies. In the mouse brain celltyping tasks, a special case is that we have tasks not only for a different subject but also for a different brain region because mouse brain has several brain regions. We count them into this category. -Inter-dataset prediction (combined reference): users have several well annotated scATAC-seq datasets and wish to use a large collection of public datasets to increase the reference data size and improve the prediction result. This is based on our previous research 5 where we found that combining individuals or datasets as reference could lead to better prediction results. -Inter-dataset prediction (combined target): users have scATAC-seq data from multiple batches and want to determine their cell types in one run using a given reference.
We have one more task design which is Inter-dataset prediction (Ground truth) where we use the FACS-sorted human PBMCs dataset as target dataset. Since the FACS-sorted human PBMCs dataset can be considered as the ground truth, we use this category to better evaluate how Cellcano predicts compared to all other methods. However, this category will not appear in real cell type prediction scenario.
Supplementary Note 2: An introduction to different ArchR gene score models
The script to generate gene score models are provided by ArchR 6 (https://github.com/GreenleafLab/ArchR_2020). In total, there are eight categories of gene score models including: (1) Model -Promoter: This class of models count the reads located on the promoter region with different window sizes.
(2) Model -GeneBody: This class of models count the reads located on the whole gene body with certain extension in up-or down-stream.
(3) GeneModel -Constant: This class of models count reads from 1K bps upstream transcription start site (TSS) and different bps downstream TSS. The constant gene model considers each read having the same weight as 1.
(4) GeneModel -TSS -Exponential: This class of models extract reads from 1K bps upstream and 100K bps downstream TSS. Gene boundaries are set so that reads from one gene body will not overlap with other gene bodies. Then, an exponential decay function is used to weight the reads from each windowed tile based on the distance to TSS. The exponential decay function is demonstrated as exp (− !"#(%&#'!()*) ,&(%-, + exp (−1)) with different window parameters. (5) GeneModel -TSS -NoBoundary -Exponential: Same as (4) except no gene boundaries are set. (6) GeneModel -GB -Exponential: Same as (4) except the distance in the exponential decay function is calculated based on the distance to gene bodies instead of TSS. Gene boundaries are set in this class of models. (7) GeneModel -GB -Exponential -Extend: Same as (6) except the gene bodies are extended. The distance in the exponential decay function is calculated based on the extended gene bodies. The gene score model recommended by ArchR lies in category (7). It integrates the signals from the gene body with TSS extended 5kb in the upstream direction. Then, it weights the reads outside the gene body region and use the window parameter as 10,000. | 1,122.8 | 2023-04-03T00:00:00.000 | [
"Computer Science",
"Biology"
] |
A Revised Mechanism for the Activation of Complement C3 to C3b
Background: An understanding of the solution structure of complement C3b is essential to understand its reactivity. Results: Ultracentrifugation and scattering revealed compact C3b structures in low salt and extended ones in physiological salt. Conclusion: The two conformations reflect Arg102–Glu1032 salt bridge formation only in low salt. Significance: The functional differences between the major C3S (Arg102) and C3F (Gly102) allotypes are explained.
The complement system comprises over 30 proteins arranged in a cascade as part of the innate immune response and is important for clearing immune complexes and cellular debris and for the elimination of pathogens (1,2). C3 (complement component 3) is the most abundant complement protein, occurring at about 1.0 mg/ml (5.3 M) in plasma and at higher levels during inflammation. All three of the complement classical, lectin, and alternative activation pathways lead to the cleavage of C3a from C3 to form active C3b, which attaches covalently to cell surfaces through a thioester bridge (1,2). The spontaneous hydrolysis of the thioester bond in C3 leads to C3u (also known as C3 H2O ), and C3u is an activator of the alternative pathway (3) but is unable to attach to cell surfaces. C3b and C3u are regulated by the cofactor Factor H and the protease Factor I, together cleaving C3b to produce the inactive C3c and C3d fragments. Because Factor H binds to host cell surfaces and not to bacteria, host cells are protected from C3b activation (4). The two major polymorphic allotypes of C3 are C3S (with frequencies of 0.79 and 0.99 in white and Asian populations, respectively) and C3F. The nomenclature is derived from their electrophoretic mobility on agarose gels, with C3S moving slowly and C3F faster (5). C3F is associated with diseases, including IgA nephropathy, systemic vasculitis, partial lipodystrophy, membranoproliferative glomerulonephritis type II, and age-related macular degeneration. C3F may increase complement activation levels (5). In complexes with erythrocytes, C3F showed a greater affinity for mononuclear cells than C3S (6). In hemolytic activity assays, C3F showed higher alternative pathway activation than C3S due to a lowered efficiency in Factor H regulation (7).
C3 belongs to the ␣ 2 -macroglobulin protein family, members of which have a reactive thioester bond and similar domain arrangements (4). The ␣and -chains of C3 (115 and 75 kDa, respectively) are arranged as 13 domains, namely eight macroglobulin domains (MG1-MG8) 2 and five single linker (LNK), anaphylatoxin (ANA; C3a), C1r/C1s-UEGF-BMP1 (CUB), C345c, and thioester-containing domains (TED; C3d) (8) (Fig. 1A). In C3, the thioester bond is buried within the TED-MG8 interface (9). During C3b activation, the release of C3a results in large conformational changes that expose the reactive TED thioester bond, which is able to bind rapidly and covalently to nucleophiles on cell surfaces (4,10). Four crystal structures of C3b employed 50 mM NaCl as the crystal precipitant (11)(12)(13)(14). All four structures showed that the TED domain made contact with the MG1 domain at the base of the C3b structure in a compact arrangement (Fig. 1C). A two-segmented structure for C3 and its homologues C4 and C5 was originally deduced using low resolution scattering modeling (15). Recent x-ray scattering and ultracentrifugation of C3u in 137 mM NaCl buffer showed that C3u has a more extended structure than in the C3b crystal structures, in which the TED and MG1 domains were well separated in C3u (Fig. 1B) (16). X-ray scattering of C3b in 200 mM NaCl with or without 5% glycerol suggested that C3b existed as multiple conformers with extended TED domains (17,18). In addition, crystal structures for ␣ 2 -macroglobulin and complement C5b in complex with C6 showed that the TED domain in these structures was also well separated from their MG1 domains (19 -21).
An understanding of the different locations of the functionally crucial TED domain in the ␣ 2 -macroglobulin protein family is essential to explain the different clinical associations of the C3S and C3F allotypes. We utilized a joint ultracentrifugation and scattering approach coupled with constrained atomistic modeling for the C3 proteins (22). Analytical ultracentrifugation established the formation of TED-mediated dimers and enabled the scattering data to be corrected for these. Improved x-ray and neutron scattering data collection in 50 and 137 mM NaCl buffers showed that the TED-MG1 domains were connected in C3b and C3u in 50 mM NaCl buffer but separated in physiological 137 mM NaCl. Crystal structures showed an Arg 102 -Glu 1032 salt bridge between the TED and MG1 domains as part of the C3b regulatory complex with Factor H. We show that the loss of this salt bridge revises our understanding of C3b activity, making the TED domain more reactive than previously thought. Because the C3S and C3F allotypes contain Arg 102 and Gly 102 , respectively, we can now explain for the first time the different functionality of the C3S and C3F allotypes.
EXPERIMENTAL PROCEDURES
Purification of C3, C3u, C3b, C3c, and C3d-C3 was purified from fresh human plasma essentially using a Q-Sepharose fast flow anion-exchange column (Amersham Biosciences) and a Mono Q 5/50 GL column (GE Healthcare) (23). The three donors were genotyped for the rs2230199 single nucleotide polymorphism (R102G) to show that all three were Arg 102 (wild-type C3S allotype). C3u was produced by incubating C3 with 200 mM hydrazine for 2 h at 37°C in a water bath and leaving at 4°C overnight. C3b was produced by treating 1 mg/ml C3 in Hepes buffer (10 mM Hepes, 137 mM NaCl, 0.5 mM EDTA, pH 7.4) with 10 g/ml trypsin (1% w/w enzyme/substrate) for 120 s at 37°C in a water bath and then adding 40 g/ml soybean trypsin inhibitor to stop further cleavage before transferring onto ice. Next, 20 mM iodoacetamide was added to the mixture to block the thioester, and then this was incubated in the dark at 20°C for 30 min (13). C3b was diluted in Tris buffer (25 mM Tris, 140 mM NaCl, 0.5 mM EDTA, pH 8.0), concentrated immediately, and passed through a Superose TM 6 preparation grade XK 16/60 size exclusion column. For C3c, outdated human plasma was incubated for 7 days at 37°C in a water bath, and then C3c was purified following the same protocol as that for C3. C3u and C3b (but not C3) were active in functional assays using Factor I and Factor H (24). Recombinant C3d was expressed in Escherichia coli with a GST tag and purified by thrombin cleavage using a GSTrap FF 1-ml column (GE Healthcare) connected with a HiTrap Benzamidine FF (high sub) 1-ml column (GE Healthcare) (25). Western blots were performed to confirm the identity of all five proteins, using an anti-complement 3 goat polyclonal antibody (Calbiochem). The absorbance coefficients for C3, C3u, C3b, C3c, and C3d (1%, 280 nm, 1-cm path length) were calculated from their compositions to be 9. 40, 9.40, 9.83, 9.21, and 13.15, respectively, assuming the presence of three high-mannose type oligosaccharides at Asn 63 , Asn 917 , and Asn 1597 in C3 (26,27). Molecular masses were calculated from compositions to be 189.0 kDa for C3 and C3u, 179.3 kDa for C3b, 135.7 kDa for C3c, and 34.6 kDa for C3d. All proteins were passed through a size exclusion gel filtration column (C3, C3u, C3b, and C3c in Superose 6; C3d in Superdex 200) to remove potential aggregates. For all experiments except for those in heavy water, the proteins were dialyzed into 10 mM Hepes, 50 mM NaCl, pH 7.4, or 10 mM Hepes, 137 mM NaCl, pH 7.4 (denoted as 50 mM NaCl or 137 mM NaCl, respectively, below). For heavy water dialysis, phosphate-buffered saline (PBS-2 H 2 O) was used (137 mM NaCl, 8.1 mM Na 2 HPO 4 , 2.7 mM KCl, 1.5 mM KH 2 PO 4 , pH 7.4). Each protein was routinely checked by SDS-PAGE before and after the ultracentrifugation and scattering experiments.
Sedimentation Velocity Data Collection and Analyses-By analytical ultracentrifugation, sedimentation velocity data were obtained on two Beckman XL-I instruments equipped with AnTi50 rotors, using two-sector cells with column heights of 12 mm at rotor speed of 50,000 rpm. The five proteins C3, C3u, C3b, C3c, and C3d were monitored using absorbance optics at 280 nm and interference optics. For 137 mM NaCl buffer, concentration series at 20°C were performed for C3b at concentrations between 0.25 and 1. The continuous size distribution c(s) analysis method was used to determine s 20,w values using SEDFIT (version 14.1) (28,29). The c(s) analyses provided size and shape data by directly fitting the experimental sedimentation boundaries to the Lamm equation for up to 300 interference scans for the five proteins; 25 absorbance scans for C3, C3u, C3b, and C3c; and 80 absorbance scans for C3d. The c(s) analyses were based on a fixed resolution of 200 and floated the meniscus, cell bottom, baseline, and average frictional ratio f/f o (where f o is the frictional coefficient of the sphere with the same volume as the hydrated glycoprotein). The starting f/f o values were 1.3 for C3, 1.4 for C3u, 1.38 for C3b, 1.35 for C3c, and 1.2 for C3d. Fits proceeded until the overall root mean square deviation and agreement between the observed and calculated sedimentation boundaries were satisfactory. The proportion of monomers and oligomers was quantitated by SEDFIT integration.
Buffer densities were measured at 20°C using an Anton Paar DMA 5000 density meter for comparison with the theoretical values calculated by SEDNTERP (30). This gave densities of 1.00487 g/ml for 137 mM NaCl (theoretical, 1.00485 g/ml), 1.00195 g/ml for 50 mM NaCl (theoretical, 1.00197 g/ml), and 1.112381 g/ml for 137 mM NaCl PBS-2 H 2 O. The viscosities of H 2 O and 2 H 2 O were taken as 1.002 and 1.251 centipoises, respectively (31). The partial specific volumes v were computed as 0.739 ml/g (C3, C3u, C3b, and C3c) and 0.747 ml/g (C3d).
X-ray and Neutron Scattering Data Collection and Analyses-X-ray scattering data were obtained for the five proteins at 20°C in concentration series in each of 50 and 137 mM NaCl buffers in one beam session on Beamline ID02 at the European Synchrotron Radiation Facility (Grenoble, France), operating at 6 GeV in 4-bunch mode to reduce the incident flux (34). The sample-to-detector distance was 3 meters, and the x-ray wavelength was 0.0995 nm. Potential radiation damage was eliminated by the continuous movement of the sample in a flow cell during beam exposure, the use of 10 time frames with durations of 0.3 s/frame or 0.5 s/frame in each acquisition, and visual on-line checks for radiation damage at low Q. No detectable effects from radiation damage were seen in the final data sets. In 137 mM NaCl, C3 was studied between 0.2 and 1.49 mg/ml (1.05-7.9 M), C3u between 0.2 and 1.5 mg/ml (1.05-7.9 M), C3b between 0.2 and 1. Other details, including data reduction, are described elsewhere (35,36).
Neutron scattering data were obtained on Instrument SANS2d at the ISIS pulsed neutron source at the Rutherford Appleton Laboratory (Didcot, UK) (37). Neutrons were derived from proton beam currents of ϳ40 A. Time of flight data were recorded with 4-meter collimation, a 4-meter sample-to-detector distance, a 12-mm beam diameter, and a wavelength range from 0.175 to 1.65 nm. The five proteins were measured in 2-mm-thick quartz circular banjo cells in a thermostatted rack at 20°C. Data acquisitions lasted 1.5-3. In a given solute-solvent contrast, the radius of gyration (R g ) corresponds to the mean square distance of scattering elements from their center of gravity and is a measure of structural elongation. Guinier plots at low Q give the R g and scattering at zero angle I(0) from the following expression (39).
This expression is valid in a Q⅐R g range up to 1.5. The I(0)/c value (c is the protein concentration in mg/ml) is proportional to the relative molecular mass M r . If the structure is elongated, the radius of gyration of the cross-sectional structure (R xs ) and the mean cross-sectional intensity at zero angle (I(Q)⅐Q) are obtained from Q values larger than those used for R g analysis.
The R xs value monitors the elongation of its cross-section shape along its longest axis. The Guinier analyses were performed using an interactive PERL script program SCTPL7 (J. T. Eaton and S. J. Perkins) on Silicon Graphics OCTANE workstations.
Indirect transformation of the I(Q) curve measured in reciprocal space into real space gives the distance distribution function P(r) and was carried out using the program GNOM (40).
P(r) corresponds to the distribution of distances r between volume elements. This offers an alternative calculation of the R g and I(0) values that is based on the full scattering curve I(Q) and not that at low Q. It also gives the most frequently occurring distance M and the maximum dimension of the macromolecule L. For the five proteins, the x-ray curves utilized up to 340 data points for Q values between 0.16 and 1.50 nm Ϫ1 . The neutron SANS2d curves utilized up to 50 data points for Q values between 0.20 and 2.1 nm Ϫ1 , whereas the neutron D22 curves utilized up to 109 data points for Q values between 0.09 and 1.96 nm Ϫ1 . Other details are described elsewhere (24,35,36). Scattering and Sedimentation Modeling of C3, C3u, C3b, C3c, and C3d-The scattering modeling of human C3, C3c, and C3d utilized their crystal structures (Protein Data Bank codes 2A73, 2A74, and 1C3D, respectively). To calculate their scattering curves, the C3, C3c, and C3d crystal structures were converted into Debye spheres. A cube side length of 0.54 nm for all three structures with a cut-off of four atoms gave totals of 2047, 1408, and 367 spheres for the neutron modeling that were within 1% of their unhydrated volumes of 1557, 1112, and 224 nm 3 , respectively. The hydration shell detectable by x-ray scattering was incorporated by adding extra spheres to the surface of the unhydrated sphere model using HYPRO (41), based on a hydrated volume of 0.3 g of H 2 O/g of glycoprotein. The optimal totals of hydrated spheres were 2047, 1408, and 367, respec-tively, for C3, C3c, and C3d. The scattering curve I(Q) was calculated using the Debye equation adapted to spheres (42). Details are given elsewhere (33,43). The four C3b crystal structures used here have Protein Data Bank codes of 2I07, 2WIN, 2WII, and 2ICF.
The constrained scattering modeling of C3, C3u, and C3b followed previous procedures (16). The x-ray modeling for C3 and C3u was extended to the neutron modeling. C3u and C3b were each considered in terms of a C3c region connected to the CUB-TED domain pair by a conformationally variable linker between the MG8 and CUB domains (Fig. 1C). The same 4650 trial models previously generated from the C3b crystal structure (Protein Data Bank code 2I07) showed linker lengths between 0.59 and 2.72 nm. The unhydrated models for C3b and C3u contained 1478 spheres, and the hydrated models contained 1945 spheres. The C3 modeling was considered in terms of varying the two linkers between the CUB-MG7 and CUB-TED domain pairs, to result in 8000 trial C3 models. The C3 models optimally contained 1559 spheres (unhydrated) and 2050 spheres (hydrated). Steric overlap between the C3c region and the CUB and/or TED domains was assessed using the number of spheres n following grid transformation. Models with less than 95% of n were discarded. The R g and R xs values of the remaining models were required to be within Ϯ5% of their experimental values. Models that passed the n, R g , and R xs filters were ranked using a R-factor goodness of fit parameter defined by analogy with protein crystallography. The curve fits used experimental curves with Q ranges extending up to 1.50, 2.1, and 1.96 nm Ϫ1 for C3b, C3, and C3u, respectively. The 10 best fit C3b models from 50 and 137 mM NaCl buffers and that for C3u in 50 mM NaCl were deposited in the Protein Data Bank with the accession codes 4MRJ, 4MRK, and 4MRL, respectively. That for C3u in 137 mM NaCl had been deposited with the Protein Data Bank code 3MMQ.
Sedimentation coefficients s 20,w for the C3, C3b, C3c, and C3d crystal structures and for the C3, C3u, and C3b best fit scattering models were calculated directly from the atomic coordinates using the HYDROPRO shell-modeling program (44). The default value of 0.31 nm for the atomic element radius for all atoms was used to represent the hydration shell.
Surface Plasmon Resonance with Mutant C3d(A1032)-Sitedirected mutagenesis was performed using the QuikChange TM site-directed mutagenesis kit (Agilent Technologies). The C3d coding region was amplified by PCR using forward 5Ј-CTG-GATGAAACGGCGCAGTGGGAGAAG-3Ј and reverse 5Ј-CTTCTCCCACTGCGCCGTTTCATCCAG-3Ј primers that incorporated the E1032A change. The PCR products were treated with DpnI restriction enzyme to remove methylated parental DNA, and the plasmid was transformed into XL-1 blue-competent cells (Agilent Technologies). Following confirmation of the correct sequence by DNA sequencing (Eurofins, MWG Operon, London, UK), the mutant recombinant DNA was transformed using heat shock into E. coli BL21 cells for protein expression (Novagen, Merck). Surface plasmon resonance binding studies were performed on a Biacore X100 system (GE Healthcare). C3c was immobilized to the flow cell of a carboxylated dextran (CM5) research grade sensor chip via a standard amine coupling procedure according to the manufacturer's protocol. 10 g/ml C3c in 10 mM acetate buffer (pH 4.5) was injected over flow cell 2 until the appropriate level of response units for kinetic analyses was attained. A control was prepared identically on flow cell 1 but without immobilizing C3c. Binding and steady state analyses with wild type and mutant C3d were performed at 25°C using appropriate Biacore X100 wizards at a flow rate of 30 l/min. Regeneration after each run was achieved by pulsing 10 mM acetate buffer, 2 M NaCl (pH 7.1) across both flow cells twice for 30 s. The analyte was passed over the chip surface in 50 mM NaCl or 137 mM NaCl HEPES buffers. For the steady state analysis of wild type C3d binding to C3c, 0.23-1.55 mg/ml (7-45 M) C3d was passed over the sensor surface in 50 mM NaCl and 0.38 -2.27 mg/ml (11-66 M) in 137 mM NaCl, respectively. For the mutant C3d, 0.24 -0.8 mg/ml (7-23.1 M) in 50 and 137 mM NaCl buffers was injected over immobilized C3c.
RESULTS
Sedimentation Velocity Analyses-C3, C3u, C3b, C3c, and C3d were studied in concentration ranges between 1 and 10 M, these being comparable with the physiological plasma C3 concentration of 1.0 mg/ml (5.3 M) (2). During gel filtration, the five purified proteins C3, C3u, C3b, C3c, and C3d (see "Experimental Procedures") each eluted as a single symmetric peak in 137 mM buffer ( Fig. 2A). Before and after experiments, C3, C3u, and C3b each migrated as a single band in non-reducing SDS-PAGE and as two bands that correspond to their ␣ and  chains in reducing SDS-PAGE (Fig. 2B). C3c migrated as a single band in non-reducing SDS-PAGE and three bands that correspond to the cleaved ␣ chain and the  chain in reducing SDS-PAGE. C3d migrated as one band in both reducing and non-reducing SDS-PAGE.
Analytical ultracentrifugation is used to observe macromolecular sedimentation under high centrifugal force in order to determine their masses and shapes (45). The main advantage of ultracentrifugation is the detection of different species in a sample from the peaks in size distribution analyses c(s). Previous c(s) data for C3, C3u, and C3d in 50 and 137 mM NaCl buffers (16,46) were extended to include C3b and C3c in the same buffers. Good absorbance and interference boundary fits were obtained (Fig. 3, A, B, E, F), and new data for C3, C3u, and C3d confirmed previous results (Fig. 4, A-C). (i) In 137 mM NaCl buffer, C3b showed primarily a monomer peak with a mean s 20,w value of 7.40 S and a molecular mass of 179 kDa (Figs. 3C (M) and Fig. 4I). A small C3b dimer peak (2% of the total) was seen at 11.4 S with the expected molecular mass of 331 kDa. C3c showed a monomer peak at 6.53 S and 135 kDa and a small C3c dimer peak (6%) at 9. peak showed a concentration dependence (Fig. 3C), attributable to fast exchange with C3b dimers. C3 and C3u also showed this fast exchange (16). Upon extrapolation of the s 20,w values to zero concentration, the s 20,w 0 value for monomeric C3b was 7.60 S (Fig. 4I) with a mass of 183 kDa. A small C3b dimer (4%) was visible at 12.4 S with a mass of 326 kDa (Fig. 3C). The 0.19 S increase in 50 mM NaCl buffer suggested that C3b is more compact in low salt. The C3c monomer and dimer peaks were unchanged at 6.53 and 10.3 S (Fig. 3G). The estimated monomer-dimer K D values were comparable at 40 Ϯ 20 M for C3b and 50 Ϯ 15 M for C3c. C3, C3u, and C3d showed well defined monomer peaks in 137 mM NaCl buffer (Fig. 4, A-C). Our improved data analyses gave s 20,w values of 8.49, 8.03, and 3.03 S, respectively, in agreement with previous results (16,46). Dimers were seen at 12.1 S (3%), 11.9 S (8%), and 5.9 S (10%), respectively. For 137 mM NaCl, the estimated monomer-dimer K D values were 70 Ϯ 10 M for C3 and 45 Ϯ 15 M for C3u. In 50 mM NaCl, C3 and C3u showed concentration dependences as before (16). The extrapolated s 20,w 0 values were 8.50 S (C3) and 8.10 S (C3u) (Fig. 4) .64 S (C3c), and 6.23 S (C3d). The dimer proportions increased to 5% (C3), 15% (C3u), 14% (C3b), 7% (C3c), and 12% (C3d). The doubling of dimer formation to 14 -15% for C3u and C3b, in contrast to the almost unchanged dimer of 2-4% for C3 and 6 -7% for C3c, indicated the assignment of the fast dimerization site in C3u and C3b to the exposed TED (C3d) domain.
X-ray and Neutron Scattering Analyses-Small angle x-ray and neutron scattering is a diffraction technique that determines overall macromolecular structures in solution (19). X-rays were used to examine the hydrated proteins in 50 and 137 mM NaCl buffers. Neutrons were used to examine their unhydrated protein structures in 137 mM NaCl 2 H 2 O buffer (33). The x-ray and neutron Guinier analyses showed high quality linear plots in two distinct Q ranges. The lowest Q values gave the R g values, which monitor the degree of elongation (Fig. 5, A-E and J-N). At larger Q values, the R xs values measure the mean cross-sectional dimensions (Fig. 5, F-I and O-R). The Guinier I(0)/c values are proportional to the relative molecular masses (39, 47). Table 1.
The distance distribution function P(r) leads to the overall length (L) at large r and the most frequent interatomic distance M. The R g and I(0)/c values from P(r) agreed with those from the Guinier analyses (Table 1). The L values for C3, C3u, C3b, C3c, and C3d in 137 mM NaCl were 16, 16, 16, 14, and 6 nm, respectively (Fig. 7). In 50 mM NaCl, the L values for C3, C3u, and C3b increased from 16 nm at 0.5 mg/ml to FIGURE 5. Guinier R g and R xs analyses for the five proteins. A-I, in the x-ray analyses, the C3, C3u, and C3b concentrations in 137 mM NaCl buffer ranged between 0.5 and 1.5 mg/ml (from bottom to top, as shown); those for C3c ranged between 0.4 and 1.1 mg/ml; and those for C3d ranged between 0.5 and 1.4 mg/ml. The Q ranges used for the R g fits were 0.18 -0.30 nm Ϫ1 for C3; 0.14 -0.30 nm Ϫ1 for C3u, C3b, and C3c; and 0.35-0.55 nm Ϫ1 for C3d. Those for the R xs fits were 0.35-0.50 nm Ϫ1 for all four proteins. The filled circles represent the data used to determine the R g and R xs values. Their values were measured within satisfactory Q⅐R g and Q⅐R xs ranges, as shown. J-R, in the neutron analyses, the C3 concentrations were 0.75 and 0.47 mg/ml; those for C3u were 0.82 and 0.52 mg/ml; those for C3b were 0.52 and 0.34 mg/ml; those for C3c were 0.91 and 0.53 mg/ml; and those for C3d were 0.55 and 0.34 mg/ml. The D22 data sets correspond to 2 H 2 O buffer. The Q ranges used for the R g fits were 0.16 -0.30 nm Ϫ1 for C3, C3u, and C3b; 0.2-0.35 nm Ϫ1 for C3c; and 0.4 -0.6 nm Ϫ1 for C3d. Those for the R xs fits were 0.35-0.50 nm Ϫ1 for all four proteins. The filled circles represent the data used to determine the R g and R xs values. 20 nm at 1.5 mg/ml and from 6 to 12 nm for C3d, reflecting dimerization (Fig. 7). C3c showed no change in L. The neutron P(r) analyses gave L values similar to the x-ray values or slightly less (Fig. 7). The M values of C3, C3u, and C3b were all 5 nm, C3c was 4.5 nm, and C3d was 2.5 nm, these being unchanged with concentration or buffer by x-rays or neutrons.
Comparison with known C3, C3u, C3b, C3c, and C3d Structures-The five molecular structures were compared with their three scattering curves in different buffers (Fig. 1 and Table 2A). For 137 mM NaCl, the C3, C3c, and C3d crystal structures gave good fits (Fig. 8). The R-factors (goodness of fit) were low, at 3.2-3.9% for x-rays and 3.9 -5.1% for neutrons (Table 2A). The four C3b crystal structures gave R-factors of 3.3 Ϯ 0.1% (x-rays) and 6.6 Ϯ 0.4% (neutrons). The C3u solution structure gave R-factors of 3.3% (x-rays) and 4.5% (neutrons). For 50 mM NaCl, because ultracentrifugation showed that dimers were present, it was necessary to extrapolate the full curves to zero concentration to eliminate the effect of dimers. The resulting C3 and C3c R-factors were 3.8 -3.9%, similar to those in 137 mM NaCl, and C3u gave a similar R-factor of 3.3%. Of importance was the much reduced R-factor of 1.5% for C3b, showing that the curve fit was improved. Other comparisons using R g values calculated from molecular structures were less precise, being sensitive to small amounts of dimer ( Fig. 6 and Table 1). For all five proteins except one, the experimental R g was between 0.04 and 0.59 nm larger than the R g values from molecular structures (Table 2A), suggesting that trace amounts of dimers perturbed the Guinier fits. The exception was C3d in 50 mM NaCl, where the experimental R g of 2.99 -3.18 nm showed a large difference from the crystal structure value of 2.11 nm, indicating that dimers were still present.
The s 20,w 0 calculations from molecular structures suggested that C3u and C3b were more compact in 50 mM NaCl and more extended in 137 mM NaCl, although the precision of the calculation was only Ϯ 0.21 S (48). C3u with its extended TED-MG1 domain pair arrangement gave a solution structural value of 7.98 S (Table 2A), in better agreement with the experimental s 20,w 0 of 8.03 S in 137 mM NaCl than that of 8.10 S in 50 mM NaCl (Table 1) Atomistic Modeling Searches for C3u, C3b, and C3-Constrained modeling searches on C3u and C3b were performed to identify the position of TED-CUB relative to C3c that best fitted the four scattering curves, starting from a library of 4650 randomized structures (see "Experimental Procedures"). When the goodness of fit R-factors were compared with the 4650 modeled R g values (Fig. 9, A-F), a single minimum was observed in all six cases, showing that a single family of conformers best fitted the data. The lowest R-factors corresponded to models with R g values similar to the experiment. In all six fits, the best fit R g values were higher for C3u compared with C3b, indicating that C3u is more elongated than C3b. (i) The x-ray fits for C3u in 137 mM NaCl replicated earlier results (16). The addition of the new neutron fits showed that five of the nine best fit models were identical to the 10 x-ray best fit models. The R-factor was low at 3.3% (x-ray) and 4.4% (neutron) ( Table 2, B and C). The calculated and experimental I(Q) and P(r) curves showed good visual agreement (Fig. 10A). The calculated s 20,w 0 value of 7.98 S agreed well with the experimental value of 8.03 S. The best fit models showed that the TED and MG1 domains were well separated by 6.0 Ϯ 0.6 nm (x-ray) and 6.2 Ϯ 0.6 nm (neutron) between their centers of mass. (ii) For C3u in 50 mM NaCl, the x-ray fits gave a low R-factor of 3.1% (Table 2D) with good visual agreement (Fig. 10C). The calculated s 20,w 0 value of 7.88 S agreed well with the experimental value of 8.10 S. Here the best fit models showed a TED-MG1 separation of 4.3 Ϯ 0.4 nm, within error of that of 3.8 Ϯ 0.4 nm in the C3b crystal structures. (iii) For C3b in 137 mM NaCl, five of the 9 -10 best fit models from the x-ray and neutron fits were identical. Good visual curve fits were obtained (Fig. 10B). The TABLE 2 Scattering and sedimentation parameters for C3, C3u, C3b, C3c, and C3d calculated from molecular modeling fits best fit C3b model has R-factors of 2.2% (x-ray) and 5.7% (neutron) ( Table 2, B and C), these being improved compared with those from the C3b crystal structures. The calculated s 20,w 0 value of 7.46 S agreed well with the experimental value of 7.40 S. The TED-MG1 separation was 5.2 Ϯ 0.5 nm (x-ray) and 5.2 Ϯ 0.5 nm (neutron), this being greater than in the C3b crystal structures. Our x-ray and neutron fits strongly indicate a single conformation for C3b, unlike earlier C3b modeling in 200 mM NaCl that suggested two different conformations (18). (iv) For C3b in 50 mM NaCl, the x-ray best fit C3b model had a low R-factor of 2.8% ( Fig. 10D and Table 2D). The calculated s 20,w 0 value of 7.60 S agreed with the experimental value of 7.60 S. The TED-MG1 separation was 4.3 Ϯ 0.4 nm, within error of the C3b crystal structures.
Control calculations showed that constrained modeling was able to replicate the C3 crystal structure. In fit searches, each of the TED and CUB domains was varied separately relative to the C3c region. Single R-factor minima were observed (Fig. 9, G-I).
The best fit R g values were lower for C3 compared with C3u and C3b, indicating that C3 has the most compact structure. (i) The x-ray fits for C3 in 137 mM NaCl replicated earlier results (Fig. 11A) (16). Four of the eight neutron best fit models were the same as the x-ray best fit models. The R-factor was low at 3.1% (x-ray) and 5.1% (neutron) ( Table 2, B and C). The calculated s 20,w 0 value of 8.57 S agreed with the experimental value of 8.49 S. (ii) X-ray fits for C3 in 50 mM NaCl showed that four of the eight best fit models were the same as in 137 mM NaCl, and the x-ray R-factor was 3.5% ( Fig. 11B and Table 2D). The calculated s 20,w 0 value of 8.57 S agreed well with the experimental value of 8.50 S.
Mutagenesis of the TED-MG1 Binding Interface-In order to verify that the structural changes in C3b and C3u between 50 and 137 mM NaCl were attributable to the Arg 102 -Glu 1032 salt bridge seen in the C3b crystal structure in 50 mM NaCl, this interaction was validated by surface plasmon resonance. When wild-type C3d was injected over immobilized C3c in 50 mM NaCl, a significant response was detected, leading to a K D value of 51 M (Fig. 12A). In 137 mM NaCl, C3d binding to immobilized C3c was much weakened, with the K D value estimated to be in the millimolar range. This indicated the loss of this salt bridge in 137 mM NaCl. The reversed experiment with C3c injected over immobilized C3d showed the same outcome (not shown). To test whether the Arg 102 -Glu 1032 salt bridge itself accounted for the observed separation of the TED-MG1 interaction in C3b and C3u in 137 mM NaCl, but not in 50 mM NaCl (above), the C3d E1032A mutant was used. Mutant C3d bound 10-fold less to C3c in 50 mM NaCl, and binding was not observed in 137 mM NaCl (Fig. 12B). The reversed experiment with the immobilized C3d mutant also showed no binding to C3c in either buffer.
DISCUSSION
In physiological 137 mM NaCl buffer, our solution structures for C3b and C3u showed that the TED domain in C3b and C3u was separated by as much as 6 nm from the MG1 domain in the C3c region (Fig. 13, C and D). This result changes our understanding of the TED domain. It had been frequently assumed that the TED and MG1 domains of C3b remain connected to form a compact structure in solution (4); however, this structure is now seen to be an artifact of crystal packing in nonphysiological conditions of 50 mM NaCl. Gratifyingly, our C3b and C3u solution structures in 50 mM NaCl (Fig. 13, A and B) agreed well with the C3b crystal structures, all four of which were crystallized in 50 mM NaCl buffer (Fig. 13E). This TED conformational change in 137 mM NaCl means that, mechanistically, C3b is now best viewed as a more reactive molecule than previously believed, with a mobile thioester group that readily binds to antigenic surfaces. If the TED-MG1 interaction was rigidly held together in C3b, this would restrict the ability of the thioester to bind to appropriate antigenic surfaces. These results may indicate a novel structural basis for the future design of improved complement inhibitors targeted against C3b in which the TED-MG1 domains become more strongly linked.
C3b forms a connected TED-MG1 structure in its regulatory complex with the Factor H SCR-1/4 (short complement regulator 1/4) domains (13). The stabilization of a less active and compact C3b conformation by Factor H is part of its regulatory mechanism, followed by the cleavage of C3b at its CUB domain by Factor I. The Arg 102 -Glu 1032 salt bridge in C3b is required for this complex (Fig. 13F). Similar residues in the human C4 sequence occur at Arg 63 and Glu 1032 but not in the sequences of human C5 or ␣ 2 -macroglobulin. Presumably, this salt bridge is conserved when a regulatory control mechanism became essential. In fact, separated TED-MG1 structures were seen to be physiologically important in the crystal structure of complement C5b (a C3b homologue) in complex with complement C6, which also showed a TED-MG1 separation of 5.4 nm (Fig. 13G) (20,21). Separated TED-MG1 structures were also seen for methylamine-induced (active) ␣ 2 -macroglobulin (19). ␣ 2 -Macroglobulin is a tetramer of C3b-like structures that is formed through contacts between pairs of TED domains and pairs of MG3-MG4 domains. The TED-MG1 separation in ␣ 2 -macroglobulin is 5.6 -5.7 nm (Fig. 13H), similar to that of 5.2-6.2 nm in our C3b and C3u structures in 137 mM NaCl. An extended C3b structure was also observed in its complex with bacterial Efb-c, where the TED-MG1 separation was about 4.9 nm (18). For C3u, a further need for the Arg 102 -Glu 1032 salt bridge may be related to the low intracellular ϳ10 mM NaCl concentration within the hepatic cells that synthesize the complement proteins. Low salt concentrations would favor the functionally inactive structure if intracellular C3u is inad-vertently formed, so that complement is not accidentally activated during its secretion.
Our finding that the TED-MG1 domains are separated in C3b in 137 mM NaCl explains the significance of the R102G polymorphism that distinguishes the major C3S and C3F allo- FIGURE 10. Scattering modeling fits for C3u and C3b in three buffers. A, the x-ray and neutron fits for C3u at 0.82-1.0 mg/ml in 137 mM NaCl buffer in light and heavy water are shown in blue and green, respectively. B, the x-ray and neutron fits for C3b at 0.52-1.0 mg/ml in 137 mM NaCl buffer in light and heavy water are shown in blue and green, respectively. C, the C3u x-ray curve in 50 mM NaCl buffer in light water was extrapolated to zero concentration. Subtraction of the fitted C3u x-ray curve in 137 mM NaCl buffer revealed a peak at 0.86 nm Ϫ1 . D, the C3b x-ray curve in 50 mM NaCl buffer in light water was extrapolated to zero concentration. Subtraction of the fitted C3b x-ray curve in 137 mM NaCl buffer revealed a peak at 1.14 nm Ϫ1 . The insets show the experimental (continuous) and modeled (dashed) x-ray P(r) curves. The right-hand panels show the 4 -6 x-ray best fitted superimposed structures in the same orientation as Fig. 1, with the best fit TED domain shown in crimson.
types. First, the loss of the Arg 102 -Glu 1032 salt bridge in 137 mM NaCl means that the positively charged Arg 102 side chain becomes fully solvent-exposed at the surface of C3S. This now explains why C3S migrates more slowly toward the anode than C3F during non-reducing SDS-PAGE; the crystal structure had predicted that Arg 102 would be buried in solution. Second, our surface plasmon resonance studies showed that the Arg 102 -Glu 1032 salt bridge is crucial for controlling the TED-MG1 separation (Fig. 12), and this is important for the regulatory C3b-Factor H SCR-1/4 interaction, which is stabilized by this salt bridge. The loss of this salt bridge in the C3F allotype is associated with age-related macular degeneration, leading to deposition of complement C3 in drusen, and with atypical hemolytic uraemic syndrome, leading to endothelial tissue damage in the kidneys.
Excessive complement activation is involved in both pathogeneses, leading to greater inflammatory damage.
This study was made possible by advances in atomistic scattering modeling that successfully identified the compact and extended structures for C3u and C3b. Analytical ultracentrifugation clarified weak dimer formation in the scattering data that required correction. The advent of high throughput scattering data collection permitted detailed data analyses for all five C3 forms in three buffers, including sufficient data to permit the curves to be extrapolated to zero concentration. Complementary x-ray and neutron scattering curves were used for modeling (22). The quality of the curve fits were similar or improved compared with other complement proteins and antibodies (47). By scattering, the crucial TED structural movement in C3u and C3b between 50 and 137 mM NaCl was most FIGURE 11. Scattering modeling fits for C3 in three buffers. A, the x-ray and neutron fits for C3 at 0.75-1.2 mg/ml in 137 mM NaCl buffer in light and heavy water are shown in blue and green, respectively. B, x-ray fit for C3 in 50 mM NaCl buffer in light water extrapolated to zero concentration. C, the best fit C3 structure is shown. For more details, see the legend to Fig. 8. FIGURE 12. Analysis of the Arg 102 -Glu 1032 salt bridge using surface plasmon resonance. A and B, wild type C3d (E1032) and mutant C3d (A1032) analytes were injected over amine-coupled immobilized C3c as ligand in 50 mM NaCl (red) and 137 mM NaCl (black) buffers. C, corresponding K D fits for 50 mM NaCl (red) and 137 mM NaCl (black). C3d (E1032) gave a K D value of 51 M in 50 mM NaCl buffer.
clearly seen by a reproducible inflection following subtraction of the 137 and 50 mM NaCl curve fits (Fig. 10, C and D). This study is thus an advance on previous x-ray scattering of C3u and C3b (16,18) that showed that C3u and C3b were more extended than the C3b crystal structures but did not explain the difference. The earlier use of 5% glycerol may have perturbed the C3b conformation through the promotion of excessive hydrogen bonding with glycerol (17,49). As an alternative method, electron microscopy showed the existence of more extended or more compact structures for negatively stained C3b and C3u in vacuo; however, the context of these two structural forms was not identified (9).
Dimer formation was an unwanted feature of the C3 proteins but required consideration for accurate modeling analyses. Ultracentrifugation showed that C3d exhibited a monomerdimer-tetramer equilibrium in 50 mM NaCl buffer but was monomeric in 137 mM NaCl (46). Two types of dimers occurred for C3 and C3u in 50 mM NaCl (16). The faster equilibrium was attributable to C3d-C3d (TED-TED) contacts, whereas the slower one was attributable to C3c-C3c contacts. The present study with five C3 proteins extended these observations. The TED domain in C3, C3u, and C3b was linked with pronounced concentration dependences only in 50 mM NaCl (Figs. 4 and 6). The resulting monomer-dimer dissociation constant K D values were 40 -70 M in both 50 and 137 mM NaCl for C3, C3b, C3u, and C3c. The K D value for C3d self-association was 25 M in 50 mM NaCl and was not observable in 137 mM NaCl. These K D values showed that these monomer-dimer equilibria were not physiologically important at typical protein concentrations of about 5 M. Nonetheless, dimer formation in the C3 proteins is important to consider in experimental studies if the buffer conditions or sample concentrations differ from their physiological ranges. | 9,647.4 | 2014-10-01T00:00:00.000 | [
"Biology"
] |
How memory effects, check dams, and channel geometry control erosion and deposition by debris flows
Debris flows can grow greatly in size and hazardous potential by eroding bed and bank material, but effective hazard assessment and mitigation is currently hampered by limited understanding of erosion and deposition dynamics. We have collected high-resolution pre- and post-flow topography for 6 debris flows over a 3 km long unconsolidated reach of the Illgraben channel in the Swiss Alps with drone-based photogrammetry. We show that the spatio-temporal patterns of erosion and deposition in debris-flow torrents are highly variable and dynamic. Check dams strongly control the spatial patterns of erosion and deposition. We identify a memory effect where erosion is strong at locations of strong deposition during previous flows and vice versa. Large sediment inputs from subcatchments initially result in new channel erosion through the subcatchment deposits and simultaneous upstream deposition, likely as a result of backwater effects. It is generally believed that erosion increases with debris-flow magnitude, but we show that there is a limit to debris-flow bulking set by channel geometry. These findings provide key guidelines for flow volume forecasting, emphasizing the importance of memory effects and the need to resolve both erosion and deposition in predictive models.
Debris flows are rapid destructive masses of soil, rock and water that are common natural hazards in mountainous regions worldwide [1][2][3] . The continued expansion of human populations into mountainous regions has greatly increased the risk associated with debris flows 4 , and hazards are expected to increase as a result of global climate change [5][6][7] . Debris flows are typically generated by small landslides or runoff in mountain catchments, after which they flow down onto (inhabited) alluvial fans and valley floors 8 . They may entrain large amounts of bed sediment when rushing down mountainsides, possibly increasing in size by several orders of magnitude [9][10][11][12][13][14] . Estimation of flow volumes is critical for assessment of flow hazard and design of mitigation measures, because flow volume is the prime control on flow velocity, peak discharge, inundation area [15][16][17] and the number of fatalities 4 . Unfortunately, limited understanding of debris-flow erosion and deposition dynamics currently hampers debris-flow volume estimation and hazard assessment and mitigation.
Resolving debris-flow deposition is often ambiguous in numerical models [18][19][20][21][22] , and few models have been explicitly tested for their performance in describing deposition due to debris flows. There has been a recent increase in the number of numerical models incorporating erosion [21][22][23] , but the inconsistency in erosion rate equations as a result of a lack of a unified theory still results in a disparity of model outcomes. Much of our understanding of debris-flow erosion stems from theoretical considerations 24 and physical scale experiments 13,[25][26][27][28] , while there is a relative scarcity of field data 11,12,[29][30][31][32] as a result of the infrequent nature of debris flows, the rough terrain in which they occur, and the high time and cost demands of field measurements. Analysis of field data is often hampered by unknown boundary conditions and material properties 11,12 , and is often based on local point or cross-section measurements 12,31 , single time-steps 32 , and measurements are typically only available for small areas 29,30 . Debris-flow torrents are often complex systems, with spatially variable channel geometries and bed materials, sediment contributing subcatchments, and in inhabited areas numerous check dams which locally stabilize the channel. To extrapolate findings from theoretical considerations and physical-scale experiments to the complex field environment and to perform accurate flow-volume estimates for hazard assessment and www.nature.com/scientificreports/ mitigation, we need to develop a detailed understanding of the largely unknown spatio-temporal erosion-deposition patterns and processes, and associated flow bulking in debris-flow torrents.
Here we present a drone-based time-and cost-effective method for monitoring topographic changes over long reaches of debris-flow channels. We present erosion and deposition patterns in the Illgraben torrent in the Swiss Alps as a result of six debris flows and subcatchment activity over a 3.3 km long unconsolidated reach with check dams, and compare these erosion and deposition patterns with in-situ flow measurements. Our work shows that erosion and deposition patterns can be highly heterogeneous, and sheds light on four crucial but generally overlooked aspects that control erosion and deposition in debris-flow torrents: (1) check dams; (2) memory effects; (3) effects of subcatchment inputs; (4) and channel geometry.
Illgraben catchment and torrent
The Illgraben catchment in the Swiss Alps (Fig. 1) has a long history of debris flows 33,34 , and an extremely high debris-flow frequency of approximately 5 debris flows and debris floods per year since 2000, annually transporting ~ 100,000 m 3 of debris to the Rhone River 35 . The catchment extends from the summit of the Illhorn mountain (elevation 2,716 m a.s.l.) to the Rhone River on the valley floor (610 m a.s.l.) 36 . The catchment has a total area of 8.9 km 2 , but all debris flows originate from a 4.6 km 2 subcatchment. This subcatchment is composed of massive dolomites on its northwest wall and layers of quartzite, conglomerates, and calcareous sedimentary rocks on its southeast valley wall 37 . Debris flows are generally triggered by intense rainfall during summer storms between May and October.
The channel stretching from the Illhorn mountain to the Rhone River has a length of ~ 6.5 km, of which the distal 4.8 km hosts 28 check dams which cause vertical drops of several meters along the channel bed. The reach with check dams has an unconsolidated bed 29 . For the most downstream 2 km the channel traverses a large alluvial fan. Just upstream of the point where the channel leaves the catchment a small (0.1 km 2 ) but recently very active subcatchment is present, the Sagenschleif, which currently contributes large quantities of sediment to the channel.
A few tens of meters upstream of the confluence with the Rhone River an automated observation station is operated by the Swiss Federal Institute for Forest Snow and Landscape Research (WSL), which records flowfront velocity, flow depth, and normal and shear forces, and collects imagery 36,37 . From these observations, it is possible to calculate the front discharge of a debris flow and to estimate its volume 38 .
Channel topography measurements
We have measured before and after channel-bed topography of the downstream 3.3 km of the channel through drone-based photogrammetry for six debris-flow events that occurred between December 2018 and July 2019 ( Table 1). These flows had maximum flow depths ranging from 0.6 to 2.6 m, flow front velocities ranging from 0.9 to 5.6 m/s, front discharges ranging from 4 to 122 m 3 /s, and volumes ranging from 3,000 to 75,000 m 3 , at the observation station (Table 1; Suppl.
Spatio-temporal patterns of erosion and deposition
The topographic changes within the Illgraben channel reveal a strong spatio-temporal heterogeneity in erosion and deposition patterns, which are also strongly affected by the check dams (Fig. 2). The flow from December 2018 shows a sharp transition from net erosion to net deposition around 1,600 m upstream of check dam 29 (Fig. 2a). The check dams cause a strong saw-tooth pattern of deposition, with most deposition just downstream of the check dams after which deposition gradually decreases in the downstream direction towards approximately Table 1. Flow characteristics of the studied debris flows measured at the observation station. Event volumes and front discharge were estimated using the methods described in 38 Figure 1) continued eroding the Sagenschleif deposits, and was also associated with sediment deposition upstream of the Sagenschleif (Fig. 2c). Interestingly, most deposition took place directly upstream of the Sagenschleif, while further upstream, where maximum deposition took place during the previous flows, there is now net erosion (~ 2,900 m upstream of check dam 29). Downstream of the Sagenschleif there is a reach of > 1,000 m over which substantial erosion and deposition simultaneously took place. This was the result of the large size of this flow, resulting in thalweg erosion and simultaneous deposition by the formation of levees and other deposits on top of overbanks (Fig. 3j- The flows of June 2 and 3, of which the flow of June 2 was of considerable size, caused relatively minor channel bed changes (Fig. 2d). These flows further eroded the Sagenschleif deposits, and also removed the backwater deposits that formed behind the Sagenschleif during the previous flows. Directly downstream of the Sagenschleif there was limited mixed erosion and deposition, shifting to net erosion around 2000 m upstream of check dam 29. Net deposition occurred downstream of 700 m upstream of check dam 29. Deposition here seems to counterbalance the erosion of the previous 3 flows that were net erosive in this reach.
Effects of check dams on erosion and deposition patterns
In the European Alps and many other mountainous areas, most debris-flow channels that pose a threat to communities and infrastructure have check dams. Check dams are generally applied to debris-flow channels to prevent vertical and lateral channel erosion and movement [39][40][41] . In the Illgraben channel the check dam surfaces are in line with the bed surface, and act as bed sills that provide a locally fixed base level. As such, the check dams exert a strong control on the spatio-temporal erosion and deposition trends observed in the Illgraben channel (Figs. 2, 3), with the most pronounced changes generally occurring directly downstream of a check dam and moving to approximately zero at the downstream check dam. Furthermore, in the study reach bed erosion was more pronounced than bank erosion, which may be the result of the check dams preventing lateral channel migration and stabilizing banks.
Severe scour has often been identified downstream of check dams, providing an engineering and safety problem 39 . Our data shows that deposition may also be very large downstream of a check dam, possibly buffering the total amount of scour over a number of events. Because of the fixed bed level at check dams, erosion automatically implies a reduction in the effective slope of the channel within the check dams. Such as reduction in channel slope then increasing the likelihood of deposition in a subsequent flow and vice versa. In contrast, erosion in unconsolidated channels without check dams lowers the channel bed more uniformly, and will thus lead to far less or no channel slope reduction. This suggests that torrents with check dams may be more prone to memory effects.
Memory effect controls on erosion and deposition patterns
Previously many flow parameters including flow volume 42 , flow depth 29,31 , flow velocity 43,44 , discharge 45 , bed slope 12,46 , shear stress 29,31,47 , grain collisional stress [48][49][50] and bed wetness 25 , have been shown to positively correlate with erosion rate. Although we do not doubt that these parameters do indeed affect erosion and deposition by debris flows, our results do suggest that memory effects may play a similarly important role in nature.
Our observations show that erosion and deposition in a debris flow strongly depend on the erosion and deposition by previous debris flows and subcatchment activity. Localities where substantial deposition occurred have a high probability of substantial erosion during subsequent flow(s) and vice versa. This might be the result of the channel-bed deviation around an equilibrium profile, partly affected by check dams as explained above. Alternatively, for some of the localities where we observe a memory effect, namely removal of deposits by the 21 June 2019 debris flow, the observed patterns might result from the short residence time of 11 days between this and the previous flow. Therefore, the removed deposits may still have had an elevated pore-water content and were easily eroded 25 . Still, our observations show that the channel carved by the 21 June 2019 flow does differ from that present before the 10 June 2019 flows. In addition, the 2 and 3 July flows did not substantially erode or deposit despite their relatively large flow depth and velocity, likely as a result of the already strongly scoured thalweg by the very large previous flow of June 21.
Furthermore, large sediment input from subcatchments that block a channel affect erosion and deposition patterns during multiple subsequent flows. We find net erosion through these deposits as a new channel is carved, but at same time upstream deposition as a result of backwater. In this particular example the backwater deposits are removed after a few flows, likely because around that time an efficient debris-conveyance corridor through the subcatchments deposits has been re-established.
A natural limit to debris flow bulking
The positive correlation between flow parameters such as volume, shear stress, and flow depth imply that there is a runaway mechanism that strongly promotes bulking during debris flow. As a debris flow erodes bed material it grows in volume, flow depth and shear stress thereby promoting even more erosion and bulking, potentially Scientific RepoRtS | (2020) 10:14024 | https://doi.org/10.1038/s41598-020-71016-8 www.nature.com/scientificreports/ leading to orders of magnitude increases in flow volume. Indeed, such behaviour has been observed in nature and is reproduced by numerical models 14,23 . Our data, however, shows that there is a natural limit to this bulking set by channel geometry. The large flow of June 21 illustrates that when a flow exceeds the channel capacity it starts forming levee and overbank deposits. This can lead to net deposition, despite substantial thalweg erosion, thereby limiting or even counterbalancing further flow bulking. Schürch et al. 29 previously found that in the Illgraben channel erosion becomes more likely with increased flow depth and shear stress, although a broad range of outcomes is possible at any given flow depth or shear stress. They further found that the transition from deposition to erosion typically occurs around a flow depth of 1.5 m, explaining the observed pattern of erosion of the deep channel thalweg and the simultaneous deposition on overbanks where flow depths are relatively small and do not exceed 1.5 m. This finding implies that channel geometry may set a limit to the maximum size that a debris flow might obtain through bulking, as material will start to deposit once the flow reaches the overbanks. Because channel geometry may also vary along a stream and over time, we can expect contrasting spatio-temporal responses. These results emphasize that understanding and resolving both erosion and deposition processes and rates is of equal importance for flow bulking and volume estimates.
Towards estimating bulking of debris flows in nature
The importance of incorporating debris-flow bulking in hazard assessments has been increasingly recognised in recent years, there has been an increasing number of studies on debris-flow erosion, and erosion equations are being incorporated into numerical models [21][22][23] . Our observations reveal critical but generally overlooked phenomena controlling erosion in debris flows torrents. These results call for more high-resolution spatio-temporal datasets of debris-flow erosion and deposition, stress that check dams should be taken into account for flow bulking estimation, call for further investigation into memory effects, and stress that we should not only strive to understand and predict erosion processes but should understand deposition processes as well 51 , before we can accurately estimate and model debris-flow entrainment and bulking. Where erosion is increasingly incorporated into models, deposition processes have to date been largely overlooked.
Methods
Point cloud generation. We have generated digital elevation models (DEMs) (5 cm ground sampling distance) and orthomosaics (2.5 cm ground sampling distance) of the channel bed with drone-based photogrammetry. We used a DJI Mavic Pro 2, with has a 1″ CMOS sensor on-board collecting raw imagery with a resolution of 20 megapixels. We collected imagery with a side overlap of 80% and a forward overlap of 70%. We took nadir images from both channel banks and the middle of the channel and images with a 25° off-nadir camera pitch over the middle of the channel looking in an upstream direction to minimize doming effects 52 . Images were obtained at an altitude of 100 m above ground level resulting in a ground sampling distance of 2.5 cm. The total survey contained 900-1,000 images. We used anthropogenic and natural terrain features as ground control points (gcps), such as manholes, road surface marks, and cobbles and boulders. Because some boulders became buried or removed over time, we used 66 gcps for the first time step in November 2018 and 51 gcps for the last time step in July 2019. Point clouds were constructed with Agisoft Metashape 1.5. Point cloud accuracy at the gcps as reported by Metashape ranged between 0.21 and 0.27 m (Suppl. Table 1), and are substantially smaller than the typical observed erosion and deposition amounts (Figs. 2, 3).
Point cloud filtering and DEM generation.
To improve accuracy of channel-bed topography we removed erroneous points and overhanging vegetation from the dense point clouds before creating the raster DEM using LAStools (rapidlasso GmbH). The filter procedure has two main steps: remove low noise and filter overhanging vegetation, but otherwise aims to retain all natural detail in the channel and avoids clipping at steep sections at the channels edge. Low noise points are typical for dense point clouds generated using UAV photogrammetry. To filter these points, we removed points more than 0.1 m below a smoothed 20th height percentile surface with at step size of 0.5 m. Overhanging vegetation was removed by classifying ground points using the lasground functionality in LAStools with 'ultra fine' settings. This setting only effectively removes overhanging and sparse vegetation in the channel, but retains most of the fine details in the channel at the expense of including dense vegetation in geomorphologically inactive areas which were not of interest to our analysis. Filtered points were rasterized into a DEM with a ground sampling distance of 5 cm.
Erosion and deposition quantification. Erosion and deposition as a result of debris-flow activity was quantified as follows (Suppl. Figure 2). Initially, we generated DEMs of difference by subtracting the pre-flow DEM from the post-flow DEM. Subsequently, we used the mudlines and levees left behind by the debris flow to manually digitize the extent of each debris flow. We used this extent to clip out the DEMs of difference, such that we only consider topographic changes in the area affected by a debris flow. This method yields a conservative estimate for erosion because it includes deposition in the falling limb of the flow hydrograph 31 . In the Illgraben torrent debris flows substantially alter channel bed topography, while floods hardly affect the channel bed-their geomorphic work is generally limited to reworking of the channel bed and winnowing of fines 31 . We therefore assume that the measured changes are largely attributed to debris-flow activity. We extracted erosion volume, deposition volume, and net volume changes for segments of 10 m along channel (Suppl. Figure 2). We did exclude areas within 2 m of check dams, because tiny (~ 20 cm) offsets between the DEMs could otherwise generate large, but incorrect, erosion or deposition volumes. | 4,401.6 | 2020-08-20T00:00:00.000 | [
"Geology"
] |
Transition Metal Catalyst‐Free, Base‐Promoted 1,2‐Additions of Polyfluorophenylboronates to Aldehydes and Ketones
Abstract A novel protocol for the transition metal‐free 1,2‐addition of polyfluoroaryl boronate esters to aldehydes and ketones is reported, which provides secondary alcohols, tertiary alcohols, and ketones. Control experiments and DFT calculations indicate that both the ortho‐F substituents on the polyfluorophenyl boronates and the counterion K+ in the carbonate base are critical. The distinguishing features of this procedure include the employment of commercially available starting materials and the broad scope of the reaction with a wide variety of carbonyl compounds giving moderate to excellent yields. Intriguing structural features involving O−H⋅⋅⋅O and O−H⋅⋅⋅N hydrogen bonding, as well as arene‐perfluoroarene interactions, in this series of racemic polyfluoroaryl carbinols have also been addressed.
Introduction
Over the past few decades,the transition-metal-catalyzed 1,2-addition of organometallic reagents to the C = Ofunctionality of aldehydes and ketones has developed as au seful method for the synthesis of substituted secondary and tertiary alcohols. [1] Numerous reagents have been used for these reactions,i ncluding organomagnesium, [2] organozinc, [1,3] organolithium, [4] organosilane, [5] organostannane, [6] organocerium, [7] and organoboron compounds. [8] In particular,o rganoboronate reagents offer significant advantages such as air and moisture stability,l ow toxicity,g ood functional group tolerance,a nd availability. [8] In 1998, Miyaura and co-workers [9] first reported the addition of arylboronic acids to aldehydes using aR hc atalyst. In subsequent studies,o ther rhodium, [10] palladium, [11] platinum, [12] nickel, [13] copper, [14] iron, [15] cobalt, [16] and ruthenium [17] complexes have been developed as precatalysts for such reactions.H owever,t ransition metals can be expensive,t oxic, and difficult to remove completely from the corresponding product. At ransition metal-free strategy would be highly desirable for these useful transformations.T he reaction products for the addition of arylboronic acids to ketones,a fter hydrolysis,a re tertiary alcohols,which are important building blocks for the synthesis of pharmaceuticals,a grochemical compounds,a nd natural products. [18] However,t he nucleophilic addition of organometallic reagents to ketones can be challenging due to the inherent steric congestion around the carbonyl group,f requently resulting in the generation of products arising from side reactions such as reduction and aldol condensation. [19] Therefore,t he development of an efficient, general, and convenient protocol for the synthesis of tertiary alcohols is of considerable interest.
Moreover,a ni deal strategy to synthesize ketones,i mportant and ubiquitous structural motifs, [20] lies in the transition metal-catalyzed replacement of an aldehydes C(O)-H group with acarbon electrophile. [21] Recently,Zheng and co-workers demonstrated the direct functionalization of aldehyde CÀH bonds with aryl halides,using aprecious metal palladium catalyst, which has proven to be aviable method to generate the corresponding ketone products. [22] Polyfluoroarenes have gained extensive attention due to their important role in pharmaceutical, agrochemical, and advanced materials. [23] Thus,identifying practical and efficient concepts for the introduction of fluorine or fluorinated building blocks is highly desirable.Several studies have been reported regarding the polyfluorophenylation of aldehydes. Fore xample,i n1 999, Knochel and co-workers [24] used fluorinated aryl bromides to perform pentafluorophenylation of aldehydes (Scheme 1a). More recently,L am and coworkers [25] used ac opper catalyst (Scheme 1b)a nd Gu and co-workers [26] (Scheme 1b)u sed an N-heterocyclic carbene (NHC) organocatalyst to obtain fluorinated aryl carbinols using polyfluorophenyl trimethylsilane as an ucleophile for the addition to aldehydes.I n2 015, Huang and co-workers [27] (Scheme 1c)reported aMg-mediated polyfluoroaryl addition to aldehydes.Although some advancements in this field have been reported, these methods suffer from the requirement for highly flammable Grignard reagents,t ransition metals or NHC catalysts.Moreover,methods reported by Lam and coworkers and Gu and co-workers are limited to pentafluorophenyl trimethylsilane or 1,4-bis (trimethylsilyl) tetrafluorobenzene as substrates.
Recently,w er eported efficient methods to generate fluorinated arylboronic acid pinacol esters (Ar F -Bpin) via C-F borylation of fluoroarenes using NHC-ligated Ni complex [28a,b] and C-Cl borylation of Ar F -Cl using Pd catalyst under base free condition. [28c] Likewise,w er eported optimized conditions for the Suzuki-Miyaura cross-coupling reaction of Ar F -Bpin compounds with ArX (X = Br,I )u sing ac ombination of CuI and 1,10-phenanthroline as ac atalyst precursor. [28d] Furthermore,w er eported the palladium-catalyzed homocoupling of fluorinated arylboronates, [28e] and the copper-catalyzed oxidative cross-coupling of electron-deficient polyfluorophenyl boronate esters with terminal alkynes. [28f] We report herein the transition metal-free polyfluorophenylation of ketones and aldehydes with fluorinated aryl boronates,which provides aconvenient and novel strategy for the synthesis of alcohols and ketones.
Results and Discussion
Addition of arylboronic acids to aldehydes using transition metal catalysts has been well developed. We expected that the use of more Lewis acidic pentafluorophenyl-Bpin with abase would generate anucleophilic intermediate in the absence of at ransition metal. To verify our hypothesis,w e initially examined the reaction of pentafluorophenyl-Bpin (1a)and benzaldehyde (2a)asamodel reaction. As shown in Table 1, secondary alcohol 3a was observed as the addition product after hydrolysis when the mixture of 1a and 2a was heated in the presence of KOMe as the base (Table 1, entry 1). Encouraged by this first result, we screened the reaction parameters,i ncluding the base and the solvent, to improve the performance of the reaction. Theemployment of K 2 CO 3 as the base dramatically increased the yield to 92 % ( 16,17). However,u nder anhydrous conditions,t he transition metal-free Scheme 1. Approachest oaccess polyfluoroaryl carbinolsv ia the addition to aldehydes. polyfluorophenylation of benzaldehyde with pentafluorophenyl-Bpin is feasible and leads to high yields of the desired product.
Using these optimized conditions,weevaluated the scope and the limitations of this reaction. As shown in Table 2, as eries of aldehydes bearing electron-withdrawing or -donating substituents at the para-, meta-, or ortho-position all worked well with pentafluorophenyl-Bpin to give the desired products (3b-3k). Notably,f or reactions employing aldehydes bearing electron-donating groups,i ncreasing the reaction temperature to 80 8 8Cf or 48 hours was required to generate the corresponding products in acceptable yields.I t should be noted that reactions using 4-(diethoxymethyl)benzaldehyde resulted in cleavage of the diethoxymethyl group yielding 3l.Furthermore,this methodology could be successfully extended to more complex aldehydes,s uch as those incorporating naphthyl and pyridyl groups (3m and 3n). The structures of compounds 3f, 3l, 3m and 3n were unambiguously confirmed via single crystal X-ray analysis (vide infra).
After ab road range of aromatic aldehydes were examined, reactions with aliphatic aldehydes were investigated using the optimized conditions.G ratifyingly,a ll reactions proceeded smoothly to afford the corresponding products (3o-3q). Importantly,a ldehydes containing ester groups,w hich are well-known to be sensitive towards Grignard reagents,a lso afforded the desired alcohols in excellent yield (3r).
We then briefly investigated the scope using simple ketones (Table 3). When reactions were performed at 120 8 8C and for prolonged reaction times,the corresponding products were provided in moderate yields (3s-3u). Modest reaction yields were obtained when sterically hindered benzophenone and (2-fluorophenyl)(phenyl)methanone were used (3v-3w). Importantly,c yclohexanone proceeded to give the desired products in good yield (3x).
To gain further insight into the aforementioned reactions, several mechanistic studies were conducted. Ther eaction of 2a with pentafluorobenzene 5 under standard conditions was examined, yet 3a was not formed in any detectable amounts (Scheme 2a), indicating that the C-Bpin moiety is essential and deprotonation of the fluoroarene or nucleophilic attack at the fluoroarene by the base is not ap lausible pathway. Interestingly,f or the standard reaction between 1a and 2a, the yield dropped dramatically if 18-crown-6 ether and K 2 CO 3 were added (Scheme 2b). This experimental result indicates that the presence of the potassium ion plays acrucial role for the outcome of the reaction. Furthermore,i ft he reaction of 1a and 2a was performed in the presence of only ac atalytic amount of K 2 CO 3 (20 mol %) (Scheme 2c), reaction rates were reduced, and aw eek was required to produce 3a in good, isolated yield. This finding again indicates that the potassium ion (or the base) plays an important role in the reaction. Substituting ortho-fluorines by ortho-chlorines,u sing either C 6 Cl 5 Bpin or 2,6-dichlorophenyl-1-Bpin as substrates,d id not yield any product as shown by in situ GCMS studies.Likewise,2,3,4-trifluorophenylBpin and 3,4,5-trifluor-ophenylBpin substrates with only one or no ortho-fluorine substituent also led to no detectable product formation. The presence of an ortho-methoxy group on the aldehyde, however, did not inhibit the reaction.
Based on previous studies [28,29] and experimental observations,amechanism for the 1,2-addition of polyfluorophenylboronates to aryl aldehydes in the presence of K 2 CO 3 as base is proposed, as shown in Scheme 3. K 2 CO 3 interacts with the Lewis-acidic Bpin moiety of substrate 1 to generate base adduct A,w hich weakens the carbon-boron bond and ultimately cleaves the BÀCb ond along with attachment of apotassium cation to the aryl group.The resulting Ar F À anion adduct B undergoes nucleophilic attack at the aldehyde carbon atom of substrate 2 to generate methanolate C.T he methanolate oxygen atom then attacks the electrophilic Bpin group to obtain compound D.T ransfer of K 2 CO 3 from intermediate D to the boron atom of the more Lewis-acidic polyfluorophenyl-Bpin 1 finally closes the cycle and regenerates complex A.T hus,t he primary reaction product is the O-borylated addition product E,w hich was detected by HRMS and NMR spectroscopy for the perfluorinated derivative (Supporting Information, section VIII).
To corroborate this mechanism, adetailed DFT study was performed on the model 1,2-addition of 1a to 2a,t he results of which are shown in Figure 1. In the initial step,K 2 CO 3 coordinates to the Bpin moiety of 1a and gives rise to the pentafluorophenyl-Bpin-basec omplex 6 with free energy decreasing by 27.2 kcal mol À1 .T he energy of compound 6 is set as the zero point of the energy profile.T he pentafluorobenzene anion (Ar F À )adduct 8 is formed endothermically by cleavage of the B-C(Ar F )bond via transition state 7-ts with an energy barrier of 26.4 kcal mol À1 .I nt he optimized structures of 7-ts,K + cations coordinate to C, Oa nd Fatoms,w hereas there is only K-O coordination in compound 6.S ubsequent cleavage of the B-C(Ar F )b ond can be facilitated by this pathway.T he separated carbonate adduct and Ar F À group in adduct 8 are connected and stabilized by K + cations. Nucleophilic attack of Ar F À at the aldehyde carbon atom via transition state 10-ts occurs to achieve the coupling intermediate 11 with an energy of 17.6 kcal mol À1 .T his low activation energy barrier can be attributed to the coordination of K + to the oxygen atom of the aldehyde,thus enhancing the electrophilicity of the aldehyde carbon atom. Subsequently,t he methanolate oxygen atom attacks the Lewis-acidic boron atom to give the corresponding compound 13 irreversibly via transition state 12-ts.T he overall energy barrier for this step is 16.2 kcal mol À1 .F inally,K 2 CO 3 in compound 13 coordinates to the boron of substrate 1a via transition state 14-ts,f ollowed by cleavage of aB À Ob ond to give 16-ts and eventually 17,regenerating the active species 6.Ass hown in Figure 1, the energy barriers for these two steps are very low, indicating that intermediate 13 transforms to product 17 swiftly.T he step from pentafluorophenyl-Bpin-base compound 6 to product 17 is calculated to be exergonic by 14.3 kcal mol À1 .T he base-assisted cleavage of Bpin and pentafluorophenyl (Ar F )i sc alculated to be the rate determining step (RDS) with af ree energy of activation of 26.4 kcal mol À1 .
As shown in Figure 1, the cation K + bonds with one or two Fatoms in these intermediates and transition states,s uggesting that the fluoride substituents possibly play an important role in the 1,2-addition of polyfluorophenylboronates to aryl aldehydes.T herefore,w ec alculated the activation free energies of the RDS using polyfluorophenylboronates with different numbers and positions of fluorine substituents as the substrate.T he results given in Figure 2c learly show that the energy barrier rises with ar eduction in the number of F substituents.T he position of the fluorine atoms also affects the energy barrier, and ortho fluorine has astronger effect on the barrier than Fsubstituents at other positions.T he barrier for 24,with an ortho-F substituent, is higher than that of 22 by 2.6 kcal mol À1 ,w hereas that of 26 with a para-F substituent rises to 39.0 kcal mol À1 .I nf act, no reaction was observed under these conditions when 26 was used as the substrate, which is consistent with our calculated results.W ec onclude that the ortho-F substituent is vital in this reaction for interaction with K + along the reaction pathway,and that other Fsubstituents also influence the reactivity for the 1,2-addition of polyfluorophenylboronates to aryl aldehydes via their electron-withdrawing effect. Thus,s tronger electron-withdrawing groups located at the para or meta carbons of polyfluorophenylboronates may promote this reaction.
To ascertain the role of the K + cation in these reactions, part of the free energy profile without the cation was also
Chemie
Research Articles calculated at the same level of theory,a nd the results are given in Figure 3. Compared with the energy profile in Figure 1, in the absence of K + ,the process of the methanolate oxygen anion 33 attack at the Lewis-acidic boron in 30 becomes improbable,w ith an activation barrier of 41.4 kcal mol À1 ,a lthough the initial cleavage of Bpin and pentafluorophenyl (Ar F )s tep has al ower free energy of activation. Upon addition of 18-crown-6 to the reaction, the yields drop dramatically.A sacounterion, K + clearly regulates the nucleophilicity of CO 3 2À ,a nd promotes the reactivity by interaction with oxygen or fluorine atoms.O ur DFT calculations indicate that both the ortho-F substituents on the polyfluorophenylboronates and the counterion K + are essen-tial for the 1,2-addition of polyfluorophenylboronates to aryl aldehydes.
Thes tructures of 3f, 3l, 3m, 3n,a nd 4d were unambiguously confirmed by single crystal X-ray diffraction. While the molecular structures are chiral (Figure 4), all the compounds represent racemic mixtures.D ue to the presence of OH groups,t he arrangement of the molecules in the crystal structures of all compounds is primarily determined by OÀ H···O or OÀH···N hydrogen bonding (Supporting Information, Table S2). Thep resence of p···p stacking interactions between pentafluorophenyl and bromophenyl or naphthyl moieties (3f and 3m), respectively,i sa lso observed in these examples ( Figure 5, Table S3). Such an attractive interaction
Chemie
Research Articles between arenes and perfluorinated arenes results from the different electronegativities of the hydrogen and fluorine atoms with respect to the carbon atoms of the aromatic rings and, hence,from opposite multipole moments of the aromatic groups.I ti sc alled the arene-perfluoroarene interaction and can be applied as as upramolecular synthon in crystal engineering. [30] This was previously confirmed by Marder and co-workers,who have shown that this type of interaction leads to the formation of highly ordered p-stacks of alternating arene and perfluoroarene molecules in co-crystals of arenes and perfluoroarenes. [30d,31] In the crystal structures of compounds 3f and 3m,t he combination of both O À H···O hydrogen bonding and areneperfluoroarene interaction leads to the intriguing formation of [O À H···] 4 hydrogen-bonded cyclic tetramers with graph set R 4 4 (8) ( Figure 5, Table S2). [32] Them olecules of the tetramer interact via arene-perfluoroarene p···p stacking between the bromophenyl or naphthyl and pentafluorophenyl moieties on the outside of the cyclic [OÀH···] 4 ring. Thei nterplanar separations (3.281(7)-3.687 (14) )are typical for p···p stacking interactions [30,31] and the angles between the interacting planes are 4.96(19)-16.8(3)8 8 (Table S3). In the higher symmetry compound 3m (space group P2 1 /c with Z' = 2, where Z' denotes the number of molecules in the asymmetric unit), arene-perfluoroarene interactions are also present between the tetramers,i na ddition to C À H···p,C À H···F,a nd F···F interactions ( Figure S6). Each tetramer of 3m is centrosymmetric and, hence,c ontains molecules of opposite chirality (RRSS), leading to aracemic mixture (Figure 5b). Tetramers are arranged in sheets parallel to theb;c-plane ( Figure S6). In contrast, compound 3fcrystallizes in the non-centrosymmetric space group P1. There are 16 symmetry-independent molecules in the asymmetric unit (Z' = 16) of 3f,which build up four symmetry-independent hydrogen-bonded cyclic tetramers ( Figure S1). Each tetramer is constituted by molecules of the same chirality (RRRR or SSSS) ( Figure 5a). Thus,t he chirality of the four tetramers in the asymmetric unit, i.e., (RRRR)(SSSS)(RRRR)(SSSS), leads to aracemic mixture,as shown in Figures 5a,S1and S3. Te tramers of mixed chirality are arranged in sheets parallel to theb;c-plane with bromine atoms all pointing up or down within the sheet (Figures S2 and S3). Parallel sheets face each other either with the bromine atoms or without. In fact, crystals of 3frepresent one of the rare class of crystals for which Z' > 1. [33,34] While searching for as tructure of higher symmetry,t he cell parameters of 3f were also determined at 200 K. As this resulted in asimilar triclinic unit-cell metric as was observed Figure 5. Compounds a) 3f and b) 3m self-assemble to form tetramers via OÀH···O hydrogen bonding and the corresponding graph set notation is R 4 4 (8). [31] p···p Stacking interactions between the bromophenyl or naphthyl and pentafluorophenyl groups, respectively, within the tetrameric unit are indicated by close C···C contacts (dashed lines). a) Each of the four symmetry-independent tetramers of 3fconsists of molecules of the same chirality (RRRR or SSSS). Only one tetramer (SSSS)i sshown here. b) In 3m,the tetramer is centrosymmetric with (RRSS)chirality of the molecules. at 100 K, the occurrence of aphase transition at temperatures between 100 Ka nd 200 Ki sunlikely.
Contrary to 3f and 3m,t he dominance of hydrogen bonding and absence of arene-perfluoroarene interactions in compounds 3l (space group P " 1), 3n and 4d(both space group C2/c)resulted in the formation of one-dimensional hydrogenbonded chains ( Figure 6). In 3land 3n,the intermolecular OÀ H···O and OÀH···N hydrogen bonding interaction takes place between the alcohol (O À H, donor) and the carboxaldehyde (O,a cceptor) and pyridyl (N,a cceptor) groups,r espectively, the latter having as tronger hydrogen bond acceptor ability compared to the alcohol group (Table S2). Depending on the position of the acceptor atom in the molecule,h ydrogenbonded chains are straight (3l, Figure 6a)orzig-zag-like (3n, Figure 6b). In 3l,e ach one-dimensional chain contains molecules of one particular chirality (either R or S), and chains of opposite chirality exhibit extensive p-stacking interaction between the phenyl groups.I nt his way,d oublestranded linear chains projecting the C 6 F 5 groups on both sides are formed, as shown in Figure 6a.The C 6 F 5 groups from neighboring strands undergo interdigitation and exhibit partial offset p···p interactions between fluorinated moieties and C À F···p interactions between phenyl and pentafluorophenyl groups ( Figures S4 and S5, Table S3). In 3n,o nedimensional zig-zag chains are formed by molecules of alternating chirality (RSRS…) (Figure 6b). Thepyridyl rings lie coplanar and the pentafluorophenyl groups interdigitate via partial offset p···p interactions to form ap arallel ribbonlike arrangement ( Figure S7, Table S3). This structure exhibits abilayer architecture as there are alternating hydrophobic and hydrophilic regions (Figures S7 and S8). [35] In 4d, corrugated one-dimensional chains are observed by the intermolecular OÀH···OÀH···hydrogen bonding interactions between the alcohol groups (Table S2), and molecules constituted of alternating pairs of same chirality (RRSSRRSS… as shown in Figure 6c and Figure S9). Other intermolecular interactions observed in 4d include CÀH···F,C ÀH···p,a nd very weak, strongly offset p···p interactions (Table S3).
Conclusion
We have demonstrated here the simple conditions for the 1,2-addition of aldehydes and ketones with polyfluorophenylboronate compounds.T his strategy has the following advantages:1 )transition metal-free catalyst system;2 )a variety of aromatic and aliphatic aldehydes were found to be suitable substrates for this reaction using pentafluorophenyl-Bpin in moderate to excellent yields;and 3) sterically hindered ketones also worked well to furnish the corresponding products.T his method also introduces the use of polyfluorophenyl-Bpin compounds instead of Grignard reagents for polyfluorophenylation of arylaldehyde and ketone substrates.Further studies of the synthesis and applications of polyfluorophenyl boronates are underway in our laboratory and will be reported in due course. | 4,621.4 | 2021-04-26T00:00:00.000 | [
"Chemistry"
] |
CONTEXTUAL CLASSIFICATION OF POINT CLOUD DATA BY EXPLOITING INDIVIDUAL 3D NEIGBOURHOODS
: The fully automated analysis of 3D point clouds is of great importance in photogrammetry, remote sensing and computer vision. For reliably extracting objects such as buildings, road inventory or vegetation, many approaches rely on the results of a point cloud classification, where each 3D point is assigned a respective semantic class label. Such an assignment, in turn, typically involves statistical methods for feature extraction and machine learning. Whereas the different components in the processing workflow have extensively, but separately been investigated in recent years, the respective connection by sharing the results of crucial tasks across all components has not yet been addressed. This connection not only encapsulates the interrelated issues of neighborhood selection and feature extraction, but also the issue of how to involve spatial context in the classification step. In this paper, we present a novel and generic approach for 3D scene analysis which relies on ( i ) individually optimized 3D neighborhoods for ( ii ) the extraction of distinctive geometric features and ( iii ) the contextual classification of point cloud data. For a labeled benchmark dataset, we demonstrate the beneficial impact of involving contextual information in the classification process and that using individual 3D neighborhoods of optimal size significantly increases the quality of the results for both pointwise and contextual classification.
INTRODUCTION
The fully automated analysis of 3D point clouds has become a topic of major interest in photogrammetry, remote sensing and computer vision.Recent research addresses a variety of topics such as object detection (Pu et al., 2011;Velizhev et al., 2012;Bremer et al., 2013;Serna and Marcotegui, 2014), extraction of curbstones and road markings (Zhou and Vosselman, 2012;Guan et al., 2014), urban accessibility analysis (Serna and Marcotegui, 2013), or the creation of large-scale city models (Lafarge and Mallet, 2012).A crucial task for many of these applications is point cloud classification, which aims at assigning a semantic class label to each 3D point of a given point cloud.Due to the complexity of 3D scenes caused by the irregular sampling of 3D points, varying point density and very different types of objects, point cloud classification has also become an active field of research, e.g.(Guo et al., 2014;Niemeyer et al., 2014;Schmidt et al., 2014;Weinmann et al., 2014;Xu et al., 2014).
Most of the approaches for point cloud classification consider the different components of the classification process (i.e.neighborhood selection, feature extraction and classification) independently from each other.However, it would seem desirable to connect these components by sharing the results of crucial tasks across all of them.Such a connection would not only be relevant for the interrelated problems of neighborhood selection and feature extraction, but also for the question of how to involve spatial context in the classification task.
In this paper, we focus on the combination of (i) feature extraction from individual 3D neighborhoods and (ii) contextual classification of point cloud data.This is motivated by the fact that such a combination provides further important insights into the interrelated issues of neighborhood selection, feature extraction and contextual classification.Using features extracted from individual neighborhoods has a significantly beneficial impact on the individual classification of points (Weinmann et al., 2014).On the other hand, using contextual information might even have more influence on the classification accuracy, because it takes into account that class labels of neighboring 3D points tend to be correlated.Consequently, this paper addresses the question whether the use of features extracted from neighborhoods of individual size still improves the classification accuracy when contextual classification is applied, and whether it is beneficial to use the same neighborhood definition for contextual classification.We propose a novel and generic approach for 3D scene analysis which relies on individually optimized 3D neighborhoods for both feature extraction and contextual classification.Considering different neighborhood definitions as the basis for feature extraction, we use a Conditional Random Field (CRF) (Lafferty et al., 2001) for contextual classification and compare the respective classification results with those obtained when using a Random Forest classifier (Breiman, 2001).As the unary terms of the CRF are also based on a Random Forest classifier, we can quantify the influence of the context model on the classification results.
After reflecting related work in Section 2, we explain the different components of our methodology in Section 3. Subsequently, in Section 4, we evaluate the proposed methodology on a labeled point cloud dataset representing an urban environment and discuss the derived results.Finally, in Section 5, concluding remarks and suggestions for future work are provided.
RELATED WORK
When focusing on point cloud classification, different strategies may be involved for each component of the processing workflow.
Fixed vs. Individual 3D Neighborhoods
In order to describe the local 3D structure at a given 3D point, the spatial arrangement of 3D points within the local neighborhood is typically taken into consideration.The respective local neighborhood may be defined as a spherical (Lee and Schenk, 2002) or cylindrical (Filin and Pfeifer, 2005) neighborhood with fixed radius.Alternatively, the local neighborhood can be defined to consist of the k ∈ N nearest neighbors either on the basis of 3D distances (Linsen and Prautzsch, 2001) or 2D distances (Niemeyer et al., 2014).The latter definition based on the k nearest neighbors offers more flexibility with respect to the absolute neighborhood size and is more adaptive to varying point density.All these neighborhood definitions, however, rely on a scale parameter (i.e.either a radius or k), which is commonly selected to be identical for all 3D points and determined via heuristic or empiric knowledge on the scene.As a result, the derived scale parameter is specific for each dataset.
In order to obtain a solution taking into account that the selection of a scale parameter depends on the local 3D structure as well as the local point density, an individual neighborhood size can be determined for each 3D point.In this context, most approaches rely on a neighborhood consisting of the k nearest neighbors and thus focus on optimizing k for each individual 3D point.This optimization may for instance be based on the local surface variation (Pauly et al., 2003;Belton and Lichti, 2006), iterative schemes relating neighborhood size to curvature, point density and noise of normal estimation (Mitra and Nguyen, 2003;Lalonde et al., 2005), dimensionality-based scale selection (Demantké et al., 2011) or eigenentropy-based scale selection (Weinmann et al., 2014).In particular, the latter two approaches have proven to be suitable for point cloud data acquired via mobile laser scanning, and a significant improvement of classification results can be observed in comparison to the use of fixed 3D neighborhoods with identical scale parameter (Weinmann et al., 2014).
Single-Scale vs. Multi-Scale Features
Given a 3D point and its local neighborhood, geometric features may be derived from the spatial arrangement of all 3D points within the neighborhood.For this purpose, it has been proposed to sample geometric relations such as distances, angles and angular variations between 3D points within the local neighborhood (Osada et al., 2002;Rusu et al., 2008;Blomley et al., 2014).However, the individual entries of the resulting feature vectors are hardly interpretable, and consequently, other investigations focus on deriving interpretable features.Such features may for instance be obtained by calculating the 3D structure tensor from the 3D coordinates of all points within the local neighborhood (Pauly et al., 2003).The eigenvalues of the 3D structure tensor may directly be applied for characterizing specific shape primitives (Jutzi and Gross, 2009).In order to obtain more intuitive features which also indicate linear, planar or volumetric structures, a set of features derived from these eigenvalues has been presented (West et al., 2004) which is nowadays commonly applied in lidar data processing.This standard feature set may be complemented by further geometric features derived from angular statistics (Munoz et al., 2009), height and local plane characteristics (Mallet et al., 2011), height characteristics and curvature properties (Schmidt et al., 2012;Schmidt et al., 2013), or basic properties of the neighborhood and characteristics of a 2D projection (Weinmann et al., 2013;Weinmann et al., 2014).Furthermore, the combination with full-waveform and echo-based features has been proposed (Chehata et al., 2009;Mallet et al., 2011;Niemeyer et al., 2011).
When deriving features at a single scale, one has to consider that a suitable scale (in the form of either fixed or individual 3D neighborhoods) is required in order to obtain an appropriate description of the local 3D structure.As an alternative to selecting such an appropriate scale, we may also derive features at multiple scales and subsequently involve a classifier in order to define which combination of scales allows the best separation of different classes (Brodu and Lague, 2012).In this context, features may even be extracted by considering different entities such as points and regions (Xiong et al., 2011;Xu et al., 2014) or by involving a hierarchical segmentation based on voxels, blocks and pillars (Hu et al., 2013).However, multi-scale approaches result in feature spaces of higher dimension, so that it may be advisable to use appropriate feature selection schemes in order to gain predictive accuracy while at the same time reducing the extra computational burden in terms of both time and memory consumption (Guyon and Elisseeff, 2003).
Individual vs. Contextual Classification
Based on the derived feature vectors, classification is typically conducted in a supervised way, where the straightforward solution consists of an independent classification of each 3D point relying only on its individual feature vector.The list of respective classification methods that have been used for lidar data processing includes classical Maximum Likelihood classifiers based on Gaussian Mixture Models (Lalonde et al., 2005), Support Vector Machines (Secord and Zakhor, 2007), AdaBoost (Lodha et al., 2007), a cascade of binary classifiers (Carlberg et al., 2009), Random Forests (Chehata et al., 2009) and Bayesian Discriminant Classifiers (Khoshelham and Oude Elberink, 2012).Such an individual point classification may be carried out very efficiently, but there is a severe drawback, namely the noisy appearance of the classification results.
In order to account for the fact that the class labels of neighboring 3D points tend to be correlated, contextual classification approaches may be applied which also involve a model of the relations between 3D points in a local neighborhood.For that purpose, statistical models of context have been increasingly used for point cloud classification, e.g.Associative and non-Associative Markov Networks (Munoz et al., 2009;Shapovalov et al., 2010), Conditional Random Fields (Lim and Suter, 2009;Schmidt et al., 2012;Niemeyer et al., 2014), Simplified Markov Random Fields (Lu and Rasmussen, 2012), multi-stage inference procedures focusing on point cloud statistics and relational information over different scales (Xiong et al., 2011), and spatial inference machines modeling mid-and long-range dependencies inherent in the data (Shapovalov et al., 2013).Some methods are based on point cloud segments, e.g.(Shapovalov et al., 2010), whereas others directly classify points, e.g.(Niemeyer et al., 2014).As segment-based methods heavily depend on the quality of the results of the segmentation algorithm, we prefer point-based techniques.Typically, statistical models for context, e.g. in a Conditional Random Field (CRF), are based on interactions between neighboring point pairs, and the considerations made about the size of a local neighborhood (Section 2.1) also apply to the selection of the set of points interacting with a given point.However, existing investigations are usually based on a radius search or on the k nearest neighbors either in 2D or in 3D, involving either a fixed radius or a fixed value for k.In (Niemeyer et al., 2011), the impact of varying the radius of a cylindrical neighborhood for defining the set of neighbors is investigated.The results indicate a saturation effect when increasing that radius, so that the average number of involved neighbors is 7, but in each experiment the radius is fixed.In this paper, we want to investigate the effect of using individual 3D neighborhoods of optimal size for defining the edges of a CRF.
METHODOLOGY
The proposed methodology for point cloud classification consists of (i) neighborhood selection, (ii) feature extraction and (iii) contextual classification.Instead of treating these components separately, we focus on sharing the result of the crucial task of neighborhood selection across all components.Details are explained in the subsequent sections.
Estimation of Optimal Neighborhoods
We start from a point cloud consisting of NP points Xi ∈ R 3 with i ∈ {1, . . ., NP }.In order to obtain flexibility with respect to the absolute neighborhood size, we employ neighborhoods consisting of the k ∈ N nearest neighbors.As we intend to avoid an empirical selection of an appropriate fixed scale parameter k which is identical for all points, we focus on the generic selection of individual neighborhoods described by an optimized scale parameter k for each 3D point Xi, where the optimization relies on a specific energy function.This strategy is motivated by the fact that the distinctiveness of geometric features calculated from the neighboring points is increased when involving individually optimized neighborhoods (Weinmann et al., 2014).
The energy functions used to define the optimal neighborhood size are based on the covariance matrix calculated from the 3D coordinates of a given 3D point Xi and its k nearest neighbors.This covariance matrix is also referred to as the 3D structure tensor.Denoting the eigenvalues of the 3D structure tensor by λ1,i, λ2,i, λ3,i ∈ R, where λ1,i ≥ λ2,i ≥ λ3,i ≥ 0, two recent approaches for selecting individual neighborhoods can be applied.On the one hand, the dimensionality features of linearity L λ,i , planarity P λ,i and scattering S λ,i with sum up to 1 and may be used in order to derive the Shannon entropy (Shannon, 1948) representing the energy function Edim,i for dimensionality-based scale selection (Demantké et al., 2011): Alternatively, we may normalize the three eigenvalues by their sum j λj,i in order to obtain the normalized eigenvalues j,i with j,i = λj,i/ j λj,i for j ∈ {1, 2, 3}, summing up to 1, and we can use the Shannon entropy of these normalized eigenvalues as the basis of the energy function E λ,i for eigenentropy-based scale selection (Weinmann et al., 2014): For each 3D point Xi, the energy functions Edim,i and E λ,i are calculated for varying values of k, and the value yielding the minimum entropy is selected to define the optimal neighborhood size.Note that minimizing Edim,i corresponds to favoring dimensionality features which are as dissimilar as possible from each other, whereas minimizing E λ,i corresponds to minimizing the disorder of points within the neighborhood.Similarly to (Weinmann et al., 2014), we vary the scale parameter k between kmin = 10 and kmax = 100 with ∆k = 1.
Feature Extraction
We involve the same feature set as (Weinmann et al., 2014) Particularly in urban environments, we may face a variety of man-made objects which, in turn, are characterized by almost perfectly vertical structures (e.g.building fac ¸ades, walls, poles, traffic signs or curbstone edges).For this reason, we also involve features based on a 2D projection of a given 3D point Xi and its k nearest neighbors onto a horizontal plane P. Exploiting the projected 3D points, we may easily obtain the respective radius r k-NN,2D,i and point density D2D,i in 2D.Furthermore, we derive the covariance matrix of the 2D coordinates of these points in the projection plane, i.e. the 2D structure tensor, whose eigenvalues provide additional features, namely their sum Σ λ,2D,i and their ratio R λ,2D,i .Finally, we derive features resulting from a 2D projection of all 3D points onto P and a subsequent spatial binning.For that purpose, we discretize the projection plane and define a 2D accumulation map with discrete, quadratic bins with a side length of 0.25 m as proposed in (Weinmann et al., 2013).
The additional features for describing a given 3D point Xi are represented by the number NB,i of points as well as the maximum difference ∆HB,i and standard deviation σH,B,i of height values within the respective bin.
All the extracted features are concatenated to a feature vector and, since the geometric features describe different quantities, a normalization [•] n across all feature vectors is involved which normalizes the values of each dimension to the interval [0, 1].Thus, the 3D point Xi is characterized by a 21-dimensional feature vector fi with which is used as input for the classification of that point.
Classification Based on Conditional Random Fields
We use a Conditional Random Field (CRF) (Lafferty et al., 2001;Kumar and Hebert, 2006) for classification.CRFs are undirected graphical models that allow to model interactions between neighboring objects to be classified, and, thus, to model local context.The underlying graph G(n, e) consists of a set of nodes n and a set of edges e, the latter being responsible for the context model.
In our case, similarly to (Niemeyer et al., 2014), the nodes ni ∈ n correspond to the 3D points Xi of the point cloud, whereas the edges eij ∈ e connect neighboring pairs of nodes (ni, nj).Consequently, the number of nodes in the graph is identical to the number NP of points to be classified.It is the goal of classification to assign a class label ci ∈ c1 , . . ., c L to each 3D point Xi (and thus to each node ni of the graph), where L is the number of classes, superscripts indicate specific class labels corresponding to an object type, and subscripts indicate the class label of a given point.Due to the mutual dependencies between the class labels at neighboring points induced by the edges of the graph, the class labels of all points have to be determined simultaneously.We collect the class labels of all points in a vector C = [c1, . . ., ci, . . ., cN P ] T .Denoting the combination of all input data by x, we want to determine the configuration of class labels that maximizes the posterior probability p(C|x) (Kumar and Hebert, 2006): Here, Z(x) is a normalization constant called the partition function.As it does not depend on the class labels, it can be neglected in classification.The functions φ(x, ci) are called association potentials; they provide local links between the data x and the local class labels ci.The functions ψ(x, ci, cj), referred to as interaction potentials, are responsible for the local context model, providing the links between the class labels (ci, cj) of the pair of nodes connected by the edge eij and the data x.Ni denotes the set of neighbors of node ni that are linked to ni by an edge.Details about our definitions of the individual terms and the local neighborhood are given in the subsequent subsections.
Association Potentials:
Any local discriminative classifier whose output can be interpreted in a probabilistic way can be used to define the association potentials φ(x, ci) in Equation 5. Note that the data x appear without an index in the argument list, which means that the association potential for node ni may depend on all the data (Kumar and Hebert, 2006).This is usually considered by defining site-wise feature vectors fi(x), in our case one such vector per 3D point Xi to be classified.We use the feature vectors fi defined according to Equation 4 as site-wise vectors fi(x), whose components are functions of the data within a neighborhood of point Xi.In our experiments, we will compare different variants of these feature vectors based on different definitions of the local neighborhood used for computing the features as defined in Section 3.1.The association potential can be defined as the posterior probability of a local discriminative classifier based on fi(x) (Kumar and Hebert, 2006): For individual point classification, a good trade-off between classification accuracy and computational effort can be achieved by using a Random Forest classifier (Breiman, 2001).Such a Random Forest consists of a pre-defined number NT of random decision trees which are trained independently on different subsets of the given training data, where the subsets are randomly drawn with replacement.The random sampling results in randomly different decision trees and thus in diversity in terms of de-correlated hypotheses across the individual trees.In the classification, the site-wise feature vectors fi(x) are classified by each tree.Each tree casts a vote for one of the class labels c l .Usually, the majority vote over all class labels is used as the classification output, because it can be expected to result in improved generalization and robustness.In order to use the output of a Random Forest for the association potential, we define the posterior of each class label c l to be the ratio of the number N l of votes cast for that class and the number NT of involved decision trees: The most important parameters of a Random Forest are the number NT of trees to be used for classification, the minimum allowable number nmin of training points for a tree node to be split, the number of active variables na to be used for the test in each tree node, and the maximum depth dmax of each tree.For our experiments, we use the Random Forest implementation of openCV 1 .
Interaction Potentials:
Just as the association potentials, the interaction potentials can be based on the output of a discriminative classifier (Kumar and Hebert, 2006).In (Niemeyer et al., 2014), a Random Forest is used as discriminative classifier delivering a posterior p (ci, cj|µij(x)) for the occurrence of the class labels (ci, cj) at two neighboring points given an observed interaction feature vector µij (the concatenated node feature vectors).Thus ψ(x, ci, cj) = p (ci, cj|µij(x)) is used to define the interaction potential.The derived results show that such a model delivers a better classification performance for classes having a relatively small number of instances in a point cloud.However, in order to apply such an approach, it is a prerequisite to have a sufficient number of training samples for each type of class transition; if the original number of classes is N l , one would need enough training samples for N l × N l such transitions, which may be prohibitive.Consequently, we use a simpler model, namely a variant of the contrast-sensitive Potts model (Boykov and Jolly, 2001) for the interaction potentials: In this equation, dij (x) 2 = fi (x) − fj (x) 2 is the square of the Euclidean distance between the node feature vectors fi (x) and fj (x) of the two nodes connected by the edge eij.Furthermore, δc i c j represents the Kronecker delta returning 1 if the class labels ci and cj are identical and 0 otherwise.The parameter σ is the average square distance between the feature vectors at neighboring training points, Na is the average number of edges connected to a node in the CRF and N k i is the number of neighbors of node ni.The weight parameter w1 influences the impact of the interaction potential on the classification results.The normalization of the interaction potential by the ratio Na/N k i is required for the interaction potentials to have an equal total impact on the classification of all nodes (Wegner et al., 2011).The model in Equation 8will result in a data-dependent smoothing of the classification results.The second weight parameter w2 ∈ [0, 1] describes the degree to which smoothing will depend on the data.
Definition of the Neighborhood:
An important question in the application of a CRF is the definition of an appropriate neighborhood Ni for each node ni.For images, one can for instance use the four neighbors defined on the image grid (Kumar and Hebert, 2006).For point clouds, such a simple definition is impossible.Typically, the definition of the local neighborhood is based on the k nearest neighbors or on all neighbors within a fixed radius of the node ni.In both cases, a cylindrical or a spherical neighborhood can be used, i.e. the search for neighbors can be carried out using a 2D or a 3D neighborhood.In case of airborne laser scanning data, it has been shown that a 2D neighborhood is to be preferred, because in an urban area building fac ¸ades will only receive a relatively small number of laser points, and the height differences between neighboring points (in 2D) carry a lot of information (Niemeyer et al., 2014).The method described in this paper is designed for data acquired by laser scanners on mobile mapping devices, where one has to deal with many points on building fac ¸ades, in which case a cylindrical neighborhood does not make much sense.Consequently, we use the k nearest neighbors in 3D of each point to define the edges of the graph.However, selecting a single value for k may not be appropriate in case of varying point density.Hence, we use the neighborhood size as defined in Section 3.1 for spatially varying definitions of the local neighborhood.For performance reasons, we have to apply stricter limits to the size of the local neighborhood than for the size of the local neighborhood used to extract the features.
Thus, if the neighborhood size determined according to one of the methods defined in Section 3.1 is larger than a threshold kmax,CRF, it will be set to kmax,CRF.In our experiments, we will compare several such definitions of the neighborhood size, some of them using a neighborhood with fixed scale parameter k.For variants with variable k, the average number Na of neighbors in Equation 8 will only be based on the actual number of neighbors per node (that is, after enforcing the threshold kmax,CRF).
Training and Inference:
In order to determine the parameters of our classifier, we need training data, i.e. a set of 3D points with known class labels.The parameters of the two types of potentials are trained independently from each other.In case of the association potentials, this involves the training of a Random Forest classifier, where we randomly select an identical number NS of training samples per class.This is required because otherwise a class with many samples might lead to a bias towards that class in training (Chen et al., 2004).Note that for classes with a small number of training samples, this might result in a duplication of training samples.For the interaction potentials, the parameter σ is determined as the average square distance between neighboring points in the training data based the same local neighborhood that is used for the definition of the graph in classification.The weight parameters w1 and w2 could be set based on a technique such as cross validation (Shotton et al., 2009).Here, they are set to values that were found empirically.
For inference, i.e. for the determination of the label configuration C maximizing the posterior in Equation 5once the parameters of the potentials are known, we use Loopy Belief Propagation (Frey and MacKay, 1998), a standard optimization technique for graphs with cycles.
EXPERIMENTAL RESULTS
In the following, we present the involved dataset, describe the conducted experiments and discuss the derived results.
Dataset
A benchmark point cloud dataset representing an urban environment has been released with the Oakland 3D Point Cloud Dataset2 (Munoz et al., 2009).The data have been collected in the vicinity of the CMU campus in Oakland, USA, with a mobile laser scanning system.This system captures the local 3D geometry with side looking SICK LMS laser scanners used in push-broom mode.After acquisition, the dataset has been split into a training set consisting of approximately 37k points and a test set with about 1.3M points.The reference class labels were assigned to the points in a semi-automatic annotation process.Thus, the classification task consists of assigning each 3D point a semantic label from the set {wire (w), pole/trunk (p/t), fac ¸ade (f ), ground (g), vegetation (v)}.The distribution of the classes in the test set is very inhomogeneous, with 70.5% and 20.2% of the data belonging to classes g and v, respectively.Class f constitutes 8.4% of the points, whereas the two remaining classes (w and p/t) only consist of 0.3% and 0.6% of the points, respectively.
Experiments
For our experiments, we use five different variants of the definition of the neighborhood for computing the features described in Section 3.2.Three variants (denoted by N10, N50 and N100) are based on fixed scale parameters (thus a fixed neighborhood) of k = 10, 50 and 100, respectively, for all points of the point cloud.For variant Nopt,dim the optimal neighborhood derived via dimensionality-based scale selection is used, whereas for variant N opt,λ the optimal neighborhood is derived via eigenentropybased scale selection (cf.Section 3.1).For each variant of the feature vectors, two variants of the Random Forest classifier based on different settings are compared.In variant RF100 the Random Forest consists of 100 trees with a maximum tree depth of dmax = 4 which are trained on 1,000 training samples per class (NS = 1,000), whereas in variant RF200 we train 200 trees with a maximum tree depth of dmax = 15 on 10,000 training samples per class.In both variants, a node is only split if it is reached by at least nmin = 20 training samples, and the number of features for each test (na) is set to the square root of the number of features, following the recommendations of the openCV implementation.The first setting is a standard one, whereas the second one is expected to lead to a slightly improved performance due to the larger number of training samples and to the larger number of trees, though at the cost of a higher computational effort.
First, we apply a classification solely based on the association potentials to the dataset, i.e. on the results of the two variants of the Random Forest classifier; the respective classification variants are denoted by RF100 and RF200, respectively.After that, we apply the contrast-sensitive Potts model in a CRF-based classification.We use w2 = 0.5, a value found empirically; in a set of experiments not reported here for lack of space, we found that changes of that parameter had very little influence on the results.The chosen value gives equal influence of the data-dependent and the data-independent terms of the interaction potential.We compare three different values of the weight w1 (w1 = 1.0, w1 = 5.0 and w1 = 10.0) to show its impact on the classification results; the respective classification variants are referred to as CRF 1 N T , CRF 5 N T and CRF 10 N T , respectively, where NT is either 100 or 200, depending on whether the association potential was based on RF100 or on RF200.The size of the neighborhood for each node of the graph is based on the one for the definition of the features, but thresholded by a parameter kmax,CRF.For variant N10, we connect each point to its 10 nearest neighbors, whereas for N50 and N100 the number of neighbors is set to kmax,CRF = 15.For the other variants, we use kmax,CRF = 25, but vary the size of the neighborhood according to the one used for the definition of the features.This results in an average number of Na = 21 neighbors for Nopt,dim and Na = 15 neighbors for N opt,λ .
As a consequence of these definitions, we carry out 40 experiments.In each case, the test set is classified, and the resulting labels are compared to the reference labels on a per-point basis.We determine the confusion matrices and derive the overall accuracy (OA), completeness (cmp), correctness (cor) and quality (q) of the results.For most experiments, we only report OA and q, the latter being a compound metric indicating a good trade-off between omission and commission errors (Heipke et al., 1997).
Results and Discussion
The overall accuracy achieved in all experiments is summarized in Table 1, whereas the quality q for the five classes is shown in Tables 2-6.Some results are visualized in Figure 1.Looking at the numbers in Table 1, one can get the impression that the classification performs reasonably well in all cases, the lowest value of overall accuracy being 85.3% (RF100, N10).The best overall accuracy is better than that by about 10% (95.5% for CRF 5 200 , N opt,λ ).However, these results are dominated by the excellent discrimination of class g from the others, which is expressed by a quality of 92.3% -98.4% for that class (cf.Table 5), which, as mentioned above, accounts for 70.5% of all points in the test set.The quality is still reasonable for class v, which contains the second largest part of the data (20.2%),though the variation is much larger (61.8% -88.7%; cf.Table 6).For the other classes, in particular for w, it is very low, and whereas for p/t and f it can be improved considerably by varying the neighborhood definitions and the classifier, for class w the best result is q = 11.7%, with a variation of about 8% between variants (cf.Table 2).The main reason for the poor quality numbers of classes w and p/t is a low correctness for these classes, i.e. there are many false positives (for an example, cf.Table 7).In both cases, this is due to a relatively large number of misclassified points that actually correspond to class f.In case of poles/trunks, structures appearing like semicolumns in the fac ¸ades are frequently misclassified as p/t.Misclassifications between f and w frequently occur at fac ¸ades that are orthogonal to the road so that they show a more sparse point distribution than those facing the roads.In any case, we have to put the relatively high values for overall accuracy into perspective: some classes can be differentiated well, independently from the classification setup, whereas wires of power lines (w) cannot be differentiated using any of the methods compared here, and the main difference between the individual experiments is in the quality of the differentiation of the classes p/t and f.
Comparing the results based on a Random Forest classifier consisting of 100 trees (RF100, CRF 1 100 , CRF 5 100 , CRF 10 100 ) to those based on 200 trees, it is obvious that using more trees and more training data leads to a slightly better classification performance.The increase in OA by using 200 trees is in the order of 0.2% -3.6% for all variants (cf.Table 1).The difference in q is largest for the variants based on a fixed neighborhood.This is particularly the case for the class f for variants N10 and N50.Here, the ordering is reversed, and the variants based on RF100 achieve a considerably better performance (cf.Table 4), though at the price of other misclassifications.However, these versions are not the best-performing ones for that class, and for the variants based on a variable neighborhood the differences in q in Table 4 are smaller, in particular for the versions based on a CRF.
Of the variants using a fixed neighborhood, N50 performs best in nearly all indices.N10 performs considerably worse in OA and particularly in the quality of classes p/t and f.This also holds for the largest constant neighborhood, N100, though to a lesser degree.A neighborhood size of 50 points seems to give a relatively good trade-off between smoothing and allowing changes at class boundaries.If no interactions are considered (RF100 and RF200), the variants based on a variable neighborhood perform slightly worse than N50 in overall accuracy, with N opt,λ performing slightly better than Nopt,dim in quality for the "small" classes (w, p/t, f ) if RF200 is used as the base classifier.
Involving contextual information in the classification process improves nearly all classification indices.The improvement in overall accuracy varies between about 1% and 5% (cf.neighborhood than for N50, in the order of 2% for the first and of 1% for the latter if 200 trees are used for the association potential.Consequently, the variant N opt,λ performs better than N50 in all cases, the margin being in the order of 1%.If RF100 is used for the association potential, this also holds for Nopt,dim, whereas in case 200 trees are used Nopt,dim performs similar to N50.Again, the differences in quality for the classes w, p/t and f show higher variations.It becomes obvious that if the better base classifier (RF200) is used, these classes are differentiated best by using an adaptive neighborhood as in variant N opt,λ , in case of class p/t by a large margin.The weight of the interaction potential does have an impact on the results, but at least in those cases where 200 trees are used for the association potentials, the effect of changing the weight in the range tested here is relatively low compared to the impact of using the interactions in the first place.The value w1 = 5.0 seems to be a good trade-off in this application.
One can see from our results that the main impact of using interactions in classification consists of a considerable improvement in the classification performance of classes that are not dominant in the data, which is consistent with the findings in (Niemeyer et al., 2014) for airborne laser scanning data.In the case of mobile laser scanning data, it might in fact be those classes one is mainly interested in.The most dominant class g can easily be distinguished from the remaining data by simply considering height, and the respective completeness and correctness numbers do not vary much.In contrast, p/t might for instance be a class of major interest for mapping urban infrastructure.When using a fixed neighborhood N50 and a Random Forest without interactions (variant RF200), the completeness and the correctness of the results are 52.5% and 42.0%, respectively, resulting in a quality of 30.4% (Table 3).Nearly half of the points on poles or trunks are not correctly detected, and more than half of the points classified as p/t are in fact not situated on poles or trunks.Using the neighborhood N opt,λ and a CRF (CRF 5 200 ), these numbers are increased to a completeness of 78.8% and a correctness of 59.7% (cf.Table 7), which results in a quality of 51.4% and certainly provides a better starting point for subsequent processes.
CONCLUSIONS
In this paper, we have presented a generic approach for automated 3D scene analysis.The novelty of this approach addresses the interrelated issues of (i) neighborhood selection, (ii) feature extraction and (iii) contextual classification, and it consists of using individual 3D neighborhoods of optimal size for the subsequent steps of feature extraction and contextual classification.The results derived on a standard benchmark dataset clearly indicate the beneficial impact of involving contextual information in the classification process and that using individual 3D neighborhoods of optimal size significantly increases the quality of the results for both pointwise and contextual classification.
For future work, we want to carry out deeper investigations concerning the influence of the amount of training data as well as the influence of the number of different classes on the classification results for different datasets.Moreover, we intend to exploit the results of contextual point cloud classification for extracting single objects in a 3D scene such as trees, cars or traffic signs.
Figure 1 .
Figure 1.Classified 3D point clouds for the neighborhoods {N50, N opt,λ } (left and right column) and the classifiers RF200, CRF 5 200 (top and bottom row) when using a standard color encoding (wire: blue; pole/trunk: red; fac ¸ade: gray; ground: brown; vegetation: green).Note the noisy appearance of the results for individual point classification (top row).
group of 3D features represents basic properties of the neighborhood such as absolute height Hi of the center point Xi, radius r k-NN,i of the neighborhood, maximum difference ∆H k-NN,i and standard deviation σ H,k-NN,i of height values within the neighborhood, local point density Di, and verticality Vi.Further 3D features are based on the normalized eigenvalues of the 3D structure tensor and consist of linearity L λ,i , planarity P λ,i , scattering S λ,i , omnivariance O λ,i , anisotropy A λ,i , eigenentropy E λ,i , the sum Σ λ,i of eigenvalues and the change of curvature C λ,i .
, which has been shown to give good results in point cloud classification.This feature set consists of both 3D features and 2D features.A
Table 1 )
. It is most pronounced for the variant having the poorest OA if no interactions are considered (RF100, N10).Apart from this single example, it is in general better for the variants having an adaptive
Table 2 .
Quality q [%] for class w achieved in all experiments.
Table 3 .
Quality q [%] for class p/t achieved in all experiments.
Table 4 .
Quality q [%] for class f achieved in all experiments.
Table 5 .
Quality q [%] for class g achieved in all experiments.
Table 6 .
Quality q [%] for class v achieved in all experiments. | 9,328.2 | 2015-03-11T00:00:00.000 | [
"Computer Science",
"Engineering",
"Environmental Science"
] |
Solving Quasiconcave Bilevel Programming Problem
In this paper, we find the solution of a quasiconcave bilevel programming problem (QCBPP). After formulating a Bilevel Multiobjective Programming Problem (BMPP), we characterize its leader objective function and its feasible set. We show some necessary and sufficient conditions to establish a convex union of set of efficient point, an efficient set at the QCBPP. Based on this result, we formulate and solve a new QCBPP. Finally, we illustrate our approach with a numerical example.
Introduction
A Bilevel Programming Problem (BPP) is a decision problem where the vector variables x and y are controlled by two decision-makers: the leader and the follower.Variables x (resp.y ) are variables of decision at the upper (resp.lower) level.This structure of hierarchical optimization appears in many applications when the strategic y of the lower level depends on the strategic x of the upper level.
Mathematically, solving a BPP consists of finding a solution of the problem at the upper level called the leader's (or outer's) problem; In the literature, the BPP and the problem with multiple objectives at the upper level or at the lower level are presented as a class of bilevel problems and are at the center of research of some authors such as [1] [2].In the quasiconcave case, Herminia et al. [3] include an optimization problem at the upper (resp lower) level, in which the objective functions are quasiconcave and linear.Fatehem et al. [4] present Particle Swarm Optimization (PSO) algorithm for solving the bilevel programming problem with multiple linear objectives at the lower level while supposing the objective function at the upper level quasiconcave.
They conclude that the feasible region of the problem consists of faces of the polyhedron defined by the constraints.O. Pieume, L. P. Fotso et al. [5] [6] study Bilevel Multiobjective Programming Problem (BMPP).For the linear case, they establish equivalence between the feasible set of a bilevel multiobjective linear programming and the set of efficient points of an artificial multiobjective linear programming problem.The same authors [5] [6] show how to construct two artificial multiobjective programming problems such that any point that is efficient for both problems is an efficient solution of a BMPP.Pu-Yan Nie [7] studies bilevel programming problem where the problem at the lower level is a multiobjective programming problem by using weighting methods to analyze the constraint conditions for multiobjective programming problem.Farahi et al. [8] extend the kth-best methods to solve multiobjective linear bilevel programming problems by using fuzzy set theory and fuzzy programming to convert the multi-objective linear bilevel programming (MOLBLP) problem to a linear bilevel programming problem.
Clearly, there are very few approaches in the literature that deal with bilevel multiobjective problems.According to Pieume et al. [5], it is not easy to find efficient solution of BMPP.In [9], the authors propose to approximate the efficient set of multiobjective programming problem by the weakly efficient set and give an approach to generate a representative subset of efficient set by using well known schemes [10] [11].
In this paper, we are interested in finding the solution of a quasiconcave bilevel programming problem (QCBPP).After the formulation of a bilevel multiobjective programming problem (BMPP), we characterize its leader objective and its feasible set.Then, we show some necessary and sufficient conditions to establish that a convex union of set of efficient point is an efficient set of the QCBPP.
Based on this result a QCBPP is formulated and solved.A numerical example is provided to illustrate our approach.
This paper is organized as follows: in the next section, we present some concepts and results in multiobjective programming.In section 3, we define and D. Balme, L. P. Fotso 123 formulate a BMPP.We give in section 4, a characterization of QCBPP.In section 5, we illustrate our approach with a numerical example.Section 6 concludes the paper.
Multiobjective Programming Problem
Here, we give some concepts and results of multiobjective programming that will be used throughout the paper.
Preliminaries and Notations
A multi-objective programming problem is formulated in general as follows: with : where the i h are the objective functions for all 1, , i Q = and n U ⊆ is the feasible set.In order to solve (MOPP), it is ne- cessary to define how objective function vectors , , , Q h x h x h x should be compared for different alternatives x U ∈ .We must define on h(U) the or- der that should be used for this comparison.Due to the fact that, for 2 Q > there is no canonical (total) order in Q .Calice Pieume and al [1] propose to define partial orders on ( ) where i λ is the weight of the i-th objective i h and defines the importance of each objective.
Geoffrion [12] shows that for 0 i λ > fixed for all i, if * x is an optimal solu- tion of the LOP then * x is efficient solution of MOPP.Greffrion [12] also shows that if U is a convex set and ( ) a concave function on U for all i then * x is a weakly efficient solution if and only if * x is an optimal solution of the LOP for a λ having positive components.If * x is Pareto-optimal then ( ) Throughout the rest of the paper, the set of efficient points of a multi-objective optimization problem defined by a vector value function h on a feasible set U with respect to a cone C will be denoted: ( )
Definition of the Problem and Formulation of BMPP
Consider the problem (1) called the leader's problem formulated as follows: where for each value of x , y is the solution of the problem (2) called the fol- lower's problem; , , , : : : are the constraint functions of the leader's problem and follower's problem respectively.
Let us consider a bilevel programming problem (BPP) that comprises at the upper level the leader's problem (1) and at the lower level the follower's problem (2).The feasible region of the BPP of the first level is implicitly determined by the follower's problem (2).This bilevel programming problem is called bilevel multiobjective programming problem (BMPP) and is defined as follows: Let define a lower level solution y for every feasible x such that: : x ∈ is a parameter of the follower's problem (2).
Let consider The bilevel multi-objective programming problem (BMPP) can be reformulated as follows: Let denote by 1 Ω the feasible space (also called induced set) of BMPP given by: The optimistic formulation of BMPP is given by:
Characterization of QCBPP
Let F be the objective function of the BMPP.
Definition 1.The objective function F of the BMPP defined on a convex subset 1 Ω of n n R + of interior non empty and ( ) Let's suppose that F is quasiconcave on the convex 1 Ω , ( ) ( ) .One has ( ) , F x y k ≥ that is to say ( ) ( ) ∈ , which is convex by hypothesis on F. Therefore, ( )( ) ( ) The lemma 1 establishes that components of F are quasiconcave functions on the convex set 1 Ω .
Theorem 1.Let 1 Ω be a nonempty convex and compact subset of Consider 1 Ω a non-empty compact set of optimal points and Let denote ( ) the optimal solution of the BMPP.
Let ( ) 1 , x y ∈ Ω .Since F is a quasiconcave function, there exists ( ) is an optimal solution of the BMPP.
Definition 2. The feasible point ( ) x y ∈ Ω is the optimal solution of the For BMPP, it is noted that a solution ( ) * * , x y is optimal for the upper level problem if and only if * y is an optimal solution for the lower level problem with * x x = .
Necessary and Sufficient Conditions
Given a fixed value of ( ) , the problem (2) can be rewritten as follows: ( , : 0 be non-empty efficient subset of 1 Ω .The following result holds. ( ) Theorem 2 permits to say that ( ) Let suppose F quasiconcave and show that ( ) and consider the following constructed follower's problem: ( ) is convex, solving the BMPP is then equivalent to solving the quasiconcave problem: is a feasible solution to the QCBPP and there are no is a optimal (efficient) solution to the QCBPP, where the binary relation ≤ defines a partial order in ( ) is an efficient solution to the QCBPP.x y ∈ Ω .Since F is a continuous function on, and by definition 3 there is no ( is optimal solution of the BMPP. Let ( ) ( )
The Efficient Subset of QCBPP
( ) Let ( ) an optimal solution of the following problem: If 0 i λ ≥ is fixed and for all i , ( ) is an optimal solution of (4) then is efficient solution of (3).If 0 i λ > is fixed and for all i , ( ) is an optimal solution of (4) then is weakly efficient solution of (3).
That is, ( ) is an efficient solution as well as weakly efficient solution in represents the efficient subset in which ( ) * * , x y is an efficient solution to the QCBPP.
The follower's problem is constructed as follows: ( is an optimal solution of (II) and is weakly efficient solution of (I).Thus,
Conclusion
In this paper, we have uniquely defined a lower level solution for every upper level feasible solution as a parameter of the follower's problem.We have formulated a Bilevel Multiple Programming Problem (BMPP), of which we have considered the quasiconcave objective function and showed that there was an extreme point of the feasible space that was an optimal solution of the BMPP.We have proven a theorem, suggesting that the optimal solution of the BMPP is an efficient solution to the QCBPP.Based on this result, we presented an efficient solution which was a weakly efficient solution in the efficient subset as well.We proved that this efficient solution was the solution of the QCBPP.Thus, we concluded that solving BMPP was equivalent to solving the QCBPP.
≤where for each value of x , y is the solution of the problem at the lower level, which is called the follower's (or inner's) problem; functions of the upper (resp.lower) level.
an arbitrary cone.They show that the binary relation C ≤ defined in C by: , achieves a partial order introduced by closed pointed convex cones that are the most used.Consider the linear optimization problem (LOP) the decision variable vectors controlled by the leader (resp.the follower).
functions of the leader's problem and follower's problem respectively; the feasible region of the problem(2).The solution set of the follower's problem denoted by: -level reaction set for each decision x of the upper level and is defined as the set of Pareto-optimal points.
∈
x F x y x F x y x F x y x x y , if y is a Pareto optimal solution of the follower's prob- lem, then ( ) 1 , x y ∈ Ω is a feasible solution to the BMPP.
1 : 1 ,
F Ω → be any function.If F is quasiconcave and continuous, then there exists an extreme point of the polyhedron ( ) * * x y ∈ Ω which is an optimal solution of the BMPP.Proof:Let suppose F quasiconcave and continuous and show that ( ) result of Theorem 4 holds by Theorem 2 and Theorem 3.
are the decision variable vectors controlled by the leader (resp.the follower).The two multi-objective problems used are: that (1.5;,1.5, 0) is an optimal solution of the BMPP if and only if it is an efficient solution to the QCBPP. is an optimal solution of the following problem: is an optimal solution of (II) and is efficient solution of (I).Also, with
( 1 .
5,1.5, 0) is an efficient solution as well as weakly efficient solution in 1.5, 0 is the solution to the QCBPP. | 2,830.8 | 2017-03-14T00:00:00.000 | [
"Mathematics"
] |
The Principle of Minimum Pressure Gradient: An Alternative Basis for Physics-Informed Learning of Incompressible Fluid Mechanics
Recent advances in the application of physics-informed learning into the field of fluid mechanics have been predominantly grounded in the Newtonian framework, primarly leveraging Navier-Stokes Equation or one of its various derivative to train a neural network. Here, we propose an alternative approach based on variational methods. The proposed approach uses the principle of minimum pressure gradient combined with the continuity constraint to train a neural network and predict the flow field in incompressible fluids. We describe the underlying principles of the proposed approach, then use a demonstrative example to illustrate its implementation and show that it reduces the computational time per training epoch when compared to the conventional approach.
I. INTRODUCTION
Before digital computing, the field of mechanics relied mostly on analytical tools, which placed limitations on our ability to solve many important problems that are described by complex multi-dimensional and/or nonlinear partial differential equations (PDEs).To tackle them, even the most astute mathematicians need to invoke simplifying assumptions often diluting the underlying physics.Motivated by the need to solve such problems, and enabled by the availability of fast computing resources, the field of computational mechanics grew rapidly over the past five decades becoming the method of choice for delineating the complex mechanics of solids, fluids, and their interactions [1][2][3][4] .To this end, various computational methods including finite element (FE) and finite volume (FV) methods, as well as spectral techniques were developed and employed 2,5,6 .
Despite such advancements, traditional computational techniques are continually challenged by several enduring complexities.Mesh generation remains a time-consuming process that is an art form for most complex domains involving interacting energy fields.Moreover, solving inverse computational problems, such as those involving unknown boundary conditions is often prohibitively expensive.Commercial software packages such as COMSOL, ANSYS, and OpenFOAM involve extensive amount of coding at the backend to handle solver algorithms, mesh generation, and to set up the boundary conditions.This makes the process of maintenance and generational updates an astounding task.Amid these enduring challenges, the field of computational mechanics has recently turned towards unconventional solutions hoping to overcome these hurdles.One such promising avenue is the application of machine learning techniques to solve complex PDEs 7 .However, largely due to the absence of advanced computational tools such as efficient deep machine learning algorithms and the advancements in GPU technology, such approaches did not gain significant traction until 2019, when Raissi et al. 8 introduced the concept of physicsinformed deep learning (PINN).In principle, the proposed method starts by using a neural network to initialize a random guess for the unknown dependent variables, then substitutes them into the governing PDEs, initial, and boundary conditions.A loss function is defined and used to measure the error between the predicted output based on the guessed variables and the target solution.The loss function is then minimized with respect to the parameters of the neural network.Iterations continue until the loss function converges to a predefined threshold around zero.
The development of PINN has inspired a number of applications in fluid mechanics involving the prediction of flows fields in various domains [9][10][11][12][13][14] .The main attractive advantage of PINNs is that a unified flexible and robust framework can be used for both forward and inverse problems 15,16 .Moreover, compared to traditional computational fluid dynamics (CFD) solvers, PINNs are superior at incorporating real-world observations of the flow quantities into the governing equations.A notable example is presented in Ref. 17 , where PINNs were proposed to infer the coefficient of the convection term in the Navier-Stokes equations (NSE) based on experimental velocity measurements for the 2D flow over a cylinder.
In the domain of fluid mechanics, the application of PINNs is predominantly grounded in the Newtonian framework, primarily leveraging NSE or one of its various derivatives 18 .Unlike in other domains where variational methods have effectively solved numerous problems 19 , the use of these principles in fluid mechanics has been somewhat limited.This limitation stems from the less mature theoretical foundation in fluid mechanics.For instance, extensive studies have attempted to extend the formulation of the NSE to align with Hamilton's principle of least action.However, these efforts have faced obstacles, largely because the Hamiltonian framework struggles to incorporate non-conservative forces like viscosity [20][21][22] .As a result, variational principles in fluid mechanics have faced constraints in effectively addressing the complexities of realworld scenarios.
In a departure from traditional approaches, Taha and Gonzales 23,24 introduced a transformative concept in fluid mechanics based on Gauss' principle of least constraint, leading to the development of the Principle of Minimum Pressure Gradient (PMPG).This principle asserts that, within any incompressible flow, the pressure gradient is minimized at all times.This framework diverges markedly from conventional methodologies, offering a unique variational approach to the problems of fluid dynamics.
Building upon this foundation, our research introduces a paradigm shift in addressing incompressible flow problems by combining PMPG with PINN, thus creating PMPG-PINN.This approach uniquely informs the neural network's loss function by focusing on minimizing a cost function derived from the PMPG that is independent of the pressure.The removal of the pressure dependence from the unknown variables reduces computational cost both in terms of reducing the size of the unknowns, and in terms of evaluating derivatives in the loss function.To the authors' knowledge, this is the first work that combines PMPG and PINN to solve incompressible flow problems.
To elucidate our proposed method, the paper is structured as follows.Section 2 provides a mathematical formulation of the principle of PMPG, shedding light into its theoretical foundation.Section 3 illustrates the conventional implementation of PINN using Navier-Stokes Equations.Section 4 demonstrates the newly proposed implementation of PINN using PMPG.Section 5 presents an illustrative example that showcases the effectiveness of the approach in solving incompressible flow problems.Finally, Section 6 presents the conclusions, and potential applications and extensions of the proposed method.
II. PRINCIPLE OF MINIMUM PRESSURE GRADIENT
To effectively present the PMPG, a foundational understanding of Gauss' principle of least constraint is essential.This principle, which forms the core of the derivation, is extensively detailed in references 25,26 .Consider the dynamics of N constrained particles, each of a fixed mass m i , whose motion can be described by the generalized coordinates, q.The dynamics of such particles is dictated by Newton's second law as: where a i is the inertial acceleration of the ith particle, F i are the external forces acting on the particle, and R i are the constraint forces, which do zero work as they preserve the constraints.
Gauss' principle asserts that the quantity is a minimum with respect to the generalized accelerations, q, at every instant of time.An equivalent S can be written in terms of the constraint forces as: where the sum of the squares of the constraint force must be a minimum.
In the absence of constraint forces, R i = 0, S assumes its absolute minimum, which is equal to zero.In other words, a particle follows the external force applied to it.However, in a constrained setting, nature acts like a mathematician.It picks, at each instant of time, an acceleration for the constrained system that minimizes, in a weighted least-squares sense, the difference between the acceleration of the free motion and the constrained system.This means that, the particle adjusts its path only to the extent necessary to meet the constraints, ensuring the least possible deviation from the unconstrained trajectory.
Gauss' principle of least constraint can be applied to any dynamical system including the two-dimensional motion of incompressible fluids described by the 2D incompressible NSE, which can be written in the following form: where u = (u, v) T , is the fluid velocity vector, p is the pressure, and ν is the kinematic viscosity.Equation ( 4) represents a balance of forces applied on the moving fluid in a domain Ω ∈ R 2 .The left-hand side of the Equation is the total acceleration of the fluid, incorporating both inertial and convective components, while the right-hand side comprises the forces acting on the fluid, which can be decomposed into either external or constraint forces.Finally, Equation ( 5) represents a mass balance often referred to as continuity.
Illustration of the evolution of the flow field in the spacetime configuration, highlighting the path traced by u with a stationary quantity ∇S = 0. .Several works 18,23 have demonstrated that in incompressible flow, the pressure gradient function acts as a constraint force, primarily enforcing the continuity.Hence, by applying Gauss' principle of least constraint to the dynamics of incompressible fluids governed by NSE, and classifying the pressure gradient as a constraint force, we can write the action, S, in Euler coordinates as: subject to the continuity and any boundary conditions defined on the boundary of the domain Γ = ∂ Ω.
As shown in Fig. 1, when the system evolves, multiple trajectories of the flow field u become possible, of which only few are shown.The trajectory followed by the system, represented by the solid line, is the one that renders ∇S = 0. Thus, the flow field will deviate from the motion dictated by the inertia and the viscous forces only by the amount necessary to satisfy the continuity.In other words, no larger pressure gradient will be generated than necessary to satisfy the continuity 24 .Thus, the name the principle of minimum pressure gradient (PMPG).
III. PHYSICS-INFORMED LEARNING USING NAVIER STOKES EQUATION
In general, PINN integrates information from the governing physical laws of a given system into the training process of a deep neural network (NN) so that the unknown variables can be approximated using a limited set of training samples.In the case of the NSE, Equations ( 4), ( 5), the unknown flow field variables, u, v, as well as the pressure, p, are attained by converting the process of finding the solution of the PDEs into an optimization problem in which a loss function (error function) is iteratively minimized by updating the parameters of the neural network.This process is denoted by NN training.
As shown in Fig. 2, the NN is a fully-connected feedforward network composed of L multiple-hidden layers.It takes a concatenation of time and the state-space z 0 = (t, x) as an input, and outputs a guess for the unknown variables (u, v, p).Each layer creates data for the next layer through a tensorial nested transformation of the form 27 : where the functions σ l are called activation functions.These can be chosen based on nature of the problem to be solved.The variables W l and b l denote, respectively, the weights and biases of each NN layer, l.These are updated after every iteration (epoch) by minimizing a loss function, L , with respect to those variables.The loss function, denoted here by L (θ ) where θ = [W l , b l ], measures the difference between the predicted solution and the target solution.When the the residual change between successive iterations is less than a predefined threshold, say ε, the training stops and the output of the last iteration is considered the solution of the PDE.Further details of this process are presented in Appendix A.
In the case of Equation ( 4), ( 5), the loss function, L (θ ), is defined as: and is formed by adding two losses subject to predefined weights (penalties).The first is the loss function, L PDE , weighted by µ p , which defines the loss in the calculation of Equation ( 4) and is given by: (10) where N B is the number of collocation points over the domain.The second is the loss function, L c , weighted by µ c , which defines the loss in the calculation of the continuity, Equation (5), and is defined as: Note that, the mean square error (MSE) is employed to compute the losses.
IV. PMPG-BASED PHYSICS-INFORMED LEARNING
Deviating from the conventional approach of including the loss function associated with NSE into the training process, the proposed PMPG-PINN searches for the velocity field u which minimizes the function S in Equation ( 6) subject to the continuity constraint (7) and the associated boundary conditions.Thus, the pressure and its gradients are completely eliminated from the training process.This can be simply achieved by converting the constrained optimization problem into an unconstrained optimization problem using the following loss function: where the integral S can be reformulated as a sum using the mean rule, leading to Here A is the area of the domain, and J is the discretized operator used to approximate the integrand in Equation ( 6), and can be written as: In comparing the definitions of the loss functions in Equation (9) and Equation ( 12), a key distinction emerges.The optimization problem in Equation ( 9) is easier to solve given that the losses associated with the PDE and the continuity, L PDE and L c , can be jointly driven to zero.This consistency enables simultaneous achievement of both constraints.In contrast, for the PMPG-PINN problem, the minimization of the loss function is achieved even when the objective function, S, is different from zero.Thus, a scheme that merely optimizes the sum of the objective and the constraint loss will usually end up in an optimum that is not a solution of the PDE 28 .
To overcome this issue, we utilize the Automated Lagrangian Method (ALM) 28 , which replaces the constrained .optimization problem with a sequence of unconstrained problems with an additional term designed to mimic Lagrangian multipliers.As such, the updated loss within this methodology at the k-th iteration can be written as: where µ c is the penalty weight, and λ k i are the Lagrange multipliers.Additional details specifying the algorithm governing the updates of λ k i following each iteration can be found in Appendix B.
V. EXAMPLE: LID-DRIVEN CAVITY
To illustrate the effectiveness of the proposed approach, we consider the widely-known lid-driven cavity problem.Among the many reasons which made this particular problem one of the most common benchmarks in CFD is the combination of its simple geometry and the presence of various corner singularities.Furthermore, only velocity Dirichlet boundary conditions are required to define the mathematical problem.We initially tackle the problem using conventional PINN.This initial step sets a baseline, enabling a comparison with the new approach.Following this, we proceed to illustrate the application of the PMPG-PINN method in two scenarios; one while employing Lagrange multipliers and the other without.Finally, we evaluate the computational efficiency of PMPG-PINN by comparing it to the conventional PINN.
As shown in Fig. 3, the computational domain is defined on the area [0, 1] x [0, 1].A no-slip condition is enforced on three walls, while the upper wall is assumed to move to the right with a velocity u = f 0 (x), where, in order to ensure continuous differentiability along the boundaries, f 0 (x) is assumed to be a parabolic function in x.Furthermore, zero-pressure is prescribed at the lower left cavity corner.The maximum axial velocity is chosen such that the flow remains laminar with a Reynolds number that is equal to Re = 100.
Since the problem considered is stationary, the time variable is eliminated and the traditional PINN takes x = (x, y) as inputs, and outputs the velocity and pressure fields as (u, v, p).To force the output solution to satisfy the boundary conditions, we modify the network and construct a new trial solu- . tion, ( û, v, p), as: During the training process utilizing the conventional PINN, the PDE and continuity were equally penalized by using equal loss weights (µ p = µ c = 2).A parametric study was then performed to determine the minimum number of neurons per layer, n, collocation points, N B , and network layers, L, that minimize computational time while leading to acceptable residual errors.This led to the following numerical values: n = 20, L = 8 and N B = 90000, which were used throughout the presented simulations.
Figure 4a presents convergence of the residual loss functions, L PDE , and L c , with the number of epochs as obtained using the conventional PINN.Both residuals demonstrate a consistent decrease until they reach an acceptable value in the order of 10 −6 after nearly 10000 epochs.The number of epochs differs based on the initialization of the PINN, yet the computational time per epoch remains nearly constant at 0.050 s.The velocity magnitude resulting from these computations is depicted in Fig. 4(b).
Next, PMPG-PINN is implemented on the same domain with the same network parameters.The first implementation does not use Lagrange multipliers in order to determine if the fixed weight scheme can result in accurate results.The value of S which minimizes the loss function is obtained for fixed values of µ c ranging between 2 and 2000.Figure 5 demonstrates that S varies considerably with µ c , which indicates that the fixed-weight training scheme results in erroneous predictions that are weight dependent.To see how far those values of S are from the reference value that the conventional PINN converges to, we substitute the velocity field obtained using the conventional PINN into Equation ( 6) and obtain a reference value of S = 0.002.It is apparent that, because the continuity constraint is not highly penalized when µ c is small, the Langrange-less PMPG-PINN yields a value of S that is much lower than the reference value.A higher penalty of µ c = 200 yields the best enforcement of the continuity constraint and a value of S ≈ 0.002.However, contrary to natural intuition, penalizing the continuity further causes the value of S to diverge from the reference value because the optimization problem becomes ill-conditioned 28 .
Next, we explore the benefits of using Lagrange multipliers in the loss function using two different penalties; namely µ c = 10 and µ c = 50.For the case µ c = 10, both the value of S and the residual of the loss constraint, L c , are plotted against the number of epochs in Fig. 6(a).The figure is sectioned into three areas, each representing one outer iteration in which the Lagrange multiplier is updated.It can be seen that adjustments to the Lagrange multipliers after every outer iteration following the procedure described in Appendix B correlates with significant reductions in the value of the residual yielding a value of S = 0.0018 and a residual of the order of 10 −6 after nearly 12000 epochs.
When PMPG-PINN is trained using µ c = 50, the value of S converges to 0.00195, while the residual of the constraint loss, L c , decreases to an acceptable value of the order of 10 −6 after nearly 11000 epochs.It can be clearly seen that the error is low, implying a good agreement between the two frameworks.In general, the accuracy of the PMPG-PINN framework can be improved further by considering higher-order integration methods of S, such as trapezoidal or Simpson's rule 29 .Beyond this, several other suggestions could be considered, such as, network architecture tuning, and learning rate optimization.When implementing PMPG-PINN, we observed a notable enhancement in the computational efficiency.Specifically, the time required per epoch in PMPG-PINN is measured at 0.044 s, in contrast to 0.050 s observed using the conventional PINN.This 12% reduction in training time per epoch was found to be consistent regardless of the value of µ c , the initial guesses of the NN parameters, or whether ALM is implemented or not.The reduction in time can be attributed to two factors: i) reduction in the number of automatic derivative calculations due to the elimination of the pressure gradients, and ii) reduction in the output variables, which means fewer parameters to adjust and consequently a smaller output space to learn.The reduction in time per epoch when implementing PMPG-PINN highlights its potential applicability in solving complex and computationally demanding problems.It is important to note, however, that this reduction in computational cost comes with a trade-off because direct calculation of the pressure field is not part of PMPG-PINN framework.Nevertheless, the pressure field can be recovered by applying the resultant velocity field, u and v, to Equation 4, followed by integrating the resulting linear, first-order, and time-independent PDE.
VI. CONCLUSIONS
In this paper, we developed a new approach to solve incompressible flow problems by integrating the Principle of Minimum Pressure Gradient (PMPG) with Physics-Informed Neural Networks (PINN) to create the PMPG-PINN framework.A critical component in the development of the new approach is the incorporation of the Augmented Lagrange Multipliers (ALM) in the loss function of the neural network.This implementation is key towards effectively balancing the objective and the constraint of the model.
Applying the PMPG-PINN framework to the stationary liddriven cavity problem yielded results that are in excellent agreement with those obtained using the conventional PINN approach.Moreover, PMPG-PINN results in nearly 12% reduction in computational time per epoch during neural network training independent of the network initialization and the penalty imposed on the loss function.The enhanced efficiency is primarily attributed to the elimination of pressure gradient calculations and the reduction of output parameter space in the neural network.
We believe that PMPG-PINN may offer a potentially important alternative to the conventional PINN based on NSE.Its apparent computational efficiency can pave the way toward solving more complex and non-stationary problems.For example, a natural progression would be to adapt the PMPG-PINN framework to transient and turbulent flow conditions.Venturing into the domain of fluid-structure interaction presents another frontier for exploration.Here, a key challenge lies in integrating the dynamics of solids into the formulation of the quantity S, which in turn can be minimized using PMPG-PINN framework.
DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request Appendix A: NN structure, activation and optimizer A fully-connected neural network with L hidden layers and n neurons per layer is employed.Because of its effectiveness, we used the tanh activation function 30 .The weights of the network are initialized with Xavier initialization 31 , while the biases are initialized as zero.No dropout methods were applied.To minimize the loss function, L , we combine ADAM 32 and L-BFGS-B 33 optimizers.We first apply the ADAM optimizer for gradient descent training and then employ the L-BFGS-B optimizer to fine tune the results.During the Adam-based training, the optimizer calculates the direction of the gradient at each iteration considering a full-batch size.The initial learning rate is 0.001.We set the stopping criterion of L-BFGS-B to the smallest positive normalized floating-point number represented in Python3.The Augmented Lagrangian Method (ALM) is an optimization technique designed to handle constraints more effectively in optimization problems.It is particularly useful in scenarios where conventional methods might suffer from illconditioning issues 28 .The ALM modifies the loss function defined in Equation ( 12) by incorporating explicit Lagrange multiplier estimates, λ k i , which transforms the original constrained optimization problem into a series of unconstrained problems, each characterized by varying coefficients of λ i .Thus, the new ALM loss function can be written as: This ALM process involves adjusting the Lagrange multipliers to ensure that the constraints are increasingly satisfied as the outer iterations proceed.During each outer iteration, the PINN algorithm seeks an approximate minimizer θ k , which will then become θ k+1 .As such, we can write the optimality condition as: In the k + 1 iteration, we anticipate that the constraint function g is nearly satisfied, and consequently rendering g 2 close to zero.Therefore, the Lagrange multipliers, λ k+1 i , must be updated to replicate the gradient in the previous iteration as: Comparison of (B3) and (B4) indicates the correct update formula for λ k+1 is: A general framework for algorithm based on the ALM can be specified as follows:
FIG. 3 .
FIG. 3. Schematic diagram of the lid-driven cavity problem showing the domain and boundary conditions..
FIG. 4 .
FIG. 4. A physics-informed solution of the Lid-driven cavity problem using conventional PINN.(a) Convergence of the residual loss functions, (dashed) L PDE , and (line) L c .(b) Heat map of the velocity magnitude:
FIG. 5 .
FIG. 5. Variation of the quantity S with the penalty weight µ c .
FIG. 6 .
FIG. 6.A physics-informed solution of the Lid-driven cavity problem using PMPG-PINN.(a) Convergence of the residual loss functions for µ c = 10, (dashed) S and (line) L c .(b) Point-wise error of the velocity magnitude between conventional PINN and PMPG-PINN with µ c = 50.. | 5,654.6 | 2024-01-15T00:00:00.000 | [
"Physics"
] |
Statistical models for identifying frequent hitters in high throughput screening
High throughput screening (HTS) interrogates compound libraries to find those that are “active” in an assay. To better understand compound behavior in HTS, we assessed an existing binomial survivor function (BSF) model of “frequent hitters” using 872 publicly available HTS data sets. We found large numbers of “infrequent hitters” using this model leading us to reject the BSF for identifying “frequent hitters.” As alternatives, we investigated generalized logistic, gamma, and negative binomial distributions as models for compound behavior. The gamma model reduced the proportion of both frequent and infrequent hitters relative to the BSF. Within this data set, conclusions about individual compound behavior were limited by the number of times individual compounds were tested (1–1613 times) and disproportionate testing of some compounds. Specifically, most tests (78%) were on a 309,847-compound subset (17.6% of compounds) each tested ≥ 300 times. We concluded that the disproportionate retesting of some compounds represents compound repurposing at scale rather than drug discovery. The approach to drug discovery represented by these 872 data sets characterizes the assays well by challenging them with many compounds while each compound is characterized poorly with a single assay. Aggregating the testing information from each compound across the multiple screens yielded a continuum with no clear boundary between normal and frequent hitting compounds.
Definitions. Compounds were considered active, inconclusive, or inactive based on assignments in the data sets. Active designations are used interchangeably with hits. Compounds were deemed frequent hitters if they hit more often than expected based on a model of active behavior. When speaking of individual compounds, we restricted discussion to compounds tested 50 or more times. This threshold is arbitrary. Hit rates vary in assays with 0.5-2% of compounds found active 24 . True positives are less common with few or none of a set of hits genuinely interacting with a target 25,26 . A compound generating a true hit once in fifty screens (2%) will be poorly characterized using 50 as a cutoff. An infrequent hitter is one found to be active fewer times than expected and a normal hitter is a compound meeting statistical expectation.
Screen models. The binomial survivor function (BSF) was proposed to identify frequently hitting compounds using information from multiple screens 6,27 . This model considers each test of a compound as a trial with some probability, p, of success. The value of p is estimated from hit rates in multiple screens. Across n such trials the probability a compound is active k times can be computed using the binomial probability mass function, Compounds with values of pBSF ≥ 2 (99% confidence) were considered diagnostic of a compound being a frequent hitter compound where pBSF = − log 10 (BSF) 6,27 . The BSF model assumes p is single valued and only considers compounds being active more times than expected (frequent hitters). Compounds can also be assigned active too few times making them infrequent hitters.
Multiple probabilities can be considered by adapting the BSF model. When N screens are performed each having a probability, p i , this can be simulated by N sets of trials with the number of trials equal to the number of compounds in each screen. The values of p i can be estimated from the fraction of active compounds in each screen. This type of process is described by the Poisson-binomial distribution [28][29][30] . Such a framework allows observed results to be simulated when the probability is not constant. There is additional complexity in the data considered here due to the number of tests on each compound varying in a complex way (see Fig. 2). For this reason, each compound was simulated as the number of successes, s, in j tests done on each of the n compounds in i screens. The probability of success for each test was drawn from a function describing the distribution of p i .
The distribution of p i values obtained from the percentage of active compounds in each screen was modelled by a generalized logistic distribution [31][32][33] (LD) after logit transformation. where θ, σ, and α are the location, scale and shape parameters.
Compound models.
Another statistical tool to model active assignments in screens is based on the Gamma distribution (GD) with the following form where α and β are parameters of the Gamma distribution and can be parameterized such that α = mean 2 /variance and β = mean/variance. Once the parameters are found, gamma distributed random numbers can be generated to simulate data and the resulting simulation compared to the binomial models and to the observed data.
Compounds have a range of attributes that combine to generate responses in screens and these attributes can be modelled using statistical mechanisms. These attributes might include approximate limits on drug-like chemical space such as Lipinski's rules of 5 34,35 . Other features might be useful building blocks 12 , privileged scaffolds 4 , PAIN motifs 9,37 , and other related chemical attributes. A bottom-up model can distribute attributes by assuming initially that a randomly selected compound has a probability, p a , of having a detectable attribute based on a binomial trial. If it has one attribute, then it has the same probability of a second, and so forth. This is like flipping coins weighted by p a and accumulating attributes while a continuous run of successes is obtained.
Scientific Reports | (2020) 10:17200 | https://doi.org/10.1038/s41598-020-74139-0 www.nature.com/scientificreports/ Assuming all attributes have a single probability is clearly an over-simplification but provides a starting framework for a parsimonious model of compound behavior. The number of successes, k, after r failures, in such a process is modeled by the negative binomial (NB) distribution.
To model the number of assayable attributes in a set of compounds we parameterized the NB to only allow a single failure, r = 1. Once the attributes are distributed, each attribute provides a chance of being designated as active in any screen based on the probability of success. This second stage considers the attributes accumulated according to Eq. (5) to be like tickets for a lottery (a screen) with a probability of success for each ticket. This model can then be compared to a rank order presentation of the number of hits a set of compounds accumulate.
Rank order models. Because the literature of drug discovery has been interested in PAINs 8,9,[38][39][40] , frequent hitters 3,6 , and related effects 23 it is convenient to look at data in rank order [41][42][43][44][45][46][47][48] to focus on compounds appearing multiple times across screens and the variability in the hit percentage across the data sets. A rank order model sorts the data by a parameter, for example the number of tests on each compound, and seeks to model the shape. A three-parameter model for such data has been proposed as an extension to Zipf 's law 49 .
In this expression, A is a scale constant, a determines the curvature at low values of r, and b describes the curvature at large values of r. Specific values of a and b are related to specific distributions. For example, when a = b = 0 the uniform distribution is obtained 47 .
Results and discussion
Overview of the 872 screens. The 872 HTS data sets examined all contained tests on 50,000 or more compounds. The data include assay results from 1,759,553 compound identification (CID) numbers. These represented approximately 1.7% of the ~ 103 million compounds on PUBCHEM and about 1.2% of the ~ 145 million CID numbers assigned as of 24/3/2020. The percentage of active compounds ranged from 0% (AIDs 592, 901, 1019, 1027, 1722, 1766) to 69.62% (AID 1996, solubility assay; 40,282 active out of 57,859 compounds tested). The six with no actives included spectroscopic profiling tests (AID 592) and assays against disease targets (AID 901, inositol monophosphatase). These 872 screens contain 223,573,071 compound assays against valid CID numbers which represented approximately 83% of the bioassays on PUBCHEM as of 24/3/2020. There were 1,559,098 assays resulting in active designations (0.782% of total) and 2,560,020 inconclusive tests (1.28% of total). These counts do not include a small number of compounds for which no CID number was included. The active designations represent 374,431 different compounds (21.2% of all compounds tested) and there were 431,189 inconclusive compounds (24.5% of compounds).
Active compounds in screens. The fraction of active compounds in screens has been considered a probability when trying to identify frequent hitters 6,27 . The wide range in the observed percentage of active compounds seen here suggested that no single probability defined the behavior over the series of screens. Although the average probability was 0.0078, this number had considerable variance. To better understand the distribution of the expected number of actives, we investigated the normal, skew normal, beta, logistic, and generalized logistic distributions as candidate distributions describing these probabilities. The logit transformed probabilities gave good correspondence with a generalized logistic distribution 33 (GLD) with location = -5.0587, scale = 0.6958, and shape = 0.6333 (Fig. 1a) in comparison to a normal distribution. The rank ordered data fit well to Eq. (5) (Fig. 1b) with parameters: a = 0.6084 ± 0.0068, b = 1.183 ± 0.0074, and A = 0.000121 ± 0.000091. Simulations using GLD random numbers followed by inverse logit transform were nearly identical to the data with reasonable recovery of rank order parameters (a = 0.64 and b = 1.01). These a and b values occupy a region that is currently not well documented 50 . Descriptively, a and b are related to the skewness and shape of the distribution with the a and b parameters related to the left and right sides of the distribution, respectively.
The distribution of active probabilities (Fig. 1a) provides the information needed for a generalization of the binomial survivor model based on the Poisson-binomial distribution to build prospective models of large HTS programs consisting of hundreds of campaigns. The shape of this distribution is critical for appropriate cost-benefit analysis of publicly funded HTS as it enables estimations of the scale in terms of campaigns required to find some number of lead compounds for an arbitrary set of diseases. In combination with a more holistic view of the entire drug discovery process, it could allow reasonable predictions of the resources needed from beginning to end of the process. To the extent these 872 screens are representative of public HTS campaigns, it is the best data available for modeling future campaigns.
Frequency analysis of tested compounds. The frequency with which each compound was tested was noted by the PUBCHEM CID number (Fig. 2). The frequency with which compounds were found to be inactive had a similar shape and structure (not shown). The rank ordered distribution shows a prominent feature due to a subset of the compounds appearing repeatedly within single data sets (Fig. 2). A total of 275 compounds were tested more than 872 times. The most tested compounds were a known approved drug: the anti-depressant maprotiline hydrochloride (CID 71,478; 1613 tests) and the serotonin reuptake inhibitor fluvoxamine maleate More problematic is the case of inositol. Inositol appears nine times in the two screens represented by AID 175 and AID 248 and 1067 times in the 872 screens. Inositol has nine stereoisomers and some of these are clearly designated in the names provided with the SIDs to PUBCHEM. PUBCHEM had eight CID numbers associated with inositol, however, six were deuterated or tritiated isotopologues. In AID 175 and AID 248 there were eight different SID numbers submitted with structures all mapping to CID 892. The names provided with the eight SIDs sometimes refer to specific forms leaving ambiguity between name and structure. The 1067 repeats assigned to CID892 may reflect this ambiguity and the tests were made on specific inositols but the stereochemical information was subsequently lost during reporting to PUBCHEM. All 9 inositols currently appear as synonyms for CID 892 along with hundreds of other names. Given the importance of stereospecific effects 51 , this loss is concerning and researchers trying to mine the PUBCHEM data base may be unaware of this problem. Submitters need to be more careful about the structural information they provide to better deal with stereochemistry and greater care would enhance the value of the PUBCHEM data base.
Scientific Reports
The compound test frequency rank order distribution (Fig. 2a) includes two other notable groups of compounds. One group of compounds is made up of the same core libraries tested over and over and another group tested 3 or fewer times. As an example of the former, there were 218,998 compounds tested over 436 times. This group represents 62% (139,313,672) of the 223,573,071 tests. The bulk of tests appear to represent a small core of compounds tested repeatedly. A similar portion (235,292 compounds, 13.4% of the total) were tested only once (0.1% of tests). It is unlikely this represents the best approach to compound library design and management. More equitable testing and expansive sampling of chemical space might help find more new drugs. This structure changed somewhat over the 15 years represented in the 872 data sets (Fig. 2b). The change we ascribe to an increased number of screening labs with more variation in libraries, however, each lab is retesting its core library repeatedly. The scale of the number of compounds tested increased roughly a factor of 3 over the period with increased diversity during the last 100 screens. Based on this evidence, we concluded that this series of HTS experiments taken in aggregate includes extensive compound repurposing at scale 23 . It also includes direct drug repurposing (e.g. the most tested compound maprotiline). In the 872 HTS data sets, core libraries are repeatedly tested against new targets. When these are looked at in isolation, it may seem reasonable to redeploy an existing library against a new target, but in aggregate the likelihood that discoveries will be made of fundamentally new classes of drugs in the core of repeatedly tested libraries is small. At the same time, more attention needs to be paid to the compounds tested a small number of times. How these are introduced and managed is particularly important as it is arguably the part of these screens representing drug discovery rather than compound repurposing at scale (Table 1).
A further issue became apparent when assessing how CID numbers were distributed (Fig. 3). CID numbers are assigned sequentially which provides an approximate timeline for when the compounds were added to the PUBCHEM data base ( Table 2). There are some limitations to this approach because some CID numbers added recently represent compounds that were used in early screens. One such example is CID 135,400,595 which was added in 2019 but used as early as AID368 (deposited in 2006). Despite this limitation, constructing a plot of the frequency of tests against CID number (Fig. 3) gives insight into the extent to which chemical space (as represented by CID numbers) is being sampled (Fig. 3a) and tested (Fig. 3b). It is neither sampled nor tested uniformly with large sections almost untested. Over 58% of compounds tested were drawn from the first 10 million compounds added. The aggregate library of 1,759,553 compounds tested in these screens is clustered into blocks as are the number of times tested. In the absence of evidence to the contrary, there is no reason to believe that the poorly sampled sections of chemical space are less likely to contain pharmacologically useful compounds. There are also very specific CID number ranges that contain nearly all excessively tested compounds (> 872 tests).
Frequency analysis of active and inconclusive compounds.
Application of the binomial survivor function 6,27 such that a frequent hitter was any compound with pBSF > 2 found 42,264 compounds (11.3% of actives) meeting these criteria with more than 50 tests. We investigated further using two binomial simulations Table 1. www.nature.com/scientificreports/ (Fig. 4a). The first model consisted of a binomial with a single probability (0.0078). The second was a Poissonbinomial model using probabilities drawn from the generalized logistic distribution. These models were used to simulate the structure of the active rank order distribution. When compared to the observed active distribution (Fig. 4a), both simulations indicated a large group of compounds hitting above expectation. The Poissonbinomial generated fewer frequent hitters but the improvement was marginal. Both models gave a poor representation of the observed data. Both models generated a much larger group of infrequent hitters. Infrequent hitters have not been described previously and they are important for understanding HTS. Specifically, binomial models based on screen level probabilities (Fig. 1) do not predict how hits distribute among the compounds in a screen. These probabilities are not equal due to the chemical properties of the compounds tested. The literature contains extensive discussion of PAINs and frequent hitters, but the chemistry of infrequent hitters also needs investigation. Similar behavior was seen in the distribution of inconclusive compounds (Fig. 4b). The scale of this group was unexpected and to our knowledge there has been little discussion of this group of compounds. A single valued binomial simulation gave a poor representation of these compounds with many "frequent inconclusive" and "infrequent inconclusive" compounds. Inconclusive compounds represent a slice of the response distribution between the threshold for being "active" and the remainder of the distribution. As such, true hitters are expected to be concentrated above the active threshold and less prevalent in the inconclusive section of the distribution and therefore fit a binomial survivor model better. There was little evidence supporting this (Fig. 4) and the binomial simulation of the inconclusive compounds was unsatisfactory. It is unclear whether inconclusive compounds are investigated seriously during lead generation but from a model building point of view they are essential. Inconclusive designations represent a different section of the distribution of measured values from the active tail and need to be included in a self-consistent model of HTS.
Models of active designations. The failure of the binomial models based on screen probabilities to provide a model of active compound frequencies required alternatives based on compound behavior. Two types of models were investigated to simulate the frequency of active designations across the 872 HTS data sets. Type I sought to build a model of compound activity giving chemical insight into what makes a compound hit. The goal was to simulate the results for randomly selected molecules drawn from a chemical space. This approach assumes that the compounds in the study are representative. Successful model construction could help decide whether HTS should grow ever larger in scale or should be smaller with more randomness in compound selection. Type II models sought improved statistical description reducing the proportion of infrequent hits.
An initial type I model was built on the following assumptions: (1) compounds have attributes that can be detected in a screen; (2) attributes are accumulated following a negative binomial process with r = 1 and p 1 = 0.50 with p 1 found by least squares minimization; (3) each attribute may be detected (result in active designation) with a binomial probability, p 2 = 0.01. A small group would be expected to get a string of 18-20 successes (attributes) in a row and 50% would be expected to fail on the first attempt. The attributes distributed in this manner might include Lipinski's five 34 , privileged scaffolds 4 , or PAIN motifs 9,37 . For example, if all compounds on PUBCHEM were sampled, there is some probability a randomly sampled compound has a molecular weight under 500. This www.nature.com/scientificreports/ model is simple but provides a framework for a probabilistic approach to estimating compound behavior in screens. This model (Fig. 5a, dotted line) gave a better approximation to the observed behavior of the compounds in rank order than the BSF model. We then investigated descriptive models for this data. A gamma distributed simulation (Fig. 5a, dashed line) parameterized by shape (mean 2 /variance = 0.0755316) and rate (mean/variance = 0.085242) worked for much of the range leaving a much smaller (~ 1000) frequent hitter group (Fig. 5b) and fewer infrequent hitters than the BSF model. The residual sum of squares for the binomial and attribute models were 27 and 3.3 times greater than the gamma model, respectively. The residuals from the gamma model could be approximated by summing attributes. For example, one to four attributes gave the results shown in Fig. 5b. While not definitive, the analysis made clear that considering the behavior of compounds rather than screens is required to understand active compounds. In the context of HTS, a clear interpretation of the gamma distribution is difficult, however, it was clearly better than any of the binomial models tested. The key limitation of the binomial models using screen derived probabilities is the assumption that all compounds have equal opportunities in a screen unless they are part of the select group of frequent hitters. This assumption cannot be supported by the data.
Compound hit probabilities. Compound behavior can be considered in terms of the total number of hits (Fig. 4) or by the fraction of screens where a compound was active. The former is a better indicator of total follow on cost and the latter a better indication of compound behavior. Individual compound probabilities can be investigated to see if a clear divide exists between normal and frequent hitters. The histogram of compound hit probabilities for p d > 0.0 (Fig. 6a) gives a spiky appearance with peaks in the histogram at 1, 0.5, 0.33, 0.25. 0.2 etc. corresponding to 1 active test in 1, 2, 3, 4, 5, … measurements. To avoid this and similar issues, the rank order presentation (Fig. 6b) was restricted to compounds tested 50 or more times (420,572 compounds). The histogram (Fig. 6a) was neither normally distributed around some constant nor obviously bimodal and there was a nearly continuous range of p d s (Fig. 6b) giving no clear boundary between normal and frequent hitting compounds. Figures 1 and 6 reveal a limitation in the current approach to HTS. A single HTS screen is an extremely poor way to characterize compounds, particularly if each is tested once. It is an extremely good way to characterize an assay response. A compound tested once is poorly understood. However, an assay challenged with 100,000 compounds has been thoroughly tested.
Characteristics of frequent hitters. The chemical characteristics of frequent hitters were assessed based on the number of times they were deemed active relative to the gamma model in Fig. 5a. We chose hit numbers, rather than hit probabilities since each hit would have required further workup and cost. For example, the most frequently active compound, 3-Methyltoxoflavin (CID 460,747, 159 times), has been patented as a treatment for tuberculosis but may have been considered multiple times for other conditions. The second most active compound (toxoflavin; CID 66,541; 151 times) features in multiple patent applications. Forty-four compounds were www.nature.com/scientificreports/ active over 100 times. Similarly, the first 100 were active over 80 times each and the 1000 most frequent hitters were all active 45 or more times. For scale, 374,431 different compounds were deemed to be active 1,559,098 times. Follow on costs associated with these active designations was probably high. Investigation of the 1000 most active compounds yielded 429 suspected PAINs, 568 not matching any PAIN 37 motif, and 4 inorganic or otherwise unclassified compounds (e.g. CID 44,202,984). Of the 568 compounds not matching current PAIN motifs, 80 contained other undesirable substructures. These 568 are of interest because they represent compounds that could be privileged structures 3,4 , PAINs 7-9 , or both. To assess this, the 541 without any suspected PAIN motifs were examined for structural similarity and presented as a heatmap (Fig. 7). Each pixel in the heatmap represents the Tanimoto similarity between two compounds with brighter pixels representing greater similarity. The heatmap shows many recognizable clusters as square along the diagonal with three examples presented in Table 2. All 568 compounds deserve further scrutiny and the structural motifs defining the clusters identified. A full annotated list the 1000 CID numbers appears as supplementary material S1.
Across the 872 screens, compounds were repeatedly tested but subsequent analysis does not appear to extend much beyond single screens. This leads to multiple retesting of frequent hitters, PAINs, potential PAINs, and otherwise promiscuous compounds as well as missing an opportunity for discovering new scaffolds. Although some of these multiply active compounds may have useful characteristics, in the absence of a more coherent approach, repeated retesting is not going to discover this. The scale of the problem needs to be understood. Within this data set there were 37,050 compounds deemed active 10 or more times. There is no chemical overview of either the modes of action of these compounds or their specificity (if any). It is unclear whether they should be purged from libraries or built into smaller scale libraries as a pre-HTS compound repurposing stage.
Conclusions
Most of the HTS data represented here is compound repurposing at scale (Fig. 2) and assay characterization. Many compounds were repeatedly retested and subsequently treated screen by screen. The multi-screen information is the most valuable for chemists. It is unclear whether industrial screening companies provide this information to their clients. It is also unclear whether large scale international screening programs are being set up so compound behavior can be transparently assessed.
Repeated retesting raises issues similar to the multiple comparisons problem; with enough tests eventually all the compounds will be designated as active. Across the 1.7 million compounds here, there is a large amount of prior information of varying quality but little evidence it is used effectively. For example, a compound with 100 active designations may be understood sufficiently for decisions to be made about its behavior. It is not clear why such compounds remain in HTS libraries. Compounds tested only a few times are not well characterized and it is not clear why so many compounds do not remain longer. Information about compounds from previous screens needs to be used more effectively.
Future studies trying to discover new drug compounds should consider sampling chemical space in a more strategic way (Fig. 3). A 1.7 million sub-sample of 100 million items could be an excellent sample if done carefully. However, to the extent that PUBCHEM represents a view of a chemical space, most compounds tested are in the first 10 million added and much of the remainder is in specific regions. These same restricted regions are most likely to be repeatedly or excessively tested. It is worth looking more closely at library construction. Is the best predictor of compound inclusion today whether it or a similar compound was there 15 years ago? If these www.nature.com/scientificreports/ isolated portions of PUBCHEM chemical space really contain the best hope for new drugs, then they should be used. However, it is unclear this hypothesis has been rigorously tested. For the remaining less tested compounds, the process by which they are included, tested a few times, and discarded is unclear. For some, a single test was considered enough. For others, no decision was made even after 1613 tests. The binomial survivor model is an extremely useful and falsifiable model. As clearly as it identifies frequent hitters, it also generates a much larger pool of infrequent hitters (Fig. 4). Considering a distribution of probabilities representing these screens rather than a single probability does not mitigate this problem. The chemical behavior of infrequent hitters deserves as much attention as frequent hitters. A self-consistent chemical model of what makes a compound hit is essential to understanding how attributes contributing to activity are distributed in chemical space. Future models should be able to identify infrequent hitter chemical motifs which may also be the keys to making compounds less toxic.
Aspects of the rank order active designations (Fig. 5) can be generated by considering attributes randomly distributed in a collection of compounds by a negative binomial process. This attribute model also generates excessive infrequent hitters. A gamma distribution parameterized by shape and rate was a better predictor of the number of active designations. Both models performed better than the BSF model indicating that compounds do not have a binomial probability of being found active in a screen. With enough tests the probability that a compound will be active can be quantified (Fig. 6), but when investigated, a continuum between normal and frequent hitting compounds was found. There was no clear boundary between them.
Nearly half of the 1000 most active compounds in the 872 screens were suspect PAINs. The remainder contained clusters of compounds which should be examined for new motifs either to decorate (privileged structures) or avoid (PAINs). In future studies, it will be useful to investigate the strength of the responses to these compounds and if they respond preferentially in particular types of assays. The presence of these compounds indicates that insufficient effort is being made to understand the behavior of compounds across multiple screens.
This study assumed that the process of assigning active compounds was done perfectly. This needs to be revisited in future work. To better understand chemistry rather than assays, future studies should be looking www.nature.com/scientificreports/ to challenge compounds with multiple assays, rather than the other way around. While there is considerable public data providing the information for some compounds, this set of 872 HTS studies were not designed to characterize chemical behavior and are doing it poorly.
Methods
Data sets. Data were obtained from PUBCHEM and included all Assay Identification (AID) numbers having more than 50,000 compounds in the data base on 19 March 2020. This yielded 872 usable data sets. A few had to be discarded because they did not include decisions on activity. The data cover a period of approximately 15 years with the first screen in the study (AID155) deposited August 15, 2004 Statistical analysis. The data were downloaded and analyzed using R 52 (Version 3.6.2) running within R-studio (version 1.2.1335) with packages dplyr 53 , car 54 , brms 55 , moments 56 , gplots 57 , ggplot2 58 , ChemMineR 59,60 , and glogis 32 . The active, inconclusive, inactive, and all CIDs were identified and collated across the 872 data sets by CID number. Tests for PAINs were done using the FAF-Drugs4 37 portal. Simulations were done using either random number generators (to simulate individual compounds in screens) or using numerical computations of the distributions. These were done using the R-packages noted and all the code needed to reproduce the study has been provided. | 7,082.6 | 2020-10-14T00:00:00.000 | [
"Chemistry",
"Computer Science"
] |
Detection and characterization of three-dimensional interconnect bonding voids by infrared microscopy
Abstract. The three-dimensional (3-D) integrated circuit relies on the stacking of multiple two-dimensional integrated circuits into a single device using through silicon vias (TSVs) as the vertical interconnect. There are a number of factors driving 3-D integration, including reduced power consumption, resistance–capacitance delay, form factor, as well as increased bandwidth. One of the critical process steps in all 3-D processes is stacking, which may take the form of wafer-to-wafer, chip-to-wafer, or chip-to-chip bonding. This bonding may be temporary, such as can be used for attaching a device wafer to a handle wafer for thinning, or permanent, incorporating direct metal bonds or solder bumps to carry signals between the wafers and oxide bonds or underfill in the regions without conductors. In each of these processes, it is critical that the bonding is executed in such a way to prevent the occurrence of voids between the layers. This article describes the capabilities of infrared (IR) microscopy to detect micrometer size voids that can form in optically transparent blanket media such as oxide-to-oxide permanent bonding, benzocyclobuten permanent bonding, or temporary adhesive bonding laminate interfaces. The infrared microscope is described, and the measurement results from a bonded void wafer set are included. The wafers used to demonstrate the tool’s capabilities include programmed voids with various sizes, densities, and depths. The results obtained from the IR microscopy measurements give an overview of the technique’s capability to detect and measure voids as well as some of its limitations.
Introduction
For many years, ongoing requirements for increased computational power at increasingly higher device-packaging densities were accomplished by shrinking the sizes of the basic devices themselves to ensure higher degrees of integration.In the last few years, three-dimensional (3-D) integration has emerged as a complementary method to feature-size scaling to achieve the performance improvement, in which integrated circuits are stacked together in order to improve power consumption, reduce resistance-capacitance delays, decrease device form factor, provide heterogeneous integration, and increase bandwidth, allowing the most efficient process technologies to be used for the various types of devices.One embodiment of this architecture stacks multicore central processing unit (CPU) units with memory devices in the same package.
Several process flows have been proposed for manufacturing 3-D integrated circuits, commonly referred to as viafirst, via-mid, and via-last. 1,2Each of these 3-D technologies requires new wafer-level process technologies to be added to the process flow.These process steps include fabrication of through-silicon vias (TSV), wafer thinning, and either temporary or permanent wafer bonding.
Motivation for Void Detection
Bonding is a key step in 3-D integrated circuit fabrication, occurring in multiple steps during the fabrication process.
There will be a permanent bonding process during final assembly-which may be a wafer-to-wafer, chip-to-wafer, or chip-to-chip procedure-this will be using a process such as oxide-to-oxide, Cu-to-Cu, or adhesive bonding.In addition, there may be a temporary bonding step, where a device wafer is temporarily bonded to a handle wafer to permit the wafer thinning and other process steps.For the vialast process, the vias are formed after the wafer is thinned (note that in the via-first and via-mid processes, the vias are formed before or during the device and interconnect process).Each bonding process requires a strong, uniform bond, which is free of voids.These voids can occur from a number of chemical or mechanical processes including trapped air, solvent evaporation, outgassing from the polymer during curing, or particulates and surface nonuniformities.Such voids can interfere with the mechanical stability of the interface, causing unwanted local topology variation, nonuniform thinning, which can affect the TSV reveal process yield, or even delamination and breakage during thinning.Additionally, voids that occur during final assembly applications may interfere with electrical connectivity.
Process Flow Description
Identification and characterization of bond voids have prompted a search for appropriate high-volume manufacturing metrology tools, to be used to scan each of the various bond interfaces described above for voids.The Inspection and Metrology Task Force of the SEMI 3D Stacked IC Committee has initiated round-robin experiments to *Address all correspondence to: Jonny Höglund, E-mail<EMAIL_ADDRESS>the capabilities of various metrology tools to detect and/or characterize voids between wafers. 3The first of these experiments uses bonded wafer pairs, produced at SEMATECH, which are patterned with sets of programmed voids.The wafer pairs are produced using the following process steps: a 5-nm oxide layer was grown on the top surface of both the patterned wafer and the cap wafer.A 70-nm SiN film, followed by an organic underlayer, was deposited on the wafer designated as the void wafer.Photoresist was deposited and after developing, the wafer was etched to one of the four predetermined depths, using the organic underlayer and SiN as hard masks.The patterned and unpatterned wafers were oxide bonded after wet clean and surface plasma activation.
Since these wafers do not have metallization or other materials that would be found on device wafers, this experiment is intended to provide a baseline for the capabilities of various metrology techniques.Most of the techniques involved in this experiment, including the one described in this article, are expected to be limited in their ability to identify voids between the interfaces in bonded stacks of patterned wafers.Also, note that these voids differ from voids caused by in-process variations, such as particles, trapped gas, etc., as described above, and the voids do not induce stress in the wafers at the region of the void.The technique described in this article does not depend on measuring stress to identify and characterize the void, allowing these patterned voids to stand in for "real" voids in this experiment.Finally, please note that none of the current metrology tools proposed for void detection are thought to be capable of detecting the actual particles that cause voids; the best expected performance is to be able to detect the voids, which are many times the size of such particles.
Instrument Description
This article reports on a technique under development using infrared (IR) microscopy 4,5 to identify and characterize the voids.The absorption edge of silicon is approximately 1 μm, which means that the IR wavelengths are necessary for the wafers to be transparent and hence for the voids to be seen by the microscope. 6Since this technique uses reflected IR microscopy, it is expected to be especially useful for temporary bonding processes where the interface can be imaged through the unpatterned carrier wafer.
The technique is particularly interesting given its relatively simple setup and fast detection with reasonable spatial and depth resolutions.As will be shown in this publication, good progress has been made in applying IR microscopy to bonding interface void detection.Good correlation to known void dimensions is demonstrated, and the technique's limit of detection is explored by varying void diameter, density, and depth.Further work is ongoing to integrate the technique into a fully automated metrology tool.
The work presented in this publication has shown IR microscopy to be a promising technique for detecting and measuring geometries of wafer-to-wafer bonding interface voids.A basic diagram of the instrument can be seen in Fig. 1.The instrument has a broadband IR light source that is incident onto the wafer, through an optical filter and a beam splitter.Optical filtering is performed using a 1-μm high-pass filter in sequence with a 1.3-μm low-pass filter, resulting in a wavelength range of 1.0 to 1.3 μm.
For detection, an indium gallium arsenide line camera is used, which has good sensitivity in the 1.0 to 1.7-μm wavelength range.Since silicon is transparent in the IR, the instrument is capable of detecting and measuring voids through the top silicon wafer, which is typically ∼775-μm thick.The objective lens assembly contains a lens exchanger, which is fitted with lenses of varying magnifications for resolving different size features.The objectives used for image acquisitions to date have magnifications in the range of 5× to apertures of 0.1 to 0.65, and field of view from 8 to 0.8 mm.The camera pixel size calibration is performed by measuring an object of known size in micrometer to determine its size in pixels.After the calibration has been performed, the lateral dimensions measured by the microscope are directly reported during the measurements.The system includes a wafer flipper, as seen in Fig. 2, in order to enable the inspection of the wafer through the bottom substrate.Handling of the wafer is done with edge gripping, and the wafer stage used is a high-precision XYZ stage.
The microscope images are acquired using the line camera, which has its detector array oriented perpendicularly to the primary scan direction of the stage.The width of the field being scanned therefore depends both on the size of the line array and the magnification of the lens objective being used.Along the scan direction, the image is created by stitching consecutive line image captures as the wafer is scanned at constant speed under the lens objective.The stage motion controller is used to generate a trigger signal for image acquisition.When the stage has moved a predetermined scan length a trigger signal is sent to the line camera, which is synchronizing image acquisitions with the movement of the stage.This predefined scan length, corresponding with the time between triggers, is set so that the pixel size in the scan direction equals the pixel size in the direction along the line camera.The acquisition of a single line starts when the trigger signal is received by the camera.The exposure time is therefore set to be somewhat shorter than the time required for the stage to travel the distance between two triggers.
The two measurement modes available are full and partial wafer scans.For partial wafer scans, the images are captured from predefined areas of the wafer, which regions may be configured to represent dies on the wafer or to cover areas of the wafer that are prone to have voids in the bonding process.The length of the scan is therefore dependent, on which one of the two measurement modes is being used.In case of full wafer scans, the scan length is determined by the wafer boundaries and is therefore dependent on how close to the center of the wafer it is being acquired, whereas for the partial scan mode it is determined by the width of the fields to be scanned.Since the size of scanned areas is generally larger than the field of view of the line camera, a final image is composed of multiple image scans that are stitched together with a small overlap, ensuring full coverage of the scanned area.
The actual time needed for a full wafer scan is dependent on which magnification the objective lens is being used, since higher magnification results in smaller pixel size and smaller field of view.Using a 5× microscope objective, a full wafer scan takes approximately 10 min, while a full wafer scan using a 20× microscope objective requires a little over 1 h.It is therefore useful to be able to control the magnification with the goal to keep the measurement time as short as possible for a given void feature detection wafer scan.
The depth of field (DOF) of the imaging system is dependent on the magnification lens objective used.For the 5× magnification, the DOF is greater than 100 μm, while for the 50× magnification it is a few micrometers.Since the wafer is supported by its edges, there is some sagging of the wafer toward its center.Experience has shown the DOF to be sufficient to resolve a complete field of view, but re-adjustment of the focus height is needed as the wafer is being scanned, and several types of autofocus systems are currently being evaluated for the final metrology tool.
The basic void detection principle is based on the fact that a reflection can occur at the transition in different refractive indices.In this case, because the void has relatively lower refractive index compared to the silicon dioxide bonding material, it gives a different reflectance than the surrounding material.As a result, the amount of light reflected at the bonding interface is different in areas where the voids are present, allowing the voids to be detected by the IR microscopy instrument.An example of this is shown in Fig. 3.
The horizontal geometry of the voids can be determined directly by measuring the voids seen on the indium gallium arsenide (InGaA)s camera.An example of intensity linescan across a void is shown in Fig. 4. As can be seen, the intensity is around its maximum value for a considerable portion of the void diameter, and there are minima near the edges of the void, which have somewhat lower intensity than the background.To determine the lateral size of a single void, we used a 30% intensity threshold edge-detection algorithm.The results obtained from the experiment using this method were found to be in good agreement with nominal designed dimensions for various void sizes, and good numerical stability was achieved when calculating the diameter of different voids of the same size.The distance found is scaled using the pixel size calibration before reporting the void size as the measurement result.
In order for voids to be measureable, they need to be resolvable and detectable.Detecting void defects can be done by inspecting an acquired image for intensity changes that are not part of the intentional pattern of the sample being inspected.In order for a detected void to be resolvable, it needs to be separated in the acquired image from neighboring voids.Examples are included in the experimental results for larger areas of multiples voids that are detectable, but with pitch too small for individual voids to be resolvable or measureable.The measurement algorithm used allows reporting of dimensions and locations for voids that are resolvable and at least three pixels in size.
Automated Void Detection Method
The approach for detecting and measuring voids is different in the case of the experiment with the wafers from SEMATECH containing intentional voids and in cases where unintentional voids are present on patterned production wafers.In both cases, a wafer coordinate system is required.By aligning the wafer using the notch and edges prior to the measurement, the sample position and orientation on the stage can be corrected, and the coordinate system with the origin in the center of the wafer can be defined.
In the case of the experiment with the SEMATECH wafers, the pattern itself consisted of the voids to be detected and a given die was requested to be scanned.After aligning the wafer and scanning the requested die, each pixel in the image can be converted into a position in wafer coordinates.Knowing the die structure from the computer-aided design (CAD) layout, the dense, semi-dense, and isolated void areas of the captured image with various void sizes can be inspected to detect the voids and measure their dimensions, and wafer or die coordinates and brightness information can be extracted.
In the case where unwanted voids are present on an intentional pattern, the structure of the pattern can also be recognized on the wafer.By scanning each of the dies to be measured and merging them into one master die, the intentional die pattern can be removed; by subtracting the master pattern from the individual die maps, residual void images are generated.Each significant feature from the residual images is treated as a defect, which can be classified according to the properties that are extracted from the captured image such as position and intensity distribution.General characteristics of voids found in these cases are approximately circular in shape and have intensity distributions that are dependent on the void size and depth.
Sample Description
The wafer set consists of four bonded wafer pairs, which are formed using standard thickness (775 μm) 300-mm wafers that are bonded together using a 5-nm oxide layer, using the process flow as described earlier.As shown in Fig. 5, one of the wafers in each pair was patterned with programmed voids.These programmed voids have diameters ranging from 0.5 to 300 μm.The voids on each of the four wafer pairs were etched to different depths ranging from 40 to 1200 nm.In addition, the voids are present in isolated, semi-dense, and dense formats, as shown in the CAD layout included in the left part of Fig. 6.
Experimental Results
The right part of Fig. 6 shows the captured image on a die with approximately 400-nm deep voids.The image capture was performed using a lens objective with 5× magnification Fig. 7 (a) Original scale images of 40-nm deep voids, captured with a 5× objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom.For this void depth, a 10-μm detection limit was observed for isolated areas, while all semi-dense and dense voids can be detected but cannot be individually resolved.(b) Original scale images of 400-nm deep voids, captured with a 5× objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom.For this void depth, all voids were detected for isolated, semi-dense, and dense areas but all of them were not individually resolved.(c) Original scale images of 800-nm deep voids, captured with a 5× objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom.For this void depth, all voids were detected for isolated, semi-dense, and dense areas but all of them were not individually resolved.(d) Original scale images of 1200-nm deep voids, captured with a 5× objective: 2.5-, 5-, 10-, and 15-μm lateral sizes from top to bottom.For this void depth, all voids were detected for isolated, semi-dense, and dense areas but all of them were not individually resolved.
or 8 μm∕px pixel size.In this configuration, the image size obtained by one scan is 3600 × 1024 pixels.To image one die, five overlapping scans were performed, of which the middle 3300 × 737 pixel area was used for each scan.Therefore, the resulting matched image (Fig. 6, right) is 3300 × 3685 pixels.From the image, all the voids in semi-dense (5∶1 space:width) and dense (1∶1 space: width) arrangements are detected, but for smaller void sizes, the individual voids cannot be resolved.From Fig. 7, which is showing images from the same image capture at original image scale, we can see more clearly how the smallest resolvable void depends on density and lateral dimensions.As the voids get smaller and thinner, they become more challenging to detect in isolated arrangement and to resolve in regions with increased void density.
A summary of the smallest detectable, resolvable, and measurable voids for samples with varying void depths is included in Table 1.For the 5× lens objective measurements, the 40-nm deep voids were observed to have a 10-μm detection limit for isolated areas.As the void depth increases to 400 nm or thicker, the smallest detectable void decreases to 2.5 μm for isolated voids, whereas in case of semi-dense and dense areas the detection limit is 0.5 μm.The smallest resolvable 40-nm deep void lateral sizes were found to be 10 μm for semi-dense and 15 μm for dense areas, while a 5-μm resolution limit was found for all other depths.The smallest automatically measurable void size for 40-to 800-nm deep voids was 25 μm for all regions.As it is seen in Fig. 4, 1200-nm deep voids have lower contrast with respect to the background, which leads to unstable measurements of the size of the 25-μm voids, and therefore a 50-μm measurement limit is recorded for this depth.Using the 50× magnification objective lens (0.8 μm∕px pixel size), voids as small as 0.5 μm can be detected.In semi-dense areas, voids as small as 0.5 μm can be resolved, while only 2.5 μm ones can be resolved in dense areas.Automatic size measurements were not performed using 0.8 μm∕px images.Lateral void dimension measurement fidelity was evaluated by correlating the lateral dimensions measured by the IR microscopy instrument with the nominal (design) dimensions.As shown in Fig. 8, the results are in good agreement with R-square >0.99 for all the void depths.From the results, we also note that the slope is generally close to unity, indicating that the instrument's pixel size calibration is in good agreement with the nominal designed dimensions.One exception is the 1200-nm deep voids, for which the slope is 0.94.This is not yet fully understood.Further work is ongoing to measure the sister wafers using other techniques, which may help to give a better understanding.While the intentional voids greater than 1-μm deep that are etched in silicon for the DOE wafers are interesting for exploring instrument sensitivity, the bonding voids that pose practical issues in the semiconductor fab manufacturing line are typically formed in the bonding interface.
In order to further explore the sensitivity to void depth, reflectance simulations were performed for void depths on the DOE wafers (40, 400, 800, and 1200 nm). 7,8For the simulations, a 0-deg angle of incidence was used with a filmstack from bottom to top as follows: infinite silicon substrate, air (varied thickness), 5-nm silicon dioxide, and 775 μm of silicon.As shown in the simulation results in Fig. 9, reflectance is expected to increase from 40 to 800 nm, but for 1200-nm depth, antireflective behavior is observed with lower average intensity.It should be noted that in this study, the focus has been on voids that are formed in the lower part of the oxide bonding layer.Additional future work exploring the effect of bonding interface void depth on instrument sensitivity could be beneficial.
Results shown in Fig. 10 indicate that the intensity contrast is sensitive to the void depth, and good correlation is observed.The contrast was calculated as the normalized range in intensity measured across the voids.A future study may be performed to determine whether the measured contrast together with precalculated reflectance versus void thickness data can be reliably used to determine the void thickness.
Conclusions
The IR microscopy has been evaluated for detecting and measuring the dimensions of voids that are formed in the oxide-to-oxide interface of permanently bonded programmed void wafers.Results have demonstrated that the technique has the required sensitivity to detect and measure the isolated and dense voids of horizontal dimensions varying in range from submicron up to hundreds of micrometers and depths varying from 40 up to 1200 nm.The capability to acquire the images, inspect them for void defects, and measure the found voids with reasonable speed makes the technique a good solution for high-volume manufacturing implementation in semiconductor fabs.
The IR microscopy instrument is currently being integrated into a fully automated platform in order to build a tool suitable for tier 1 fab production implementation.The method is expected to have immediate application to temporary bonding applications, where the imaging can be done through the carrier wafers.Additional work will be needed to investigate the applicability to permanent bonding applications, where the presence of surface metallization and TSVs (for via-first and via-mid processes) may prevent imaging of the bond plane.Work is also being pursued to explore whether combining IR microscopy with other techniques, such as photoluminescence or model-based Fourier transform infrared (FTIR) reflectometry, 9 will be beneficial for improved inspection and metrology capability in a manufacturing metrology tool.Future work to take advantage of algorithms previously developed is desirable to more accurately report dimensions. 10,11onny Höglund is managing the Semilab USA applications group and is working with metrology and inspection equipment used for material characterization and process monitoring, primarily in the semiconductor and photovoltaic industries.In 2004, he joined Philips AMS, where he worked with surface acoustic wave and MBIR metrology, and he is working with Semilab since the merger in 2009.He has been active in the semiconductor industry for 13 years and previously worked with ASML.
Zoltan Kiss joined the optics group of Semilab in 2010 as a developer working with metrology and inspection equipment used for material characterization and process monitoring, primarily in the photovoltaic and semiconductor industries.Between 2001 and 2010, he worked in Hungarian Astronomical Research Institutes on Measurement Automation, Development, Image and Data Analysis in international cooperation.He gained a PhD degree on related topics in the Eötvös University, in 2009.
Gyorgy Nadudvari leads the optics development team at Semilab Inc., working on optical inspection and metrology technologies both in semiconductor and photovoltaic industries.In 2007, he joined Semilab, and runs projects including infrared inspection of bulk microdefects in silicon, photoluminescence, and photomodulated reflectance metrology.Earlier, he worked in the area of optical development, inspection, and manufacturing in other companies, including Philips Optical Storage.He holds an MSc degree in engineering physics.
Zsolt Kovács has worked at Semilab, Hungary, since 2012 in the optics group and is dealing with the development of optical methods for the characterization of material and manufacturing defects in semiconductor or photovoltaic samples.Chris Moore received his PhD in physics in 1983 and after a brief period of teaching, where he concentrated on instrument development and measurement technologies, he was one of the founders of Waterloo Scientific Inc. in 1985.He has worked in all facets of the metrology industry from product development through marketing and sales to technical and business management.His last position was with Semilab USA as president and CEO.
Victor Vartanian is a metrology engineer at SEMATECH in the 3D Interconnect Division.Prior to joining SEMATECH, he worked at Motorola/Freescale on applications of strained silicon to transistors.Previously, he worked on applications of FTIR and mass spectrometry to environmental issues in semiconductor manufacturing and in process development.He received his BS and PhD degrees in chemistry from the University of Texas at Austin.He has numerous publications and 10 patents in the semiconductor field.
Richard A. Allen is a physicist at NIST, where his research focuses on metrology for 3D stacked ICs and MEMS.He leads the SEMI 3DS-IC Standards Committee and was NIST assignee to SEMATECH's 3D Enablement Center from 2011 to 2013.Prior to joining NIST, he developed test methods for in situ characterization of space radiation effects at JPL.He received the BS and MS degrees from RPI and his MBA degree from Columbia Union College.
Fig. 2
Fig. 2 Wafer flipper used to facilitate the imaging through the backside of wafers.
Fig. 3
Fig. 3 Principle of IR microscopy: the mismatch in refractive index due to the presence of a void results in greater reflection.
Fig. 6
Fig.6CAD layout and captured image of designed defects using a lens objective with 5× magnification, corresponding to an 8 μm∕px pixel size.
Fig. 8
Fig. 8 Correlation between lateral dimensions measured by the IR microscopy instrument versus the nominal dimensions based on the CAD layout.These results are based on data of isolated voids.Measurements on top left voids (and their nearest neighbors) in semi-dense and dense arrangements resulted in same values within measurement error.
Fig. 9
Fig. 9 Simulation results showing spectral sensitivity to void depth.
Fig. 10
Fig. 10 Correlation between simulated reflectance and measured contrast.
From 2009 to 2012, he worked at the Research Institute for Solid State Physics and Optics in Budapest, dealing with optical investigation of flowing phenomena in granular materials.He gained the MSc degree as a physicist in the Eötvös University, in 2012.Szabolcs Velkei has managed the software development group in Semilab since 2013.He is heavily involved in R&D of image processing techniques and data processing.He established his own company in 2005 and is primarily concerned with artificial intelligence development.He joined Semilab in 2012.
Table 1
Summary of the smallest detectable, resolvable, and measurable voids for samples as a function of void depth and density. | 6,284.6 | 2014-01-01T00:00:00.000 | [
"Engineering",
"Materials Science",
"Physics"
] |
Reservoir-engineered spin squeezing: macroscopic even-odd effects and hybrid-systems implementations
We revisit the dissipative approach to producing and stabilizing spin-squeezed states of an ensemble of $N$ two-level systems, providing a detailed analysis of two surprising yet generic features of such protocols. The first is a macroscopic sensitivity of the steady state to whether $N$ is even or odd. We discuss how this effect can be avoided (if the goal is parity-insensitive squeezing), or could be exploited as a new kind of sensing modality to detect the addition or removal of a single spin. The second effect is an anomalous emergent long timescale and a"prethermalized"regime that occurs for even weak single-spin dephasing. This effect allows one to have strong spin squeezing over a long transient time even though the level of spin squeezing in the steady state is very small. We also discuss a general hybrid-systems approach for implementing dissipative spin squeezing that does not require squeezed input light or complex multi-level atoms, but instead makes use of bosonic reservoir-engineering ideas. Our protocol is compatible with a variety of platforms, including trapped ions, NV defect spins coupled to diamond optomechanical crystals, and spin ensembles coupled to superconducting microwave circuits.
I. INTRODUCTION
Among the most sought-after states in quantum metrology are spin-squeezed states, highly entangled states of spin-1/2 ensembles that enable parameter sensing with a sensitivity better than the standard quantum limit, even reaching fundamental Heisenberg-limit scaling [1,2]. The standard approach for producing these states is to unitarily evolve an initial product state under a collective spin-spin interaction Hamiltonian. While many interactions are possible, the most widely studied one is the one-axis twist (OAT) Hamiltonian [1], which has been realized in a number of ground-breaking experiments [3][4][5][6]. It unfortunately is not capable of achieving Heisenberg-limited squeezing even in the ideal case [2]. An alternate, more complex interaction Hamiltonian is the two-axis twist (TAT) Hamiltonian [1,[7][8][9][10], which, while more resource intensive, allows achieving Heisenberg-limited scaling.
While easy to understand, tailored unitaryevolution is not the only approach to spin squeezing. An alternative is to use the general strategy of reservoir engineering [11], where tailored dissipation is exploited to both produce and stabilize a nontrivial state of interest, i.e., a spin-squeezed state (see Fig. 1). The dissipative approach in principle has several advantages: the spin-squeezed state is stabilized in the steady state (as opposed to just prepared at a specific instant of time), the stabilization is largely insensitive to the initial state of the ensemble, and one can achieve Heisenberg-limited scaling.
FIG. 1. Schematic representation of a generic approach to generate dissipative spin squeezing by coupling spins to a bosonic mode that interacts with a squeezed reservoir. The squeezing rate experienced by the cavity is governed by the parameter κsqz, while g represents the spin-cavity coupling strength. Limiting factors to the protocol's performance are the intrinsic photon-loss rate κint, the local spin-relaxation rate γ rel , and the local spin-dephasing rate γ φ .
The dissipative stabilization of bosonic squeezed states has been studied extensively both theoretically [12][13][14] and experimentally [15][16][17][18][19][20]. Corresponding schemes for spin squeezing have also been studied theoretically. The earliest works analyzed schemes where atoms are directly illuminated with squeezed light. Both the cases of two-level atoms [21][22][23] and V -type multilevel atoms [24] were studied. More recently, it was shown theoretically that the same effective dissipative dynamics could be realized by using Raman processes in driven multilevel atoms coupled to a lossy cavity [9,25].
In this work, we revisit the dissipative approach to spin squeezing. Our work complements previous studies both by discussing a powerful, alternative method for implementing these schemes, as well as describing surprising phenomena that had not been fully analyzed in the past. In terms of implementation, we analyze a very general hybridsystems approach that harnesses bosonic dissipative squeezing. We consider a spin ensemble which is resonantly coupled to a cavity mode (via a standard Tavis-Cummings [26] interaction), which is in turn coupled to an effective squeezed reservoir (see Fig. 1). Previous proposals [22,24] suggested to implement this squeezed reservoir by driving the cavity with squeezed light, an approach which is limited by losses associated with the transport and injection of an externally prepared optical squeezed state. We show that there are also simpler methods to generate the effective squeezed reservoir, which can be implemented using only classical optical or microwaves drives by harnessing existing dissipative bosonic squeezing schemes. Specifically, we consider coupling the cavity to an auxiliary lossy degree of freedom (two level system or bosonic mode) which is driven simultaneously with imbalanced red-detuned and blue-detuned sideband drives. Such schemes produce an effective squeezed dissipator for the cavity and have been experimentally implemented in wide variety of platforms, including optomechanics [15], trapped ions [16] and superconducting circuits [20]. Since only classical radiation is required, this approach is insensitive to the aforementioned transport and insertion losses of squeezed radiation. We demonstrate that this hybrid-systems approach to dissipative spin squeezing can reach the Heisenberg limit, and also outperform OAT in the presence of single-spin T 1 decay. Note that unlike the Raman scheme of Ref. 25, which requires atoms with a specific four-level configuration, the approach here only requires standard two-level atoms, making it compatible with a wide variety of systems (including possibly solid-state systems such as ensembles of NV defect spins in diamond [27]).
Our work also analyzes surprising phenomena that were not fully discussed previously. Perhaps most striking is the extreme sensitivity of dissipative spin squeezing to the parity of the total number of spins N : the steady state is macroscopically different for N spins versus N + 1 spins. While this effect was implicitly contained in the results of Agarwal and Puri [21,22] (see Sec. VIII for a detailed discussion of the relation to previous works), we provide here a fully qualitative and quantitative analysis. We discuss how this effect can be avoided (if one wants strong spin squeezing independent of parity), and how it could also be exploited as a new kind of sensing modality. We also make a surprising connection to a non-dissipative many-body system, the antiferromagnetic Lipkin-Meshkov-Glick (LMG) model [28,29].
A second surprising and new phenomenon we describe is the interplay between collective dissipative-spin-squeezing dynamics and noncollective singlespin dephasing. As we show, this results in an extremely long relaxation timescale in the system (i.e., inverse dissipative gap) which grows with system size N . At a fundamental level, the effect has parallels to prethermalization behavior observed in weakly nonintegrable systems (see, e.g., [30,31]). At a practical level, we show that even infinitesimally weak single-spin dephasing dramatically impairs the steady-state spin squeezing to at most −3 dB. However, we also show that this need not be a limitation: large amounts of squeezing are possible in the prethermalized regime, i.e., at transient times parametrically shorter than the timescale required to reach the steady state, or by deliberately adding very small levels of single-spin relaxation. This effect may allow one to quickly generate strong squeezing even in parameter regimes that had previously been discarded based on the low level of steady-state squeezing.
Finally, we analyze the impact of imperfections in the reservoir-engineering process that lead to the engineered squeezed dissipation having a non-zero impurity and effective thermal occupancy. We reveal a striking sensitivity of spin squeezing to such imperfections if the squeezing strength is made too large.
The remainder of this paper is organized as follows: In Sec. II, we outline the key idea behind the standard approach to dissipative spin squeezing as well as summarize our generic protocol. In Sec. III, we explore the even-odd effect and briefly discuss connections to sensing. In Sec. IV, we carefully analyze the performance of our dissipative spin squeezing protocol in the presence of single-spin dissipation, showing that the steady-state squeezing it generates can outperform the transient squeezing produced by standard OAT. In Sec. V, we discuss the emergence of anomalously slow relaxation times and we introduce a dynamical-decoupling protocol to cancel the effect of inhomogeneous broadening. In Sec. VI, we analyze imperfections in the reservoir-engineering process, while in Sec. VII, we discuss in more detail how our protocol could be implemented in a variety of different physical systems. In Sec. VIII, we review previous works on dissipative spin squeezing and discuss their relation to our new findings. Conclusions and a summary are presented in Sec. IX.
II. MODEL AND THE BASIC DISSIPATIVE SQUEEZING PROTOCOL
The reservoir engineering approach to spin squeezing requires one to construct a nontrivial dissipative environment for the spins. In this section, we review the idealized spin-only quantum master equation that describes the needed dissipative dynamics [21,22]. We then present a more realistic model that corresponds to the generic, experimentally-friendly hybrid-systems setup sketched in Fig. 1, where a spin ensemble is coupled to a cavity (or other bosonic mode), which is in turn coupled to an engineered squeezed reservoir.
Throughout this paper, we quantify the amount of metrologically-useful spin squeezing (i.e., as relevant to a standard Ramsey measurement) using the Wineland parameter [2,32]. It is defined as where ∆Ŝ 2 ⊥ is the minimum variance in a direction perpendicular to the direction of the mean of the collective spin andŜ ≡ (Ŝ x ,Ŝ y ,Ŝ z ) is the vector of spin operators.
A. Idealized spin-only model
We consider the following quantum master equation acting on the Hilbert space of N spin-1/2 particles,ρ where we introduced the operator Σ[r] = cosh(r)Ŝ − − sinh(r)Ŝ + .
Here, Γ is the coupling rate to the engineered reservoir, r characterizes the squeezing strength, and D[ẑ]ρ =ẑρẑ † − {ẑ †ẑ ,ρ}/2 is the standard Lindblad dissipative superoperator. We also introduced the collective spin operatorsŜ ± =Ŝ x ± iŜ y witĥ denotes a standard Pauli matrix acting on the jth spin. Here,Σ[r] is analogous to a standard bosonic Bogoliubov annihilation operator, where bosonic raising and lowering operators have been replaced byŜ + andŜ − respectively. Similar to reservoir-engineered bosonic squeezing [13], the desired squeezed state will correspond to the vacuum of this operator and the squeezing parameter r characterizes the amount of squeezing, e −2r , of the vacuum fluctuations.
To be more explicit, Refs. 21 and 22 showed that, for even N , Eq. (2) has pure steady states that correspond to zero-eigenvalue eigenstates (i.e., "dark states") ofΣ[r],Σ Since Eq. (2) conserves the total angular momentum j, there is a dark state for each allowed value of j. Each |ψ dk [j; r] has a mean spin polarization in the z direction, and exhibits squeezing (anti-squeezing) ofŜ y (Ŝ x ). The choice of the squeezing axis is determined by the relative phase between theŜ + and S − terms in Eq. (3), which is chosen here to be −1.
If the system is initialized in an arbitrary state with a definite value of j, the dissipative dynamics will relax the system to a dark state in this subspace. For states in the maximum-angular-momentum subspace j = j max = N/2, the relaxation timescale (i.e., the inverse dissipative gap of the Liouvillian) is ∝ 1/N Γ, see Sec. III. Note that the dark states with j < j max are not unique, since the corresponding angular-momentum subspaces are degenerate [33]. However, if the initial state and the dynamics are invariant under permutation of spins, the system will only explore permutationally invariant states [34], and there is a unique dark state for each j subspace, see App. E 1.
As detailed in App. D, the dark states can be expressed in the form [22,23] where |j, m y denotes an eigenstate ofŜ 2 andŜ y , N (r) is a normalization constant, and we defined θ = ln tanh(r). In terms of the eigenstates |j, m ofŜ 2 andŜ z , these states read as follows.
where every second coefficient is nonzero, for k ∈ {0, ..., j}, and all other coefficients vanish, c −j (r) serves as a normalization constant. The parameter r controls the amount of squeezing in the steady state. If we initialize the system in an arbitrary state with j = N/2, the resulting pure steady state is squeezed, with ξ 2 R → 2/(N + 2) in the large-r limit. This corresponds to Heisenberglimited spin squeezing, and thus outperforms both the standard quantum limit (i.e., ξ 2 R = 1) as well as the maximum squeezing possible with an ideal OAT interaction (ξ 2 R ∝ 1/N 2/3 ). Note that a standard leading-order Holstein-Primakoff approximation could be used to map Eq. (2) to a bosonic squeezing dissipator; however, this would not let one understand the ultimate saturation of squeezing (with increasing r) to the Heisenberg-limited value.
B. Hybrid-systems approach to dissipative spin squeezing As noted in the introduction, previous studies have analyzed methods for realizing the dissipative dynamics in Eq. (2). These methods either required direct driving of spins with squeezed light [22,24] (which is experimentally challenging), or the use of Raman processes in structured four-level atoms [9,25] (which is not applicable to generic two-level systems). We present here an alternate, generic method that takes a hybrid-systems approach: a cavity (or other bosonic mode) is coupled both to an ensemble of two-level systems, as well as to an engineered, bosonic squeezed reservoir (see Fig. 1). As discussed, such a bosonic squeezed reservoir can be realized using only classical driving fields, and has been implemented in a variety of different experiments [15,16,20]. We discuss specific implementation strategies of this general approach in Sec. VII; here, we present the general structure of the overall quantum master equation.
To this end, we consider a spin ensemble that is resonantly coupled to a bosonic mode (with lowering operatorâ). In the rotating frame, the Hamiltonian isĤ where g is the spin-cavity coupling strength. We further assume that this mode is coupled both to an engineered squeezed reservoir (with coupling rate κ sqz and squeezing parameter r) as well as subject to unwanted zero-temperature loss (at rate κ int ). The quantum master equation is theṅ We have also included standard single-spin decay and dephasing dissipators (at rates γ rel and γ φ , respectively).
At a heuristic level, the cavity serves as a transducer that allows the spins to inherit the squeezed fluctuations produced by the bosonic squeezed reservoir. As the squeezed reservoir is engineered, we will treat r and κ sqz as tuneable parameters that can be optimized. In contrast, we will take the coupling g and the unwanted dissipation (i.e., κ int , γ φ , and γ rel ) to be fixed. This then motivates introducing singlespin cooperativities η k and collective cooperativities C k via: where k ∈ {φ, rel}. The goal will be to understand the optimal squeezing possible for a fixed value of C k . As we will show in Sec. IV, in the case where single-spin relaxation dominates over dephasing, the optimized dissipative scheme achieves steady-state squeezing scaling as ξ 2 R ∝ 1/ √ C rel . This is significantly better than the optimized transient OAT squeezing in this regime, which only scales as [35]. To connect our setup to the simpler quantum master equation (2), we consider the regime where the condition √ N g κ int +κ sqz holds, and we adiabatically eliminate the cavityâ. We obtain (see App. A) where we have defined and We see that the internal loss of the cavity results in a collective relaxation process for the spin ensemble; this is similar to OAT-based protocols that are derived using a strongly detuned cavity-spin ensemble system (in contrast to the resonant regime considered here).
A. Basic effect
A striking feature of the purely dissipative dynamics described by Eq. (2) is an extreme sensitivity to the parity of the number N of spins. As we will see, the steady state can be macroscopically different for N spins vs. N + 1 spins. While early work noted that the form of the steady state depends on parity [21,22], subsequent studies on achievable squeezing focused on the even-N case [23,25]. Our work reveals important new aspects of this parity effect. We show that by appropriate parameter tuning, one can avoid this effect, allowing steady-state squeezing that is near Heisenberg limited regardless of the parity of N . We also discuss a different regime where the even-odd effect could be used for a new sensing modality based on the macroscopic sensitivity to spin-number parity. Crucially, we show that there is Sketch of the steady state for (a) even N and (b) odd N . The size of the black circles represents the population of a level |j, m . For even N , a pure dark state exists for any squeezing parameter r because the jump operatorΣ leads to destructive interference between adjacent levels (blue arrows) such that every second level is unoccupied (dashed red lines). For odd N and large r, the interference condition cannot be satisfied for all levels (brown flashes) and the steady state is mixed. The two pure-state contributions with largest statistical weight are sketched here.
no long timescale associated with the emergence of this sensitivity to the addition or removal of a single spin. Note that the even-odd effect in dissipative spin squeezing has no counterpart in bosonic dissipative squeezing. We start with a simple intuitive picture that explains why the steady state of Eq. (2) is so sensitive to the parity of N . Recall that pure bosonic squeezed states are fully paired: they are superpositions of states having even photon numbers only [36]. A similar structure holds in our spin problem. We can think of the fully polarized state |j, m = −j as being the "vacuum", and a state |j, m = −j + q as having q excitations (i.e., q flipped spins). We thus see directly from Eq. (7) that, like bosonic squeezed states, the spin dark states |ψ dk [j; r] also only involve even numbers of excitations q.
Formally, in both the bosonic and spin problem, this paired structure leads to destructive interference that makes the state dark. WhenΣ[r] acts on a paired state, it creates a state having only odd number of excitations. For a given odd excitation number q odd , achieving a dark state requires destructive interference between the two pathways leading to q odd :Ŝ − could have acted on the state with (q odd + 1) excitations, orŜ + could have acted on the state with (q odd − 1) excitations. These destructive interference conditions can be directly used to derive the coefficients in Eq. (7) that determine |ψ dk [j; r] . This structure is shown schematically in Fig. 2
(a).
With this picture in mind, it is easy to see why we cannot have a pure dark state for odd N . In this case, the maximum number of excitations q max is odd. As such, the needed destructive interference is impossible to achieve. Starting with a fully paired state, we can create a state with q max excitations by acting withŜ + on |j, −j + (q max − 1) . However, there is no complementaryŜ − process, as there is no state with q max + 1 excitations. The best one can then do is to construct fully paired states that are only approximately dark due to this incomplete destructive interference [see Fig. 2 The net result of this "frustration" is dramatic: for odd N and large r, the dissipative steady state of Eq. (2) is impure and, moreover, exhibits no spin squeezing for large r. More specifically, for odd N , the steady-state squeezing diverges in the large-r limit, while the purity tends asymptotically to 1/3. This behaviour is shown explicitly in Fig. 3. One also sees that, for modest r, there is no appreciable even-odd effect: the odd-N steady state is almost pure and has the same squeezing as the even N case. This also follows from our heuristic picture: for small enough r, there is very little probability to have a large number of "excitations", and hence one is almost insensitive to the frustration resulting from the cut-off on maximum excitation number.
While our discussion has been focused on the ideal quantum master equation (2), the even-odd effect persists even in the presence of single-spin relaxation and dephasing [as described by Eq. (11) in the limit κ int → 0]. As discussed in App. D, observing the even-odd effect in the steady state requires the single-spin cooperativities η rel and η φ defined in Eq. (10) to be order unity or larger.
Finally, we note that the even-odd effect discussed here is distinct from the sensitivity to parity exhibited by unitary evolution under a OAT Hamiltonian H OAT = χŜ 2 x [37][38][39]. The unitary evolution generated byĤ OAT for a time π/2χ maps the initially fully polarized state |N/2, −N/2 to Greenberger Horne Zeilinger (GHZ) states oriented along orthogonal axes in phase space, depending on the parity of N . This coherent effect results in a strong sensitivity to parity at a particular instant in time; in contrast, in our system, we have a dissipative effect where the sensitivity manifests itself in the steadystate of the system. Moreover, in our case, the even vs. odd states are not equivalent up to a rotation, but differ both in their purity and the magnitude of their fluctuations. N and its purity (dash-dotted red line) approaches 1/3. Inset: Minimum Wineland parameter for even N (crosses) and odd N (dots) obtained at an optimal squeezing strength ropt. The squeezed spin component is alwaysŜy.
B. Parity-independent Heisenberg-limited squeezing
In most experimental situations, the even-odd effect will be a nuisance: one aims for strong steadystate squeezing without needing to control N at the single particle level. We therefore derive a quantitative estimate on the maximum squeezing parameter r that can be used without any parity sensitivity. For small r, the system and its steady state are well described by a Holstein-Primakoff approximation [40]; one recovers bosonic squeezing physics [41], which is independent of the parity of N . However, the correspondence between bosonic squeezing and spin squeezing will break down if the populations of the states |j, m ≈ j become nonzero. Using the steady-state occupation number of the Holstein-Primakoff bosons b †b ss = sinh 2 (r), one can estimate that this breakdown happens if the condition b †b ss ≈ N/2 holds. This yields the breakdown criterion which provides an estimate for the maximum squeezing parameter r possible with no even-odd effect. While working with e 2r N avoids parity effects, one might worry that this constraint precludes ever reaching Heisenberg-limited scaling of the steadystate spin squeezing. This is not the case. As shown in the inset of Fig. 3, the minimum Wineland parameter for odd N (obtained at a squeezing parameter r opt , which depends on N ) exhibits Heisenberg-like scaling, and the spin squeezing differs only by a constant prefactor ≈ 2.6 from the maximum achievable spin squeezing of ξ 2 R,HL = 2/(N + 2) for even N , which is obtained in the limit e 2ropt N .
C. Connections to the LMG Model
Despite first appearances, the extreme even-odd effect of our system is more than a nuisance. At a fundamental level, the effect has a surprising connection to a seemingly unrelated closed-system manybody model, the LMG model [28]. To see this, recall that in a quantum trajectories formulation of the master equation in Eq. (2), the evolution of a state vector in the absence of quantum jumps is governed by the non-Hermitian Hamiltonian (−i/2)Ĥ LMG , wherê x + e 2rŜ2 y +Ŝ z . (15) H LMG is precisely the Hamiltonian of the anisotropic antiferromagnetic LMG model [28], a generalized transverse field Ising model with all-to-all Ising couplings. For even N , we are thus dissipatively stabilizing the many-body ground states |ψ dk [j; r] of the antiferromagnetic LMG model [29,42] and converge to one of them depending on the total angular momentum j of the initial condition. Note that the physics here has crucial differences from the more studied ferromagnetic LMG model, which is also known to exhibit spin squeezing in its ground state [43,44] (but has no simple connection to a dissipative protocol). Focusing on the case where N is odd and r > 0, H LMG is positive and the steady state of Eq. (2) in a given total-angular-momentum subspace j can be written as (see App. D) where λ k and |ψ k are the ordered eigenvalues and eigenvectors ofĤ LMG . We can thus directly connect the properties of the odd-N steady state to the spectrum of the LMG Hamiltonian. Consider first the limit r → 0, whereĤ LMG →Ŝ 2 −Ŝ 2 z +Ŝ z . Then, the Hamiltonian has a unique ground state |ψ 0 → |ψ dk [j; 0] = |j, −j . Moreover, the groundstate energy is zero for any N and the gap to the double-degenerate first excited states is finite, i.e., lim r→0 λ 0 = 0 and lim r→0 λ 1,2 = 2j. As a result, the steady state is approximately pure even when N is odd, as |ψ 0 dominates the sum in Eq. (16).
In the opposite limit r → ∞, the LMG Hamiltonian is dominated by theŜ 2 y term,Ĥ LMG ≈ e 2rŜ2 y , and its eigenvalues are the eigenstates |j, m y ofŜ y with energy λ m ≈ m 2 e 2r . Now, there is no zero energy ground state for odd N (because m takes halfinteger values), the ground state is double degenerate, and the steady state converges to an incoherent mixture ofŜ y eigenstates, A direct computation shows that the purity converges to lim N →∞ Tr (ρ (N/2) ss ) 2 = 1/3. In the limit r → ∞, there is no mean spin polarization, but the variance ofŜ y remains finite, Ŝ 2 y ≥ 1/4. As a result, the Wineland parameter will diverge as shown in Fig. 3.
The connection to the LMG model thus provides useful intuition into the odd-N steady state. For even N , the dark state |ψ dk [j; r] remains an exact zero mode ofΣ † (r)Σ(r) for any value of r and interpolates smoothly between the limits |ψ dk [j; 0] = |j, −j and lim r→∞ |ψ dk [j; r] = |j, 0 y . In terms of the LMG model, this implies that for even N , the ground-state gap does not close as a function of r [45]. This feature of the antiferromagnetic LMG model has been discussed previously in the context of a closed system quantum phase transition [29,42,46].
D. Enhanced sensing
The dramatic even-odd sensitivity of the steady state, which has no counterpart in bosonic spin squeezing, could enable a new kind of sensing modality: it provides a means for detecting changes in N at the single-spin level. This kind of sensing has long been of interest for both fundamental studies and applications [47][48][49][50][51][52]; a recent experiment has even used dispersive sensing to measure real-time changes in atom number in an atomic ensemble dispersively coupled to a cavity [53]. Our dissipative setup could provide an alternative route for an analogous kind of sensing.
As discussed above, for a large squeezing parameter e 2r N , the squeezing of the collective steady state depends exponentially on the parity of N (see Fig. 3). A simpler quantity, the variance ofŜ y , also exhibits this strong sensitivity in the large r limit: for even N , it vanishes like N 2 e −4r /8 whereas, for odd N , it converges to the constant value N/π 2 if N 1. We thus see that measuringŜ 2 y provides a direct means for estimating the parity of N . Such collective spin fluctuation measurements have been implemented in variety of systems [54][55][56][57][58][59][60].
While the parity sensitivity is in principle a steady-state effect, the relatively fast relaxation timescale here means that it can be harnessed for real-time sensing. We stress that the strong sensitivity to parity does not come at the expense of a vanishingly small bandwidth: if a spin is suddenly lost, the relaxation time to the new opposite-parity steady state is (at worst) set by the inverse coupling rate 1/Γ. This timescale does not grow with system size [see inset of Fig. 4(a)]. The relaxation is even faster if one is in the maximum-j subspace; here, the relaxation rate is collectively enhanced by a factor of N .
We thus have a means for detecting spins leaving or decoupling from the cavity one by one, as each such event causes a large change inŜ 2 y [see Fig. 4(b)]. Note that the variance detection requires multiple repetitions of the measuement to distinguish even N from odd N : although the probability to obtain a measurement result with |m y | > 1/2 is negligible for even N in the large-r limit, it is only between 15 % and 19 % for odd N [cf. Eq. (16) and App. D]. One thus has to wait for a probabilistic measurement outcome with sufficiently large |m y | (the value depends on the detector resolution) to determine the spin-number parity unambiguously.
Imperfections of the squeezing process, e.g., an impure engineered reservoir, and local dissipation will reduce the visibility of the even-odd effect (see Sec. VI and App. D, respectively). In Sec. VII D, we discuss that even-odd effect can be observed in a state-of-the-art trapped-ion platform for N 10. This opens the exciting possiblity to experimentally verify the even-odd effect in spin squeezing, which has no counterpart in bosonic squeezing.
IV. ENHANCED PROTECTION AGAINST SINGLE-SPIN RELAXATION
The dissipative approach to spin squeezing also provides strong advantages when unwanted singlespin dissipation is included. In this section, we focus on the case where local relaxation is dominant, i.e., we study Eq. (11) in the limit γ rel = 0, γ φ → 0. For atomic systems, this can be viewed as a fundamental limit arising from spontaneous emission, whereas single-spin dephasing is a technical imperfection. As noted in Ref. 35, in this limit, standard OAT achieves an optimized squeezing that yields the scaling ξ 2 R ∼ C −1/3 rel for large N . This work also introduced an alternate Hamiltonian protocol involving two mutually-interacting spin ensembles, which could achieve a more favourable ξ 2 y is plotted for a system described by the ideal quantum master equation (2) with r = 2.5, starting from the state |N/2, −N/2 . The evolution is interrupted at randomly chosen times (black triangles), where a single (randomly chosen) spin is removed from the system. These spin-loss events cause the system to relax to a new steady state, leading to dramatic swings in the value of Ŝ 2 y after each loss event. Note the logarithmic scale used for the y axis.
at a specific optimized time. As we show below, our dissipative approach can achieve an identical scaling, but now for the steady state, and only using a single ensemble of standard two-level systems. We also show that this enhanced performance over OAT holds even for small-N ensembles. Note that singlespin dissipation was also studied in Ref. 25, but only for spontaneous emission in an ensemble of four-level atoms with a specific structure. This is distinct from the more generic model Eq. (11) we study.
Focusing on the limit of large N and a small singlespin cooperativity, we can approximate our system well using a standard mean-field theory based on linearizing the equations of motion for the system's covariance matrix. Solving these in the steady state and considering the limit of a sufficiently large r (see App. C), one finds that the steady-state squeezing is The numerator here describes unwanted heating by both single-spin relaxation and the collective decay γ coll associated with internal cavity loss. The only parameter left to optimize over is κ sqz , the coupling between the cavity and the squeezed reservoir, which enters Eq. (18) via Eqs. (12) and (13). There is a non-trivial minimum here. Suppression of unwanted collective heating requires a large κ sqz , as this reduces the ratio γ coll /Γ. In contrast, suppressing the effects of γ rel requires a large Γ and hence small κ sqz .
Minimizing with respect to κ sqz , we find where the optimal value of κ sqz satisfies We thus obtain an optimized squeezing that scales significantly better with collective cooperativity in this relaxation-dominated regime than the OAT result of ξ 2 rel . In App. C 2 we show numerical simulations of a more accurate non-linear mean-field theory that confirm these results. As we have stressed, the squeezing here is also achieved in the steady state (and not just at one optimal time). While we assumed a large value of r to derive these results, in practice one only needs exp(−2r) 1/ √ C rel for this scaling to hold.
The advantage over OAT in this relaxationdominated regime also persists for smaller-sized spin ensembles. To study this regime, we numerically solve Eq. (11) for the steady state. Figure 5 shows the obtained results for the steady-state squeezing (orange curve) as a function of N , where we have fixed g, κ int , and γ rel so that the single-spin cooperativity is η rel = 2. For each value of N , we optimize the parameters of the squeezed reservoir (κ sqz , r) to minimize the steady state ξ 2 R ; the optimized values are presented in App. F. For comparison, we also plot the optimized transient squeezing achievable using OAT (blue curve) in an identical cavity-spin system [27,35] (see App. G for details). For the OAT setup, there is no squeezed reservoir, κ sqz = 0, and there is a large detuning ∆ between the spins and cavity, which is optimized for each value of N . Figure 5 shows that, even for small N , the dissipative protocol yields an advantage over OAT. While for these small values of N and large η rel , the linearized mean-field theory scaling predictions are not expected to hold exactly, there is a qualitative agreement with the predicted power laws (as indicated by black dashed lines).
In App. H, we provide a brief performance analysis of a special case where κ int = 0. Mathematically, such a scenario is equivalent to a setup where one directly shines squeezed light onto the spin ensemble. We show that in the limit of large spin number, one can achieve the scaling of ξ 2 R ∝ (N Γ/γ rel ) −1 , although naturally, having either κ int = 0 or irradiating a spin ensemble directly, would likely be difficult to realize experimentally.
A. Pre-thermalization and emergent slow timescales
We now consider the effects of weak single-spin dephasing [i.e., the γ φ term in Eq. (11)] on our dissipative spin squeezing protocol. For very weak dephas-ing, such that γ φ < γ rel /N holds, the mean-field theory results of the previous section still provide a good description; one simply substitutes γ rel → γ rel + 2γ φ in Eqs. (19) and (20). The more interesting case is when dephasing is the dominant form of singlespin dissipation, but is still weak compared to the rate Γ associated with the collective spin squeezing dissipator (i.e., Γ γ φ γ rel ). In this case, the dynamics is surprisingly rich, exhibiting features reminiscent of prethermalization behavior observed in weakly nonintegrable systems [30,31]. Prethermalization is associated with approximately conserved quantities that can only be dynamically randomized on extremely long timescales; this results in an intermediate-time quasi-steady state whose form is contingent on the initial value of the conserved quantities. Here, a similar phenomenon arises, with total angular momentum playing the role of the approximately conserved quantity. We discuss this more in what follows.
Starting from an initial product state, we find that a seemingly tiny amount of single-spin dephasing is enough to completely destroy spin squeezing in the eventual steady state. Using a mean-field analysis, one can show that in the presence of arbitrarily weak but non-zero single-spin dephasing (and γ rel = γ coll = 0), the steady-state squeezing is bounded by −3 dB in the large N limit: where the optimal value is achieved with r = 1 8 ln N . Despite this, there exists an extremely long-lived intermediate-time regime (a quasi-steady state) where strong spin squeezing is observed. The system's dissipative dynamics is thus characterized by two vastly different timescales, as shown in Fig. 6. The system first evolves into a transient spinsqueezed state on a fast timescale ∝ 1/N Γ. In contrast, the eventual relaxation to the true steady state (which has minimal squeezing) occurs on a much slower timescale ∝ N/γ φ . For a large system size N , the ratio of these timescales can be dramatic. We also note that the slow relaxation time is parametrically slower than the single-spin dephasing time 1/γ φ .
The emergence of this surprisingly long timescale, and the corresponding fragility of the steady state to weak dephasing, are both surprising; we stress that single-spin relaxation (as discussed in the previous section) does not give rise to an analogous behavior. In App. E, we analyze this effect using Liouvillian perturbation theory [61] and develop an intuitive physical picture: Single-spin dephasing enables transitions between subspaces of different total angular momentum [34] such that an initial state in the (11) for weak local dephasing, r = 1.0, and N = 50 spins (thick dashed orange line, γ φ /Γ = 0.005, γ rel /Γ = 0, and γ coll /Γ = 0). The final amount of steady-state spin squeezing is indicated by the thin dash-dotted orange line. Local dephasing deteriorates the amount of steady-state spin squeezing compared to an ideal system without local dissipation (solid blue line, γ φ /Γ = γ rel /Γ = γ coll /Γ = 0). Local relaxation counteracts this effect and partially restores the steady-state spin squeezing (dotted green line, γ φ /Γ = 0.005, γ rel /Γ = 0.001, and γ coll /Γ = 0). Note that the transient state is strongly spin squeezed even in the presence of local dissipation since the collective dissipator Σ induces spin squeezing on a short timescale ∝ 1/N Γ whereas the system approaches its steady state on a longer system-size-dependent timescale ∝ N/γ φ . (b) Wineland parameter calculated using the mean-field equations of motion detailed in App. B for N = 1000 spins and the same sets of dissipation rates as in (a). j = j max subspace evolves into a steady state populating subspaces with j < j max . The degeneracy of the j < j max subspaces gives rise to anomalously small matrix elements between the subspaces, which represent bottlenecks for the relaxation to the steady state.
We stress that the surprising impact of dephasing need not be problematic for experiments. The spin squeezing exhibited by the Wineland parameter ξ 2 R in the "prethermalized" intermediate-time regime is comparable to ξ 2 R of the steady state obtained in an ideal system without single-spin dissipation, as long as the conditions γ φ Γ, N γ φ are satisfied. Moreover, there is a simple but effective way to improve the spin squeezing of the steady state by deliberately adding a competing single-spin relaxation process γ rel . If this relaxation rate satisfies the condition γ rel γ φ /N , population will be pushed back to the large-angular-momentum subspaces, which decreases the steady-state Wineland parameter significantly and increases spin squeezing beyond the −3 dB limit, as shown in Fig. 6.
B. Inhomogeneous broadening and dynamical decoupling
In addition to the Markovian mechanism described in the previous section, in some platforms dephasing due to inhomogeneous broadening of the spin ensemble can also play a role. A major advan-tage of spin squeezing generated by OAT dynamics is that it is compatible with dynamical decoupling and thus allows for an effective cancellation of the impact of inhomogeneous broadening by a simple sequence of π pulses about the x axis [27]. At first glance, this does not seem to be the case for our dissipative scheme. However, we will show here that our dissipative scheme is in fact compatible with a slightly modified dynamical decoupling sequence.
Our starting point is a generalization of the Hamiltonian of Eq. (9), where ∆ j = ω j − ω 0 denotes the detuning of spin j from the resonance frequency of the bosonic modê a due to inhomogeneous broadening. Using average Hamiltonian theory [62], we will now derive an effective Hamiltonian for the two different decoupling sequences shown in Fig. 7, which are designed such that the effects of the ∆ j terms cancel out on average. Instead of transforming the state of the system at each decoupling pulse, it is more convenient to consider a Heisenberg picture where the Hamiltonian changes at each pulse, the so-called toggling frame [62].
A single π pulse about the x axis will flip the sign ofσ erators in the interaction term of Eq. (22), The first effect is desired and will cancel inhomogeneous broadening, but the second effect is unwanted because it will turn damping of the Bogoliubov modê Σ(r) into anti-damping. If one can control the coupling constant g as a function of time, one can switch off the undesired interaction after every second π pulse, as shown in Fig. 7(a), and obtains the average Hamiltonian where the inhomogeneous broadening has been canceled at the cost of a reduction of g by a factor of 2. Experimentally, the coupling g could be switched off by detuning the spins rapidly from the cavity. Instead of switching off the interaction between the spins and the bosonic mode for half of the period T , one could also use the dynamical decoupling sequence shown in Fig. 7(b). By applying a π rotation about the z axis, one can flip the sign of the second term without disturbing the first one, The sequence is terminated by a π rotation about the y axis which reverts all signs and restores the original Hamiltonian (22). If the waiting times between the pulses have a ratio of 2 : 1 : 1, both inhomogeneous broadening and the unwanted interaction terms will be canceled in the average Hamiltonian, Generating the additional π pulse about the z axis may seem challenging because it requires a controlled detuning of the spins from the cavity such that the accumulated phase is exactly π. Experimentally, this would likely be even more difficult than turning the coupling off for half a period, and one may conclude that this scheme is harder to implement than the first one. However, it is well-known from NMR that pulses about the z axis can also be realized using a so-called composite pulse which is a suitable combination of x and y rotations [63]. Specifically, a π pulse about the z axis can be decomposed into pulses about the x and y axes as follows: Dynamical decouping sequences to cancel inhomogeneous broadening in Eq. (22). Each sequence has a total duration time T , where κsqzT 1 and κintT 1, and is repeated multiple times. (a) A simple sequence of π pulses about the x axis will generate unwantedΣ † (r) antidamping terms [cf. Eq. (23)]. Therefore, the coupling to the spins, g(t), must be switched off after every second πx pulse. (b) By adding two additional π pulses about the y and z axes, the detrimental impact of the first πx pulse can be cancelled and the coupling g(t) can be kept constant.
which does not require any detuning between the spins and the cavity. Higher order contributions to the average HamiltonianĤ will become negligible if the period T of the decoupling sequence satisfies the conditions κ sqz T 1 and κ int T 1.
C. Non-uniform single-spin couplings
Another experimentally very relevant source of imperfections are inhomogeneities in the coupling strength g between the spins and the common bosonic mode, i.e., one has to replace the Hamiltonian in Eq. (9) witĥ This breaks the permutational symmetry of the spins reflected in the collective spin operatorsŜ ± . The standard strategy to analyze this effect is based on an expansion of the mean-field equations of motion around the average couplingḡ = N j=1 g j /N [64]. Variants of this approach have been applied to study superradiance [64,65], microwave quantum memories [66], and spin squeezing [67][68][69][70][71]. Nonuniform couplings then lead to an approximate collective model with a renormalized coupling parameter and a reduced effective length of the collective spin vector. Similar results will hold in our case if the squeezing parameter is not too large, e 2r N .
To analyze the case of strong squeezing, e 2r N , more powerful theoretical methods need to be de-veloped, which is an interesting and relevant subject for further study. We expect non-uniform single-spin couplings to reduce the visibility of the even-odd effect and to influence the prethermalization physics, since the suppression of transition rates crucially relies on the indistinguishability of the spins in the ensemble.
VI. IMPURE ENGINEERED RESERVOIR
In this section, we explore a different kind of imperfection that has not been studied in previous discussions of dissipative spin squeezing: the engineered reservoir may have an imperfect purity (or equivalently, mimic thermal squeezed light rather than vacuum squeezed light). At the most basic level, this corresponds to modifying our ideal quantum master equation (2) bẏ where n th ≥ 0 parameterizes the effective temperature of the squeezed reservoir. We will discuss below in Sec. VII and in App. I how this generic model can be related to more microscopic mechanisms (including collective decay, γ coll > 0). Even though the hybrid-systems reservoir-engineering approach we focus on is not limited by losses associated with the transport and injection of an externally prepared optical squeezed state, many reservoir engineering techniques will inevitably result in a n th = 0, hence it is important to understand the impact of this unwanted heating. Note that for n th = 0, the steady-state of the spin ensemble will necessarily be impure. This can have a deleterious impact on Ramsey-type sensing experiments if one is interested in signal phases that are not infinitesimally small (as has been discussed in the context of OAT [72,73]). For e 2r N , the spin squeezing described by Eq. (29) is well-approximated by a linearized bosonic master equation (via use of the Holstein-Primakoff approximation [40]). In this limit, one thus expects that a small n th will have only a small impact on the steady-state squeezing [13], i.e., The interesting question is whether n th also has innocuous effects for the larger values of r needed to approach the Heisenberg limit. Fig. 8(a) shows that this is not the case. The linearized bosonic theory breaks down if the squeezing parameter r is too large, with dramatic consequences: even small imperfections, n th 1, cause the steady-state Wineland parameter to strongly deviate from its ideal limit ξ 2 R,HL = 2/(N + 2). Further, one finds that the steady state Wineland parameter increases with increasing r (irrespective of the parity of N ), i.e., the steady-state squeezing is a non-monotonic function of r and exhibits a minimum at an optimal value r opt (which depends on N and n th ).
We thus have an important caveat: if the engineered squeezed reservoir has a non-zero effective temperature, increasing the reservoir squeezing parameter r does not result in ever-increasing steadystate spin squeezing. We stress that this is true even when N is even. Numerical simulations indicate that the minimum Wineland parameter at r opt almost follows a Heisenberg-like scaling with N , as shown in Fig. 8(b), but with a significantly larger prefactor than the ideal result ξ 2 R,HL = 2/(N + 2). Further numerical results exploring the parameter dependence of the optimal squeezing parameter r opt are given in App. J, while App. K briefly discusses scaling obtained using a mean-field theory approach.
Interestingly, Fig. 8(c) shows that, although the ratio ξ 2 R (r opt )/ξ 2 R,HL deviates from the bosonic expectation, the purity of the steady state at the optimal squeezing parameter r opt does closely follow the corresponding relation (2n th + 1) −1 valid for bosonic dissipative squeezing.
At a heuristic level, the much stronger sensitivity of dissipative spin squeezing to n th > 0 and the similar purity (in comparison to dissipative bosonic squeezing) can be understood by the following simplified picture (see also App. J). Similar to dissipative bosonic squeezing, the dominant contributions to the mixed steady state of Eq. (29) are the dark state ofΣ, |χ 0 = |ψ dk [r] , and the "first excited" state |χ 1 ∝Σ † |ψ dk [r] . Their statistical weights are in a thermal ratio p 1 /p 0 = n th /(n th + 1), which explains the similar n th -dependence of the purity for r ≤ r opt . However, the Ŝ 2 y variance of |χ 1 differs strongly from its counterpart in a bosonic squeezed state for r r opt : in this limit, the dark state converges to the m y = 0 eigenstate ofŜ y , |ψ dk [r] → |j, 0 y [cf. Eq. (5)], with a vanishing Ŝ 2 y variance. The operatorΣ † expressed in theŜ y basis contains spin raising and lowering operators, i.e., the state |χ 1 is an equal superposition of the |j, ±1 states and has a finite Ŝ 2 y variance. This leads to an increase of the Wineland parameter with increasing r as soon as the Ŝ 2 y variance approaches to its nonzero minimum value.
Thus, the exponential increase of the Wineland parameter for r ≥ r opt has a very similar origin as the corresponding effect in the the odd-N zeron th case discussed in Sec. III. These results are yet another demonstration of the fact that dissipative spin squeezing is more complicated than dissipative bosonic squeezing, due to the finite-dimensional Hilbert space and the intrinsic nonlinearity of spin systems.
A consequence of these findings is that, for large squeezing parameters r r opt , an impurity of the engineered reservoir reduces the even-odd effect. We will discuss the consequences for the even-odd effect below in Sec. VII D, where we will show that the even-odd effect can nevertheless be observed on state-of-the-art experimental platforms.
VII. HYBRID-SYSTEMS IMPLEMENTATION USING DISSIPATIVE BOSONIC SQUEEZING
As discussed in Sec. II B, the dissipative spinsqueezing setup described by the general quantum master equation (11) can be realized using standard two-level systems (unlike the more structured fourlevel atoms in Refs. [9,25]), and without requiring the use of non-classical squeezed input light. Instead, one harnesses a standard (resonant) Tavis-Cummings coupling between a spin ensemble and a bosonic mode, along with the dissipative squeezing of this bosonic mode which is engineered by coupling the bosonic mode to a lossy auxiliary mode that is driven by imbalanced red-detuned and blue-detuned sideband drives. The second element here has been experimentally realized in a variety of systems. In this section, we provide more details on the physical implementation of our hybrid-systems approach to dissipative spin squeezing in three promising platforms: trapped ions, solid-state optomechanical devices, and superconducting circuits.
A. Trapped ions
In trapped ions, the relevant spin degree of freedom usually corresponds to two metastable internal states (spin or orbital) of each individual ion. In contrast, the bosonic "cavity" mode corresponds to a collective motional mode of the ions [74] and the coupling parameter g now characterizes the spin-phonon coupling. Recent experiments have already utilized the spin-motion coupling for over 50 ions in a 2D Penning trap [57] and 1D linear Paul trap [75]. The desired Tavis-Cummings coupling is commonly realized by applying a laser field that is resonant with the red motional sideband of the spin-level transition (see, e.g., Ref. 76). Motional dissipation is, in turn, mediated by coupling the motional mode to a dipole-allowed transition of an ion.
To realize dissipative spin squeezing with N spins, we imagine a setup that consists of N + 1 ions. N of these ions make up the spin ensemble that we wish to squeeze; the remaining additional ion serves as a "cooler" ion that is used to dissipatively squeeze the collective motional mode, as shown in Fig. 9. A squeezed bath can be engineered by applying two laser fields that are resonant with the red and blue sideband transitions of the cooler ion [12], leading to an effective Hamiltonian where G ion ) is the red (blue) sideband coupling, andσ − is the lowering operator of the coolerion transition. The squeezing strength can be controlled by the ratio of the couplings, i.e., tanh(r) = |G ion |, and the squeezed axis is determined by their relative phase. We stress that the sideband transitions are implemented by classical drives, i.e., no squeezed radiation is required for the bath engineering. Reading out the spin-squeezed state can be performed by individually measuring each ion in the Pauliσ y basis. The collective spin variance can then be obtained from the statistics of the measurement outcomes collected in multiple runs.
To assess the practical requirement of our proposal, we adopt the state-of-the-art system parameters in the dissipative motional squeezing experiment by Kienzler et al. [16]. This experiment employs the bath engineering techniques described above to prepare the motion of a single trapped ion in a −12.6 dB squeezed ground state with up to 12% infidelity. Utilizing this scheme for spin squeezing requires two additions.
First, the reservoir engineering technique has to be extended to the collective motion of an ion chain.
Since the center-of-mass mode frequency is not altered by the number of ions in the trap, and the frequencies of other motional modes are well resolved [77], the same sideband drives used in the single-ion case can be applied to engineer the same squeezed bath for the collective center-of-mass mode. Note that, for incoherent electric field noise, the centerof-mass heating rate does not depend on the number of ions in the trap [78]. Second, spin-motion coupling has to be applied between the center-of-mass mode and the N system ions with a spin-phonon coupling strength g. To implement the collective spin dissipator in Eq. (11), a sufficient but not necessary condition is that the coupling strength g is sufficiently weak to guarantee the adiabatic elimination of the motional mode, while sufficiently strong such that qubit decoherence does not deteriorate the spin squeezing, i.e., κ sqz / √ N g √ κ sqz γ φ . From the time evolution of squeezed-state pumping in Ref. 16, we estimate the experimentally realized squeezed-reservoir coupling rate to be κ sqz ≈ 0.5 kHz. Moreover, γ φ ≈ 0.1 Hz has been realized in many experiments (e.g., Ref. 79). With these realistic parameters, a coupling g ≈ 30 Hz should fit in this regime for a modest chain with N 10 ions. Since spin-motion coupling as strong as kHz has been routinely implemented in trapped-ion experiments (e.g., Ref. 80), our desired range is thus well achievable by simply using a weaker drive.
The remaining issue is whether the thermal excitation due to motional heating will mask the desired even-odd effect. Motional heating leads to collective spin excitation and relaxation processes at equal rates γ heat . As discussed in App. I, this can be mapped onto a quantum master equation of the form (29) with an effective squeezing parameterr, i.e.,Σ(r) →Σ(r) wherẽ and with an effective thermal occupation number n th given by Kienzler et al. achieved a squeezed ground state with a fidelity larger than 88 % at −12.6, dB of squeezing [16], which corresponds to n th ≈ 0.14 or, equivalently, a motional heating rate γ heat /Γ ≈ 0.017. Note that this is a worst-case estimate of the fidelity since its reported value includes measurement errors. Consequently, the true value of γ heat in the experiment is likely lower. However, even with this pessimistic values the even-odd effect can be observed in experiments, as we will show below in Sec. VII D.
B. Solid-state spins in an optomechanical crystal
In solid-state platforms, the spin ensemble in our scheme could be realized using defect centers in a semiconductor, e.g., NV-center defect spins in diamond. These spins can be implanted in a structure which in turn realizes an optomechanical crystal: a patterned photonic crystal beam with a defect that localizes both a mechanical mode and an optical mode [81]. We note that high-Q diamond optomechanical crystals have been realized experimentally [82], with a recent experiment even integrating such a system with NV-center defect spins [83]. The localized mechanical mode plays the role of the bosonic "cavity" in Eq. (11). The spins and mechanical motion exhibit an intrinsic coupling due to the strain dependence of spin-level transitions, and the coupling could be further enhanced by incorporating the high strain sensitivity of excited states through phonon-assisted Raman transitions [83][84][85].
In this kind of setup, the optomechanical coupling between the localized mechanical and optical cavity mode provides a mechanism for the dissipative squeezing of the mechanical mode. If one is in the sideband-resolved regime (where the mechanical frequency is larger than the optical cavity decay rate), then this dissipative mechanical squeezing can be realized by driving the optical cavity by two control lasers that are resonant to the red and blue motional sidebands, respectively [41]. We stress that these drives are classical, coherent state drives. Ignoring the non-linear coupling that is usually negligibly weak in most platforms, the optomechanical coupling is well approximated bŷ where G OM |, which can be tuned by varying the amplitude of the driving tones. We note that this kind of dissipative squeezing of mechanical motion via optomechanics has been realized in several experiments [15,[17][18][19]. Our protocol thus provides a means of harnessing this capability to generate spin squeezing. Finally, in solid-state settings inhomogeneous broadening of the spin ensemble is almost always an issue; this is typically mitigated by using dynamical decoupling techniques. For spin-squeezing protocols based on OAT dynamics, a very simple decoupling sequence can be used, which repeatedly applies π pulses about the x axis [27]. This very simple strategy fails in our case because it transforms theΣ decay term in Eq. (11) into aΣ † anti-damping term. However, as discussed in Sec. V B, this unwanted excitation dynamics can be canceled by two additional π pulses about the z and y axis, which makes our protocol compatible with dynamical decoupling.
C. Superconducting microwave cavities
Superconducting microwave cavities and circuit QED are another promising class of systems for implementing our ideas. Our basic building block of a bosonic mode coupled to a spin ensemble could be realized by coupling a single microwave cavity mode to either a set of superconducting qubits [86][87][88][89], or to electronic spins in substrate (e.g., Bi donors implanted in Si [90,91]). The second ingredient, a mechanism for the dissipative generation of microwave squeezing, could also be implemented in different ways. One approach is to inject squeezed microwave radiation directly into the cavity using the output of a Josephson parametric amplifier [92][93][94]. This has already been achieved experimentally in Ref. 90, in a system where a cavity has been coupled to a spin ensemble. An alternate approach which has the advantage of not being limited by insertion losses (associated with transporting a squeezed state) is to mimic the same dissipative squeezing protocols used in optomechanics to squeeze a mechanical mode. This can be accomplished by coupling three microwave modes via a Josephson ring modulator [95], which generates a three-wave mixing term (p+p † )(â+â † )(b+b † ) between the modesâ,b, andp [96]. By driving the pump modep coherently at the sum and difference frequency of theâ andb modes, ω ± , one can engineer an interaction of the form of Eq. (34), where the prefactors G (±) OM depend on the strength of the drives at ω ± , respectively. Adiabatic elimination of the strongly-dampedb mode generates an effective squeezed bath for theâ mode as shown in Eq. (9). A recent experiment implementing this approach has demonstrated up to −8 dB of intracavity squeezing of theâ mode [20].
D. Experimental viability of the even-odd effect
Given the results of the Sec. VI, one may worry that a finite effective temperature of the engineered (a) Ratio between Ŝ 2 y for N = 8 vs. 9 as a function of the motional heating rate γ heat (red triangles). The dotted red line is a guide to the eye to highlight a ratio of unity, i.e., no even-odd effect. For each value of N and γ heat , we optimize the squeezing parameter over the range 0 ≤ e 2r ≤ 12 dB to minimize the variance Ŝ 2 y (blue data points). (b) Ratio of Ŝ 2 y for N vs. N + 1 spins as a function of N for an experimentally realistic parameter of γ heat /Γ = 0.017 (red triangles). Again, the dotted red line is a guide to the eye to highlight a ratio of unity, i.e., no even-odd effect. For each N , we optimize the squeezing parameter over the same range 0 ≤ e 2r ≤ 12 dB as in (a) to minimize the variance Ŝ 2 y (blue stars).
reservoir will substantially decrease spin squeezing and mask the even-odd effect introduced in Sec. III.
Here, we show that this is not always the case and, in particular, that the even-odd effect can be observed in experimentally accessible parameter regimes. For concreteness, we focus on the trapped-ion platform introduced in Sec. VII A, which is most mature. With a first experimental observation of the even-odd effect in mind, we consider a modest number of spins, N 10. In this platform, motional heating due to classical trap noise is the dominant source of imperfections of the engineered squeezed reservoir. As discussed in Sec. VII A and App. I, its impact can be modeled using Eq. (29) with an effective squeezing parameterr and an effective thermal occupation number n th given by Eqs. (32) and (33), respectively.
Compared to the bare squeezing parameter r (which is related to the amplitude of the red and blue sideband drives), the effective squeezing parameter strengthr is reduced by the motional heating rate γ heat . The effective thermal occupation number n th grows with increasing r, i.e., a large squeezing parameter reduces the purity of the squeezed state. For a given number N of ions, the variance Ŝ 2 y will thus take a minimum value at an optimal squeezing parameter r opt , whose value depends on N and γ heat /Γ, as shown in Fig. 10(a). Comparing these minimum variances for N = 8 vs. 9 ions, we find an even-odd difference of more than 20 % over a wide range of motional heating rates γ heat . The even-odd difference decreases with effective temperature n th and will disappear if the impurity of the squeezed reservoir is sufficiently large. Note that even for the upper bound of the motional heating rate derived in Sec. VII A, γ heat /Γ = 0.017, the even-odd difference is clearly visible.
An alternative way of probing the even-odd effect is shown in Fig. 10(b). There, the motional heating rate γ heat is kept fixed but the number of ions in the trap is varied. Again, we optimize r for each data point individually to minimize the variance Ŝ 2 y . The ratio between these variances shows pronounced oscillations for N ≤ 10 ions. Note that we restrict the optimal squeezing parameter e 2ropt in both scenarios to be smaller than the experimentally achievable −12.6 dB. Thus, our results suggest that the even-odd effect is realizable on state-of-the-art trapped ion platforms.
Instead of varying r for each N individually to minimize the variances of even and odd N independently, one could also use the same squeezing parameter r for a pair of N and N + 1 spins and maximize the ratio of their variances Ŝ 2 y (r) N / Ŝ 2 y (r) N +1 . In this case, the even-odd effect would be even more pronounced than the data shown in Fig. 10.
VIII. CONNECTION TO PREVIOUS WORKS
Here we review previous works on dissipative spin squeezing [9,[21][22][23][24][25], summarize their results and point out the differences to this work.
Agarwal and Puri discussed the idealized spinonly quantum master equation (2) with an additional collective decay term D Ŝ − ρ and a coherent drive ΩŜ + +Ω * Ŝ − [21,22]. They derived explicit expressions for the steady state and pointed out that the steady state of Eq. (2) is pure for even N (wherê Σ has a zero eigenvalue) and mixed for odd N (wherê Σ has only nonzero eigenvalues) [21]. Moreover, they discussed the pairwise excitation structure of the even-N steady state illustrated in Fig. 2(a) and showed numerical results for the odd-N population distribution in the regime e 2r N [21]. In a follow-up article [23], they calculated the Wineland spin-squeezing parameter ξ 2 R of a state of the form |ψ ∝ exp θŜ z exp −iπŜ y /2 |j, m , which contains the even-N steady state (5) as a special case, but not the odd-N steady state. They also discussed squeezing in the presence of a coherent drive [22].
Unlike our work, Agarwal and Puri did not evaluate the spin-squeezing properties of the odd-N steady state. Consequently, they did not find the dramatic difference in spin squeezing between the undriven even-N and odd-N steady states for e 2r N , which is one of the central results of this work. Note that the spin-squeezing parameter r required to see an even-odd effect will decrease if the number N of spins is lowered, which makes the regime e 2r N attainable on state-of-the-art experimental platforms, as discussed in Sec. VII. Therefore, another new and experimentally relevant result of our work is that the even-odd effect is not a mere mathematical complication, but a real and testable physical effect.
Agarwal and Puri suggested to generate spin squeezing by illuminating the spins with squeezed radiation, which is experimentally very challenging. In a later work, Kuzmich et al. proposed a related alternate scheme to generate spin squeezing using Vtype atoms illuminated by squeezed light [24]. They did not discuss dissipative spin squeezing and, consequently, did not comment on the even-odd effect at all. In contrast to the proposals by Agarwal et al. and Kuzmich et al., our work does not require any injection of nonclassical light, which lowers the experimental challenges significantly.
Dalla Torre et al. proposed another method to implement dissipative spin-squeezing dynamics in a specific multilevel atomic system using Raman transitions [25]. They use a pure dark state of the type of Eq. (6) to analyze the ideal situation when there is only collective loss of the form of Eq. (3). In addition, they discuss the impact of dissipation due to single-atom Raman scattering into free space. They do neither comment on an even-odd effect nor mention that their pure-state analysis is strictly speaking only valid in the case of even N .
Our work proves that, while this mathematical treatment is indeed admissible if squeezing is not too strong, e 2r N , substantial changes to the spinsqueezing physics will show up at e 2r N , which have not been discussed in the literature before. Regarding experimental versatility, Kuzmich et al. and Dalla Torre et al. discuss very specific implementations, which are not applicable to a generic ensemble of two-level systems. In contrast, we propose a more generic mechanism that has not been discussed in the literature before. Importantly, our approach is compatible with a wide range of systems including solid-state implementations (see Sec. VII) and is perhaps the most flexible and experimentally viable implementation. This versatility allowed us to discover the prethermalization physics discussed in Sec. V A, which is not present in the scheme of Ref. 25 because of their very specific decay mechanism that couples collective and local dissipation.
Finally, Borregaard et al. considered Λ-type atoms driven by multiple laser drives and identified a dissipative spin-squeezing scheme as its resonant limit [9]. They use this observation to interpret their results, but they do not discuss the dissipative spinsqueezing scheme in detail nor explore any of its consequences revealed in this work, like the even-odd physics and prethermalization.
Finally, we note that in contrast to our work, the previous works on dissipative spin squeezing reviewed above did not discuss or analyze the consequences of an imperfect engineered reservoir (see Sec. VI).
IX. CONCLUSIONS
In this work, we have revisited the reservoirengineering approach to preparing and stabilizing spin-squeezed states. We analyzed in detail a particular implementation strategy that had not previously been studied, but that is compatible with a number of experimental platforms: employ a hybrid-systems approach where one first uses bosonic reservoir-engineering techniques to stabilize a bosonic squeezed state, and then uses this state (via a standard Tavis-Cummings-type coupling) to dissipatively squeeze a spin ensemble. We also discussed how this approach compared favourably to the standard one-axis-twist method for spin squeezing in the presence of single-spin relaxation.
Our work also addressed fundamental aspects of dissipative spin squeezing, with a focus on two general but surprising phenomena. The first was an extreme, macroscopic sensitivity of the steady state to the parity of the number N of spins in the en-semble. We analyzed both how this effect could be avoided (if the goal was to generate spin squeezing without any parity sensitivity), and how it might be harnessed for a powerful new sensing modality. The second general effect we studied was the emergence of a surprisingly long slow timescale and "prethermalization" behavior when weak single-spin dephasing is added to our model. While the steady state in this regime exhibits at best limited squeezing, the intermediate time quasi-steady state can be highly squeezed. Moreover, the reduction of steady-state spin squeezing can be almost completely suppressed by deliberately introducing a small amount of singlespin relaxation.
Finally, we investigated the impact of an engineered reservoir stabilizing an impure steady state. We discovered a strong sensitivity of the Wineland parameter to impurity if the squeezing parameter r is large.
We hope our work will lay the groundwork for near-term experimental implementations of reservoir-engineered spin squeezing in a variety of systems. In future theoretical work, it will be interesting to explore extensions of the models analyzed here. For example, it is well known that collective Hamiltonian interactions that are not truly infinite range can still generate large amounts of spin squeezing [97,98]. Is the same true with dissipative spin-spin interactions, and if so, are the requirements more or less forgiving than in the coherent case? It would also be interesting to study in more detail the effects of disorder, e.g., due to inhomogeneous broadening, both on spin squeezing and on the parity-sensing scheme proposed here.
ACKNOWLEDGMENTS
This work was primary supported by the DARPA DRINQS program (agreement D18AC00014). We also acknowledge partial support by the University of Chicago Materials Research Science and Engineering Center, which is funded by the National Science In this Appendix, we outline the derivation of the effective quantum master equation (11) of the main text. Our starting point is Eq. (9) describing a collection of spins interacting with a squeezed bosonic mode. For the moment, we ignore the terms in Eq. (9) describing local dissipation of the spins, Assuming that the cavity evolves on a much shorter timescale than the spins, we adiabatically eliminate the cavity by a projection operator technique [99] similar to the calculation outlined in Ref. 100. To this end, we split the quantum master equation into two superoperators, where L intρ is considered to be constant on the timescale defined by L cavρ . Using this approximation, we can formally solve Eq. (A3), Performing a Born approximation, we decompose the state asρ(t) ≈ρ sp (t) ⊗ρ ss cav , whereρ sp (t) is the reduced density matrix of the spin system andρ ss cav is the steady state of L cav . The equation of motion of the reduced spin density matrix is Inserting the explicit form of L int , we find that the integral on the right-hand side of Eq. (A5) depends on the cavity correlation functions Tr cav â † e Lcavtâρss cav = Tr cav âe Lcavtρss These correlation functions decay fast compared to the timescale on whichρ sp evolves, therefore, we can perform the Markov approximationρ sp (t ) ≈ρ sp (t) and rewrite Eq. (A5) as follows: Taking into account the remaining terms in Eq. (9) describing single-spin dissipation, we recover Eq. (11) of the main text.
Appendix B: Mean-field theory equations of motion
In this section, we provide the set of nonlinear equations of motion for symmetrized products of spin operators, for the effective spin-only model considered in the main text, namely Eq. (11). While such a system of equations is not closed, we neglect third-order cumulants (equivalently performing a 2 nd -order cumulant expansion) [101,102], which lets us approximate the third-order expectation values of various operators as We stress that the above approximation is applied to expectation values of symmetrized operators, defined according to the following convention: Given the initial state with spins completely polarized along the −z direction (i.e., Ŝ z = −N/2), the evolution is governed by the equations where Ĉ ZZ = Ŝ 2 z − Ŝ z Ŝ z . We stress that if we assume that Eq. (11) is a result of coupling the spin system to a cavity interacting with an engineered squeezed reservoir with photon loss κ int , then we have and as discussed in the main text and shown in detail in App. A.
Appendix C: Cooperativity scaling of the ξ 2
R parameter
In this Appendix, we provide a derivation of the cooperativity scaling of the Wineland parameter ξ 2 R . We concentrate our analysis on the weak dephasing limit, and start with the case where γ φ = 0 and where only the local decay γ rel as well as the collective cavity-induced decay γ coll are present. A scenario where local spin dephasing is dominant can lead to substantially altered behavior of the system, and is the subject of Sec. V A and App. E.
Analytical derivation
We begin by linearizing the mean-field-theory equations of motion shown in App. B by focusing on the limit where Ŝ z stays fixed at −N/2. This approximation closely reflects the true system dynamics when the spin number N is large and when the single cooperativities η φ or η rel are not much larger than unity, resulting in effective spin squeezing that is far from the Heisenberg limit. Hence, taking Ŝ z = −N/2 (i.e., spins keeping their polarization throughout the evolution and in the steady state), the Wineland parameter takes a simple form, which we can write using the results in App. B as Note that ξ 2 R gets smaller as r increases and hence, in what follows, we will take the limit r → ∞. It is worth pointing out, however, that choosing a finite r which satisfies exp (−2r) 1/ √ C rel is sufficient to reproduce the scaling of ξ 2 R derived below. In the large-r limit, we find Next, we use Eqs. (12) and (13) of the main text to rewrite the above expression as We consider a limit where N → ∞, while G = √ N g stays fixed. In such a case, the last term of the numerator can be dropped. Here it is crucial to point out that in an experimental setting, one will typically not have much control over κ int and γ rel , while κ sqz can be tuned at will through appropriate reservoir engineering (see Sec. VII). Hence, it is important to understand what value of κ sqz should be chosen to maximize the amount squeezing that this protocol can achieve. At first glance, one might think that choosing κ sqz as large as possible (i.e., κ sqz → ∞) is ideal as that maximizes the amount of bosonic squeezing that the spin-coupled cavity experiences. From Eq. (12), however, we see that such a choice will actually limit the value of Γ, which directly impacts the strength of squeezed-vacuum reservoir that the spins see [see Eq. (11) in the main text], resulting in the squeezing performance being strongly limited by the value of γ rel . Hence, as we shall see shortly, the right thing to do is to still choose κ sqz κ int , γ rel , but yet not too large, so that the Γ-controlled process is dominant over the local spin decay γ rel . To see this explicitly, we minimize Eq. (C4) with respect to κ sqz . Assuming N 1, this leads to where, in the second line, we used Eq. (10) of the main text to express the result in terms of the collective cooperativity C rel and then expanded in the limit of large C rel . Our above expression shows the C −1/2 rel cooperativity scaling for the dissipative protocol, which outperforms the C −1/3 rel behavior of the OAT method [35], in the case where spin decay is the dominant local noise process. The optimal value of κ sqz that results in Eq. (C5) reads which confirms the need for κ sqz κ int (given large C rel ), while also showing that it should not be infinitely large. Finally, we stress that ξ 2 R is ultimately limited by κ int /(κ int + κ sqz ), hence it is crucial that an appropriate κ sqz can be realized in an experimental setting.
While the above result has been calculated in the limit where γ φ = 0, a similar expression is valid when some local dephasing is present (i.e., γ φ = 0). In such a case, one can simply assume γ rel → γ rel + 2γ φ in Eq. (C4). As discussed in more detail in Sec. V A of the main text, however, this is only true when γ φ is not too large, namely when γ φ N γ rel . Otherwise, a local dephasing process can have a significant impact on the evolution and therefore dramatically limit the steady-state performance of the protocol.
Mean-field theory simulations
In this section, we present mean-field-theory simulations of the dissipative protocol obtained using R as a function of collective cooperativity C rel . The blue curve corresponds to ξ 2 R calculated by evolving the full (nonlinear) mean-field equations of motion for the dissipative system (see App. B). Here κint = 500g, γ rel = 0.04g, giving η rel = 0.2. The number N of spins is changed in order to vary C rel . At each blue point, both r and κsqz are optimized in order to minimize ξ 2 R . The orange dashed curve shows the corresponding fit (calculated over the three data points with the largest C rel ). The black dashed line describes the optimized squeezing of the engineered bosonic reservoir. The solid black line shows an ideal Heisenberg scaling 2/(N + 2). Finally, the black dotted curve shows the OAT scaling as calculated in [35].
the full (nonlinear) equations shown in App. B. We consider the case where local spin decay dominates over local spin dephasing, and in particular work in the limit of γ φ = 0.
The plot in Fig. 11, shows the scaling of the Wineland parameter as a function of the collective cooperativity C rel . The parameters are κ int = 500g, γ rel = 0.04g, giving η rel = 0.2, while the number N of spins is varied in order to modify C rel . At each blue point, both r and κ sqz are optimized in order to minimize ξ 2 R . The orange curve shows the corresponding fit, which is calculated using the three data points with largest C rel . We see good agreement with the cooperativity scaling discussed in the main text and derived in detail in the section above (where we have linearized the equations of motion). For comparison, the black dashed line describes the optimized squeezing of the engineered bosonic reservoir. The black solid line shows an ideal Heisenberg scaling 2/(N + 2). We also plot the dotted black curve, which corresponds to the squeezing one would get from the OAT protocol in the limit where γ rel dominates over γ φ (in the large C rel limit) -see Ref. 35. The simulations confirm that the dissipative protocol can indeed outperform the OAT approach.
Appendix D: Even-odd effect
In this Appendix, we briefly review previous results on the dissipative steady state of Eq. (2) in the main text and we derive Eq. (16) of the main text. We then comment on variance detection measurements required to use the even-odd effect as a sensor, and we discuss the impact of local dissipation.
Properties of the steady state
Agarwal and Puri derived that the steady state of Eq. (2) isρ if the Hermitian operatorΣ †Σ is invertible [22].
If N is even,Σ(r) has a zero eigenvalue in each subspace of angular momentum j and the associated eigenstates are the dark states |ψ dk [j; r] ∝ e θ(r)Ŝz |j, 0 y , 2. Using the even-odd effect for sensing As described in Sec. III D of the main text, the sensitivity of the steady state on the parity of the number N of spins can be used for sensing. Experimentally, sudden changes in the parity of N can be induced by various mechanisms. Trapped atoms can be physically lost from the trap by collisions with background gas, internal collisions, and photon-assisted processes [104]. If the spin-1/2 degree of freedom is a subspace of an atomic multi-level structure, undesired internal transitions can occur, which take the atom out of the spin-1/2 subspace and effectively remove it from the collective dynamics even though it may still be trapped [25]. Moreover, one could devise a system where the coupling strength of a single spin to the cavity and, thus, to the collective spin depends on an external parameter. A change of this single-spin coupling strength modifies the number of collective spins, which is collectively amplified and yields a large change of the steady state.
Note that such effective atom loss events do not change the collective expectation value Ŝ y = 0 of the distribution. However, the statistics of the fluctuations Ŝ 2 y depends on the parity, as shown in Fig. 12. The parity of N can thus be inferred by imposing a threshold condition on the variance Ŝ 2 y measured using spin-noise spectroscopy [54][55][56][57][58][59][60].
Impact of local dissipation
So far, our analysis of even-odd effects in the steady state has focused on the idealized case without any single-spin dissipation: γ rel = γ φ = γ coll = 0. We found that the Wineland parameters for even and odd N differ strongly in the regime e 2r N , as shown in Fig. 3 of the main text. Figure 13 shows that if local dissipation is taken into account, spin squeezing is reduced but the ratio between the Wineland parameters for even and odd N remains large. Moreover, for a fixed value of the squeezing parameter r, the ratio of the Wineland parameters in the presence of local dephasing can even be larger than the corresponding ratio obtained for γ φ = 0. For fixed local dissipation rates, the ratio is largest around the onset of the even-odd effect. At this optimum squeezing parameter r max , effective single-spin cooperativities much larger than unity, Γ/γ φ 1 or Γ/γ rel 1, are required to observe a ratio of the Wineland parameters greater than two.
Appendix E: Liouvillian perturbation theory of the slow timescale In this Appendix, we use Liouvillian perturbation theory [61] to analyze the emergence of the long relaxation timescale in the presence of local dephasing, which has been discussed in Sec. V A of the main text. We also provide a simple physical argument to understand this effect.
1. Hilbert space of N spin-1/2 systems and permutational invariance Addition of angular momenta of N spin-1/2 systems gives rise to N/2 + 1 subspaces of total angular momentum j, where j takes values between j max = N/2 and j min = 0 (1/2) if N is even (odd) [105]. For N > 2, all but the maximum-angularmomentum subspace are degenerate since there are multiple ways to combine N spin-1/2 systems to a total angular momentum j < N/2 [33] (for an illustration, see, e.g., Ref. 106). If local dissipative processes act identically on each spin-1/2 system, the equations of motion are invariant under permutation of the spins [34]. Consequently, if the system is initialized in a permutationally invariant state, e.g., any state in the subspace j = j max , the collective and local dissipative processes will preserve the permutational symmetry. Exploiting this symmetry, one can derive an effective quantum master equation which requires significantly less degrees of freedom to describe the system [34] and gives rise to efficient numerical simulation of large spin ensembles [107].
Analysis of the slow timescale
Our starting point, the quantum master equation (11) of the main text, belongs to the class of permutationally invariant systems described above. In the following, we focus on the case γ coll = γ rel = 0. Introducing the dimensionless time τ = Γt, the equation can be rewritten in the form dρ/dτ = L 0ρ + εL 1ρ , where we introduced the dimensionless superoperators and the dimensionless perturbation strength ε = 2γ φ /Γ. In the absence of local dephasing, ε = 0, the superoperator L 0 has N/2 + 1 different steady statesρ (j) 0 , each of them living in a different subspace of collective angular momentum j. Weak local dephasing, γ φ Γ, enables incoherent transitions between adjacent angular-momentum subspaces [34], which can be visualized as trajectory in a triangular (j, m) state space [108]. This perturbation lifts the degeneracy of the steady states and opens a new dissipative gap that determines the relaxation timescale towards the new, unique steady state.
The first-order corrections to the vanishing eigenvalues of L 0 are given by the eigenvalues of the tridiagonal matrix containing the transition rates j → j between collective angular momentum subspaces. Here,1 (j) is the identity operator in the angular-momentum subspace j. For even N , the transition rates are shown in Fig. 14(a). They depend on the structure of the dark state (6) given in the main text, where c (j) m are the expansion coefficients of the dark state and Γ (5,6) j;m,m are the transition rates derived in Ref. 34 using the notation introduced in Ref. 107. Note that, in our case, L 1 is dimensionless such that the transition rates (E4) and (E5) are dimensionless, too. The asymptotic decay rate, i.e., the absolute value of the smallest gap in the spectrum of M , is shown in Fig. 14(b).
For r = 0, the unperturbed steady states are the ground states of each angular-momentum subspace, ρ For r = 0, transition rates Γ j→j−1 are nonzero and dominate over the rate Γ N/2−1→N/2 if the condition r > 1/ √ N holds. As a consequence, an initial state in the maximum-angular-momentum subspace j = N/2 will undergo a directed hopping process towards lower angular momentum subspaces until it reaches a subspace j 0 where "downward" and "upward" rates are balanced, Γ j0→j0−1 ≈ Γ j0−1→j0 . Note that the downward rates Γ j→j−1 are almost constant as a function of j whereas the upward rates Γ j→j+1 depend strongly on j, as shown in Fig. 14(a). The asymptotic decay rate towards the steady state is proportional to 1/N if the transition rates in the vicinity of the equilibrium point j 0 scale proportional to 1/N . Inspection of the rates Γ (5,6) j;m,m listed in Ref. 107 shows that this is the case if j N/2 and m 0. For a given squeezing parameter r, these conditions can be fulfilled if N is sufficiently large, as shown in Fig. 14(b). Numerically, we find an exponent a ≈ 5, see inset of Fig. 14(b).
In the limit r → ∞, the asymptotic decay rate converges to the constant value 1/2.
Physical argument for the slow timescale
The existence of a bottleneck relaxation rate causing a 1/N scaling of the asymptotic decay rate for local dephasing can be understood by an intuitive argument. To explain it, we focus on the transition rate Γ N/2−1,−N/2+1→N/2,−N/2+1 , which is the bottleneck determining the asymptotic decay rate in the limit r < 1/ √ N . The states that are involved in this transition can be parametrized as [109,110] where p ∈ {0, . . . , N − 1}. Here, |j denotes the Nparticle state where the jth spin is in the excited state and all others are in the ground state. The p = 0 state has total angular momentum j = N/2, i.e., we can identify it with the state in the maximum-angular-momentum subspace. In contrast, the N − 1 states with p > 0 have total angular momentum j = N/2 − 1. Therefore, the index p > 0 allows us to label the N − 1 degenerate states in the j = N/2 − 1 subspace, |p ≡ |N/2 − 1, −N/2 + 1, p for p > 0 .
Local dephasing of spin n changes one sign in the superposition (E7), 1 2σ and thus creates an overlap between the orthogonal states |0 and |p > 0 that is proportional to 1/N , For identical dephasing processes on all N spins and for a collective initial state, i.e., a uniform statistical mixture of all N − 1 states |N/2 − 1, −N/2 + 1, p , the total upward transition rate between the two collective angular momentum subspaces is which is the bottleneck of the relaxation process and features the 1/N scaling with system size. Note that the corresponding downward rate is of the order of unity because we have to sum over all N − 1 possible target states |p > 0 , too, Also note that local relaxation does not lead to a similar emergence of the slow timescale because the overlap corresponding to Eq. (E11) will only be proportional to 1/ √ N and is thus canceled by the summation performed in Eq. (E12).
Appendix F: Optimal parameters in master equations
In this section, we show how the optimal protocol parameters vary as a function of increasing system size N in simulations from Fig. 5 of the main text. In the case of the dissipative protocols, the optimization included varying both r as well as κ sqz , whereas in the case of OAT, the spin-cavity detuning ∆ c (see App. G) is was varied. The results are shown in Fig. 15.
Appendix G: Effective One-Axis-Twist quantum master equation In this Appendix, we present the effective model that we consider when discussing the OAT protocol both in the main text and in App. C 2. In particular, following [27,35], we envision an ensemble of spins dispersively coupled to a bosonic cavity. After adiabatically eliminating the cavity, the spin-only quantum master equation can be approximated by [35] with and with ∆ c representing the cavity-spin detuning, g the cavity-spin coupling strength, κ int the decay rate of the cavity, and γ rel the local spin decay. We point out that we assume in the simulations that ∆ c is a tunable parameter, over which we optimize in order to maximize the amount of spin squeezing that the protocol can achieve.
Appendix H: Scaling of the Wineland parameter ξ 2 R in the limit η rel → ∞ When analyzing the performance of the dissipative spin-squeezing protocol in the main text, as one means of implementation, we envisioned engineering the required dissipator by coupling a spin ensemble to a lossy cavity that in turn interacts with an appropriately engineered squeezed bath. Furthermore, in our cooperativity scaling analysis (see Sec. IV and App. C) we investigated the limit of weak single spin cooperativity η rel ≤ 1. It is also interesting to consider a different asymptotic regime, where the internal cavity loss κ int is negligible, giving an extremely large η rel . We focus on the specific case where κ int = 0, and the only undesired dynamics is due to single-spin relaxation at a rate γ rel . Such a situation could be realized without any cavity, by directly irradiating an ensemble of two-level atoms with squeezed light. While this situation was analyzed in Refs. [22,24] the impacts of single spin relaxation were not studied.
The master equation in our chosen limit is thuṡ The key dimensionless parameter that describes the competition of the desired collective dissipative dynamics and the unwanted relaxation is Once again concentrating our attention on the large-N limit and fixing Ŝ z = −N/2, we can approximate the Wineland parameter using the mean-field equations of App. B as In the above expression we have already taken the limit r → ∞, which minimizes ξ 2 R . As one would expect, achievable squeezing increases asη gets larger, but more importantly we have that ξ 2 R ∝ 1/N . We can also define a quantity analogous to a collective cooperativity in this simplified system, which then lets us write assumingη ≤ 1 andC 1. In this Appendix, we show that the generalized model (29), describing an engineered reservoir that stabilizes an impure steady state, can capture the impact of collective excitation and relaxation if the squeezing parameter r and the effective thermal occupation number n th are adjusted properly. To proof this, we start with the fully general quantum master equationρ = ΓD Σ ρ + Γ D Σ † ρ which can model, e.g., interaction with an impure squeezing reservoir and a finite-temperature collective-decay reservoir if the ratios Γ/Γ and γ/γ are chosen properly. Equation (I1) is equivalent to a quantum master equation of the form given in Eq. (29),ρ =ΓD Σ ρ +Γ D Σ † ρ , where we defined a new squeezing parameterr and new decay ratesΓ,Γ as follows: 1 tanh(2r) = 1 tanh(2r) Γ = (Γ + Γ ) cosh(2r) + γ + γ 2 cosh(2r) Condition (I4) can be satisfied for arbitrary nonnegative rates Γ, Γ , γ, and γ . Collective excitation or decay, γ = 0 or γ = 0, respectively, will decrease the squeezing parameter, i.e., we always haver ≤ r. Sufficient (but not necessary) conditions to obtain nonnegative decay ratesΓ andΓ are Note that these conditions are satisfied if γ = γ , which is the case for the trapped-ion implementation dicussed in Sec. VII. There, we have Γ = 0 and γ = γ = γ heat Γ (see Sec. VII). Expanding the general results (I4) to (I6) in the small parameter γ heat /Γ, we find This corresponds to an impure squeezed reservoir with reduced squeezing parameterr < r, effective thermal occupation number n th = cosh(2r)γ heat /Γ, and unchanged decay rate Γ, as shown in Eqs. (32) and (33).
Similarly, for a perfect squeezing reservoir and small zero-temperature collective decay, i.e., Γ = i.e., the presence of collective decay can be understood in terms of an impure squeezing reservoir with decreased squeezing parameterr < r, effective temperature n th , and an enhanced decay rate Γ + γ coll . Ratio between the minimum steady-state Wineland parameter obtained by solving the mean-field equations of motion for the system governed by Eq. (K1), and the ideal Heisenberg-limited value ξ 2 R,HL = 2/(N + 2). Each dot corresponds a solution with optimized squeezing strength r, while the dashed curves are fits to a horizontal line (taken over the largest four values of N for each n th ).
of the main text, which we reproduce here for completeness: ρ = Γ(n th + 1)D Σ ρ + Γn th D Σ † ρ. (K1) Following App. B, we can once again write the corresponding mean-field equations setting third cumulants to zero, which can be readily solved numerically. Of particular interest is the scaling of the Wineland parameter ξ 2 R , with the spin number N . The results are presented in Fig. 17, where we show the ratio between the minimum steady-state Wineland parameter obtained by solving the meanfield equations of motion, and the ideal Heisenberglimited value ξ 2 R,HL = 2/(N + 2) (which full theory predicts at n th = 0). Each dot corresponds a solution with optimized squeezing strength r, while the dashed curves are fits to a horizontal line taken over the largest four values of N for each n th . From the plot it is clear that the mean-field theory predicts ∼ 1/N scaling at large N which is weakly skewed at smallest values of N that we consider. These results are consistent with our full master equation simulations presented in Sec. VI of the main text. Similarly to a simple bosonic theory prediction, we also find that the minimum steady state value of Wineland parameter scales linearly with growing n th . In particular, to a good approximation it satisfies the phenomenological equation ξ 2 R ≈ 2.8(1 + 2n th ) exp(−2r opt ), (K2) with r opt separately optimized for each n th . Finally, we point out that, as can be seen from the blue curve of Fig. 17, the mean-field theory solution does not correctly predict the prefactor of the 1/N in the case of n th = 0 (i.e., where the corresponding ratio ξ 2 R (r opt )/ξ 2 R,HL should be equal to 1). This is a somewhat expected behavior, as in parameter regimes where the ξ 2 R may be either at, or near its Heisenberg limited value, the state of the spin ensemble is not Gaussian, which may substantially lower the ability of mean-field theory to quantitatively describe its behavior. | 22,990.8 | 2021-04-21T00:00:00.000 | [
"Physics"
] |
Continuous and time-discrete non-Markovian system-reservoir interactions: Dissipative coherent quantum feedback in Liouville space
Based on tensor network realizations of path integrals reducing exponential memory scaling to polynomial efficiency and a Liouville space implementation of a time-discrete quantum memory, we investigate a quantum system simultaneously exposed to two structured reservoirs. For this purpose, we employ a numerically exact quasi-2D tensor network combining both diagonal and off-diagonal system-reservoir interactions with a twofold memory for continuous and discrete retardation effects. As a possible example, we study the non-Markovian dynamical interplay between discrete photonic feedback and structured acoustic phonon modes, resulting in emerging inter-reservoir correlations and long-living population trapping within an initially-excited two-level system.
In this work, we present a numerically exact tensor network-based approach allowing for the first time to describe two non-Markovian processes simultaneously, namely continuous and discrete retardation effects where interactions with both diagonal and off-diagonal system couplings are taken into account, i.e., couplings without or with energy exchange between system and reservoir. Recently established matrix product state (MPS) techniques to implement a time-discrete memory were aimed at quantum systems where decoherence and dephasing effects are not a key factor, and thus were based on a wave function ansatz to solve the quantum stochastic Schrödinger equation [27][28][29]. We extend this approach to a time bin-based density matrix description in Liouville space to include Markovian and non-Markovian decoherence effects. In a second step, it is combined with a tensor network-based real-time path integral method to describe interactions with a continuous structured reservoir [30][31][32][33][34], resulting in a quasi-2D tensor network formalism [35][36][37][38]. This architecture enables non-Markovian simulations of quantum systems coupled to two structured reservoirs, see Fig. 1, maintaining the relevant entanglement information and capturing both diagonal and off-diagonal system-reservoir interactions on equal footing. Possible applications include setups of waveguide-QED with dephasing [25,[39][40][41], e.g. realized by additional decay channels, or multiple spatially separated solid-state quantum emitters initially prepared in a dark state and interacting with their environment [42]. Here, we specifically consider a twolevel quantum system coupled to a structured reservoir of independent oscillators and subject to time-discrete coherent quantum feedback, extending the paradigm of the Spin-Boson model to the feedback realm [43]. We demonstrate that non-Markovian interplay between relaxation and decoherence processes results in a dynamical protection against destructive interference and thereby enables population trapping. This expands upon the widely-discussed localized phase stabilization in the spinboson model [31,43] from an incoherent feedback-induced perspective, replacing the coherent driving with another structured reservoir.
The paper is organized as follows: In Sec. I, a recently established MPS-based path integral implementation for continuous reservoirs is discussed. Afterwards, we introduce an MPS implementation of a time-discrete quantum memory in Liouville space in Sec. II. In Sec. III, we combine the two algorithms to form a quasi-2D tensor network, enabling numerically exact calculations of two non-Markovian system-reservoir interactions, before demonstrating its capabilities in Sec. IV, where we find a dynamical protection of coherence in the presence of two non-Markovian reservoirs. Lastly, we summarize our findings in Sec. V.
I. PATH INTEGRAL FORMULATION FOR CONTINUOUS RESERVOIRS
We start with the implementation and evolution of a system subjected to a continuous harmonic reservoir. For a numerically exact description, our theoretical approach is based on a real-time path integral formulation [9][10][11][12][13][14][15][16][17]. In recent breakthroughs, path integrals have been implemented in a tensor network approach based on MPS [31,32], allowing to solve non-Markovian dynamics and providing an efficient representation using highdimensional tensors with restricted correlations [31]. In the following, we briefly recapitulate the algorithm introduced by Strathearn et al. [30,31,34], which is employed as a part of our solution to multiple non-Markovian system-reservoir interactions.
Our goal is to employ path integrals for a numerically exact solution of the von-Neumann equation given a Hamiltonian H(t) describing a time-dependent systemreservoir interaction [44,45], with ρ(t) denoting the density matrix and L(t) the Liouvillian superoperator. As a main challenge, the evaluation of path integrals becomes increasingly expensive over time, since the history of all preceding paths at times 0, . . . , t n−1 must be taken into account for the calculation of the current time step t n . However, in case the system-reservoir correlations are finite in time, the augmented density tensor scheme can be introduced for improved numerical accessibility [16,17]. Exploiting the finite reservoir memory length, only the last n c time steps are taken into account for the calculation of the current path. This treatment is known as the finite memory approximation and results in the augmented density tensor representation as a solution to the system part of Eq. (1) with traced out reservoir contributions, which reads at and constitutes a discrete path integral formulation where indices i (′) n contain the left (right) configuration of the system at time t n = n∆t. The field transformation matrix M inin−1 e.g. accounts for the action of an external driving field Ω 0 . For the case Ω 0 = 0 considered below, it takes the simple form M inin−1 = 1δ in,in−1 . The influence functional is given by with and φ(τ − τ ′ ) the reservoir autocorrelation function [10,11,43]. Using an improved finite memory approximation, for n − m ≡ n c all former paths up to t nc = n c ∆t are additionally incorporated in the integration, i.e. η nc := η n−m + n−nc−1 k=1 η n−k [30]. Under this approximation, it is possible to restate the augmented density tensor and its time evolution efficiently as a tensor network [31]: First, Eq. (2) is mapped to a vector ρ jn in Liouville space, with I(j n , j m ) := jn−1M jnjn−1 exp (S jnjm ). Here, left and right system indices i k , i ′ k have been combined to a single index j k for each time step, resulting in Liouville space representationsM jnjn−1 andS jnjm of the field transformation matrix and the influence functional, respectively. Afterwards, the augmented density tensor is rewritten as an MPS, storing the present and up to n c − 1 past states in individual tensors with the oldest state located at the left end of the MPS. In this representation, tensor compression by consecutive applications of the singular value decomposition [46] reduces the memory requirements to polynomial rather than exponential scaling with respect to n c [31].
The time evolution is carried out by a network of matrix product operators (MPOs), shown schematically in Fig. 2(a) (dark grey shapes). During the first time step, the initial system state ρ j0 (0) (red shape) is contracted with the first MPO in the network [dashed frame in Fig. 2(a)]. As a result, the system state is updated and the preceding path is stored to its left, increasing the length of the MPS by one. Fig. 2(b) shows the MPS after completion of the first time step. Once step n = n c is reached, the oldest path in the MPS is summed over by application of a delta tensor [semicircular shape in Fig. 2(a)], corresponding to the improved finite memory approximation [30,31]. At this stage, the MPS length is fixed for the rest of the time evolution. Moreover, for time-independent problems, apart from the index nomenclature the structure of the MPO remains unchanged for all time steps n ≥ n c , resulting in an additional performance gain.
To provide an example of a continuous reservoir of noninteracting harmonic oscillators, we consider the Hamiltonian [47] corresponding to diagonal system coupling without inducing system transitions, with system operators σ ij = |i j|, bosonic annihilation (creation) operators b ( †) q of reservoir modes with frequency ω q = c s |q|, c s the sound velocity, and a mode q-dependent system-reservoir coupling amplitude g q . The corresponding correlation function reads with temperature T and k B the Boltzmann constant.
In the following, we choose a generic coupling element e.g. describing acoustic bulk phonons interacting with a quantum emitter [48]. As a benchmark of the tensor network-based path integral implementation, we first calculate the analytically solvable Independent Boson model, consisting of a single two-level emitter subjected to pure dephasing by a structured harmonic reservoir, as described by Eq. (6). In Fig. 2(c), we prepare the initial polarization at ρ 01 (0) = 0.5i and calculate the resulting dynamics at varying temperatures T (solid lines). The corresponding analytical solution (dashed grey lines) is given by [49] ρ 01 (t) = exp (8) exhibiting excellent agreement with the numerical results at all considered temperatures. In addition, the employed method features very high performance, enabling reservoir memory depths of n c = 100 and beyond. To exemplify the capabilities of the tensor network implementation, we calculate the time evolution dynamics of the Spin-Boson model, corresponding to Eq. (6) with an additional continuous driving field term at amplitude Ω 0 , Fig. 2(d) shows the resulting dynamics at parameters n c = 100, Ω 0 = 0.5 ps −1 and T = 77 K for the memory, driving field and temperature, respectively.
II. TIME-DISCRETE MEMORY IN LIOUVILLE SPACE
As a second non-Markovian reservoir, we consider a discrete time-bin based quantum memory. Recently established implementations rely on an MPS-based wave function ansatz to compute the quantum stochastic Schrödinger equation [27][28][29]. However, this formulation is inherently incompatible with the previously introduced path integral formulation. As a solution to this problem, we present an MPS implementation of a time-discrete quantum memory in Liouville space. Here, the dynamics of the system density matrix is prescribed by a Liouvillian superoperator [see Eq. (1)], with a Hamiltonian H D containing the time-delayed system-reservoir coupling, such that interactions occurring at time t couple back into the system and affect its state at a subsequent time t + τ , with τ the retardation time. Such a time-discrete coupling e.g. arises in a two-level emitter with states |0 , |1 at an energy difference ω 0 , placed in front of a mirror with a round trip time τ . The corresponding Hamiltonian reads describing off-diagonal system coupling leading to energy exchange between system and reservoir and system phase relaxation, with system operators σ ij = |i j|, bosonic annihilation (creation) operators r ( †) k of photon modes with frequency ω k = ck, c the speed of light, and a constant electron-photon coupling amplitude Γ. For an efficient evaluation, the dynamics imposed by L(t) is translated in a time bin-based MPS formalism [27][28][29]37] which maintains the relevant system-environment correlations, scaling with τ . In case of additional phenomenological dissipative channels, the Liouvillian can be extended by the standard Lindblad operator [44,45].
For the MPS implementation of the time-ordered Liouvillian, we start from the formal solution of the system part of Eq. (1) for the density matrix, with T the time-ordering operator. For an MPS-based approach and in analogy to the time-discrete path integral formulation, we restate Eq. (10) in a time-discrete basis, which reads at time t N = N ∆t with at time discretization ∆t and with time-bin normalized operators For the MPS evolution of the density matrix during each time step n, the discrete Liouvillian time step operator L(n, n − 1) is approximated as a tenth order series expansion, i.e., Figs. 3(a)-(e) show the tensor network scheme for the implementation of the time-discrete memory. The square red tensor in Fig. 3(a) contains the system density matrix at the initial time t = 0. To consider a time-discrete memory, here n d = 4 circular tensors to its left store the reservoir state in Liouville space at preceding times, with the oldest state located on the left end of the MPS (blue). The reservoir states for all future time steps are initialized to the right of the system bin, containing full reservoir entanglement e.g. at finite temperature. In the following we assume an initial vacuum state. Therefore, during each time step a new empty reservoir bin (orange) is added to the MPS from the right, representing the present reservoir state [see Fig. 3(a)]. The memory loop realization explained in detail below introduces the retardation time τ = n d ∆t by the number of initial memory bins n d .
The first step of the time evolution is carried out as follows: By applications of the singular value decomposition algorithm [46], the first memory bin (blue) is pushed to the left of the system (red) while maintaining relevant entanglement information in the swapping procedure [see Fig. 3(a)]. The Liouvillian operator L(1, 0) for the first time step is then applied to the system bin, current memory bin and present reservoir bin, as shown in Fig. 3(b). Afterwards, the processed memory bin (grey) is swapped back to its original position and stored for the rest of the time evolution. The updated present reservoir bin (green) is pushed to the left, taking the role of a memory bin [see Fig. 3(c),(d)]. Fig. 3(d) shows the MPS after completion of the first time step. The second time step is carried out in the same fashion, as shown in Fig. 3(e). After completion of n d time steps, all initial memory bins have been processed. At step n d + 1, the reservoir bin modified during the first time step [green bin in Fig. 3(d)] becomes the current memory bin, containing information of a previous system state and setting off reservoir-induced memory effects in the system in complete agreement with the time-ordered problem.
As a first benchmark for the presented time-discrete quantum memory in Liouville space, we calculate the system dynamics imposed by Eq. (9). Fig. 3(f) shows the unfolding emitter population dynamics at a feedback time τ = 3.0 ps and Γ = 31.6 ps −1 calculated using the MPS implementation (solid blue line) and compared to its analytical solution up to t = 3τ (dashed grey line). The latter is given by σ 11 (t) = | σ 10 (t) | 2 only valid in the single-excitation regime with (14) as calculated in [50][51][52] and shows excellent agreement with the numerical result. In this regime, i.e. Γτ ≫ 1, the delay in the amplitude governs the dynamics and leads to re-excitations at multiples of the round trip time τ . The phase of the amplitude φ = ω 0 τ , however, loses importance in the first τ -intervals due to a stronger decay of the mixing terms in the absolute square of Eq. (14). The advantage of our implementation of quantum feedback in Liouville space becomes evident if the impact of phase destroying processes is in question. Up until now, this impact has only been investigated for a special case, finding the emergence of an Ornstein-Uhlenbeck process during the first τ -intervals [52]. These results have been obtained via analytical calculations, limiting the investigation to a small number of feedback intervals. Steadystate scenarios, however, are out of reach in this case as the evaluation of the phase-noise kernels must be done analytically. In our method, these limits have been overcome. Due to the here presented Liouville architecture, additional Lindblad-based dissipation can be easily implemented without increased numerical expense. In Fig. 3(g), we present the dynamics of a decaying, intially excited two-level emitter under the influence of quantum coherent feedback and additional phenomenological dephasing at rate γ, realized by adding a Lindblad dissipator to the Liouvillian [Eq. (1)] and H = H D , [44,45] with a redefined system operator in full configuration space,σ 11 = 1 D σ 11 1 D , including the time-discrete reservoir basis 1 D = dk ∞ n=0 |{n k } {n k }|. The time trace shows long time calculations of the emitter population, comparing the cases γ = 0 (blue line) and γ = 0.5 ps −1 (orange line). Most importantly, we see that the pure dephasing process becomes important only after the feedback signal re-excitates the emitter and the stabilization of the incoming and outgoing phase comes into play. In the presence of an additional pure dephasing γ = 0, the initial decay process is unchanged but the re-excitation becomes less efficient until only incoherent re-excitation takes place, leading to a faster decay to zero without population trapping, regardless of the choice of φ.
This important result sheds light on the robustness of quantum feedback processes in the presence of additional Markovian dissipation channels. As expected, additional Markovian decoherence leads to a washing out of the signal since a loss of quantum feedback-induced coherence is inevitable. However, this does not have to be the case in the presence of an additional non-Markovian dissipation channel, which we discuss in the following.
III. QUASI-2D TENSOR NETWORK
As a next step, we expand upon the MPS architecture for time-discrete quantum memory in Liouville space by combining it with the previously discussed tensor network-based path integral implementation for continuous harmonic reservoirs, resulting in a quasi-2D tensor network. The technical connection of the networks via link indices which store arising entanglement information enables the numerically exact description of correlation buildup in between the reservoirs. When considering a scenario involving two non-Markovian reservoirs not isolated from each other, such inter-reservoir correlations may have fundamental impact on the system dynamics, therefore prohibiting a strict truncation of the arising inter-reservoir entanglement, e.g. in the form of a low Schmidt value cutoff precision d cut . As a result of not only two system-reservoir interactions but additionally arising reservoir-reservoir entanglement, the overall grade of entanglement in the system rises intensively with respect to the twofold single reservoir case. On the other hand, in setups where two non-Markovian reservoirs are present but do not crucially interact, e.g. via dynamical decoupling, a much more restrictive truncation is possible without cost of accuracy. For the presented results, we have employed a high Schmidt value cutoff precision d cut = 10 −12 , such that no relevant entanglement information is lost during the time evolution. The dynamical interplay between the two reservoirs with the system and with each other poses an immense numerical challenge and strongly limits the accessible memory depths in the here considered system: While the two presented algorithms by themselves enable simulations of a single reservoir with deep memories, their combination is accompanied by limitations due to the arising interreservoir entanglement. As a result, the combined number of memory bins in the quasi-2D network is limited to n c + n d < 20 for our model of choice, as is the case in traditional path integral implementations for a single reservoir [11][12][13][14][15]43]. However, we stress once more that this limitation is a natural consequence of the high grade of entanglement in between the two reservoirs and the system. With the presented quasi-2D network architecture, we take first steps to unravel the mostly unexplored field of multiple interacting non-Markovian reservoirs by explicitly considering memory-enabled information backflow in between them.
The construction of the quasi-2D network is sketched in Fig. 4: The two tensor networks for the continuous [Figs. 4(a),(b)] and time-discrete reservoirs [ Fig. 4(c)] are connected to each other via the common tensor representing the current system state in both MPS algorithms (red shape). The system state tensor is employed to act as a junction connecting the two reservoirs [dashed circles in
Figs. 4(b),(c)]
and thereby enables the buildup and storage of inter-reservoir correlations in the connecting link indices. The time evolution of the quasi-2D network is carried out as follows: During each time step, the system is first evolved under the influence of the continuous reservoir by a single contraction of the network, as shown in Fig. 4(a) (dashed frame). Afterwards, the new current system state [red shape in Fig. 4(b)] is subjected to the second tensor network algorithm accounting for the time-discrete reservoir [ Fig. 4(c)]. In consequence, the quasi-2D network stores the history of both interactions, maintaining crucial entanglement information and enabling the calculation of two dynamically interacting time-delayed processes. As an example application for the quasi-2D network, in the following we investigate the interplay of off-diagonal coherent quantum feedback and a diagonal reservoir of independent oscillators. As illustrated below, for certain memory depths and initial states, this results in dynamical protection of coherent quantum feedback properties in the open system.
IV. MEMORY-INDUCED DYNAMICAL POPULATION TRAPPING
Coherent quantum feedback mechanisms exhibit a rich variety of non-Markovian phenomena [51,[53][54][55][56][57][58], e.g. enabling coherent population trapping [28,42,[59][60][61][62], Ornstein-Uhlenbeck-type events in the presence of white noise [52], and formation of large entangled photon states [63]. However, so far these effects have not been explored in the presence of additional non-Markovian decoherence or dissipative channels. To investigate the impact of dephasing on feedback-induced decoherence, we consider a two-level emitter placed in front of a mirror with a round trip time τ , taking the role of a time-discrete reservoir (see Fig. 1) described by Eq. (9). The photoninduced feedback imprints a time-delayed coherence in the form of a feedback phase ϕ = ω 0 τ /(2π) on the system, critically influencing its dynamics. It is given by the delay time τ and the transition frequency ω 0 of the electronic coherence operator σ 12 . To study a pronounced quantum optical effect, we consider the case of an initially excited two-level system where coherent population trapping occurs as a result of a bound state in continuum at feedback phases ϕ ∈ Z [28,42,52,[59][60][61][62]. The unfolding dynamics are evaluated for parameters Γ = 0.9 ps −1 , τ = 1.2 ps and n d = 4, resulting in a time discretization ∆t = 0.3 ps, e.g. typical for semiconductor quantum dot based devices [58,61,62]. Moreover, we get Γτ ≈ 1.1, corresponding to the strong non-Markovian regime [27,52,64]. The employed series expansion of the Liouvillian up to tenth order [see Eq. (13)] justifies this coarse time discretization, making n d = 4 time bins sufficient for our investigation while resulting in convergent results (see Appendix A).
To illustrate the power of our method, we compare the cases of Markovian and non-Markovian dephasing introduced by an additional diagonal system-reservoir coupling, representing the continuous reservoir in Fig. 1. As a first step, we calculate the system dynamics given by Eq. (9) in the presence of phenomenological dephasing at rate γ, introduced by the Lindblad dissipator stated in Eq. (15). Fig. 5 shows resulting excited state population dynamics at varying γ and feedback phases ϕ. At γ = 0 there exists a periodic ideal feedback phase ϕ ∈ Z such that the system decouples from its environment by constructive interference, resulting in coherent population trapping (solid blue line in Fig. 5). Choosing a nonzero dephasing γ = 0.001 ps −1 has no impact on the population dynamics until feedback sets in, since the radiative decay is frequency-independent until t = τ (dashed blue line). Thereafter, phenomenological dephasing destroys the phase interference and with it the trapping mechanism, resulting in an asymptotic decline of occupation to zero. At a nonideal feedback phase, here ϕ = 1.17, and no dephasing, destructive interference leads to an asymptotic decline towards zero as well (solid orange line). The decay is further accelerated by setting γ > 0 (dashed orange line), since any phenomenological decoherence attacking the phase relation ϕ results in faster decay. In conclusion, choosing a feedback phase ϕ / ∈ Z without a structured phonon reservoir inevitably results in asymptotic population decay via thermalization, and a Lindblad formulation of decoherence never preserves quantum correlations between the reservoir and system states.
As a next step, we show this is not necessarily the case if the decoherence process itself is the result of a non-Markovian reservoir interaction. Using the quasi-2D tensor network, we calculate the emitter dynamics imposed by H D + H C [Eqs. (6), (9)] at n c = 4 and leaving all remaining parameters unchanged. Fig. 6 shows resulting population dynamics at different temperatures. For the chosen parameters, we find population trapping for the non-ideal feedback phase ϕ = 1.17, i.e. ϕ / ∈ Z, at T = 4 K (solid line). Time-delayed excitation backflow from the continuous reservoir of oscillators to the system enables a decoupling from destructive interference with the time-discrete photon environment. This information backflow results in correlation buildup and information exchange between the reservoirs, dynamically protecting feedback-induced coherence in the system for long times. As long as diffusion processes at finite temperature take place on a comparable time scale as the coherent feedback dynamics, we always find dynamical population trapping by tuning of ϕ after a typical excitation backflow time. At higher temperatures, it is to be expected that thermal properties of the phonon reservoir start to dominate the dynamics: Dashed and dotted lines in Fig. 6 show corresponding thermalization dynamics at T = 30 K and T = 77 K, respectively, exhibiting population decay. The temperature dependence clearly shows that correlation lengths within the full system-reservoir dynamics are of importance, and the observed effect allows to probe these otherwise inaccessible microscopic environmental properties. This formation of self-stabilizing dissipative structures is closely related to a localized phase stabilization in the coherent driving case Ω 0 = 0 at Ohmic spectral densities and without photons [31,43]. There, above a critical coupling strength the system transitions into a localized phase with nonzero steady state population rather than decaying to zero. In our case, a similiar phenomenon is established with incoherent feedback instead of coherent external driving, addressing the localized phase stabilization process from a dissipative non-Markovian side. The complexity of this phenomenon is illustrated in the inset in Fig. 6, showing the dimension of the link index connecting the current open system bin to the discrete memory MPS over time (see Fig. 4). The red line shows the case including the continuous reservoir. After slowly increasing during times t < τ , it exhibits a vast increase once feedback sets in and quickly reaches a maximum due to finite memory. The high grade of entanglement between the two reservoirs even at arguably low finite memory sizes n d = 4, n c = 4 underlines the crucial role of the interplay between the two non-Markovian processes for the observed protection of coherence. Switching off the phonon coupling, g q = 0, results in a much lower maximum link dimension (blue line), as no entanglement between the reservoirs arises. For phenomenological dephasing (orange line), the link connecting system and memory bins has an even lower dimension due to the highly decreased complexity of the then 1D network.
V. CONCLUSIONS
We have presented an MPS algorithm for the description of a time-discrete quantum memory in Liouville space. By combining this technique with a path integral tensor network implementation for continuous non-interacting harmonic reservoirs, we have established a quasi-2D tensor network, allowing for simulations of quantum systems subject to two non-Markovian environments while maintaining crucial entanglement information in the coupled system with both diagonal and off-diagonal system-reservoir interactions. Due to arising reservoir-reservoir correlations, system correlations scale intensively with respect to the twofold single reservoir case. In consequence, the achievable memory depth is limited to n c + n d < 20 in our study. However, appropriate tuning of the relevant system and reservoir time scales via the employed parameters still opens up a wide array of accessible systems and scenarios where numerical convergence can be achieved. The next step will be to trace out the time-discrete feedback bins as well after their interaction with the systems' degrees of freedom to further improve numerical efficiency. This will allow for longer delay times and therefore increased time discretizations without changing the qualitative results. Hence, the presented quasi-2D tensor architecture is a first step towards Fig. 6 with respect to (a) the continuous reservoir memory depth nc and (b) the Schmidt value cutoff precision dcut applied during the singular value decomposition. In (a), the left inset shows a zoom-in on the long term dynamics, with the right inset depicting the resulting dimension of the link index connecting the two tensor networks. The inset in (b) shows a zoom-in on the long term dynamics.
unraveling the mostly unexplored field of multiple interacting non-Markovian reservoirs in a numerically exact fashion.
As an example application, we have demonstrated that the interplay of a structured phonon reservoir and photon feedback can dynamically protect the system from destructive interference by time-delayed backflow of coherence, resulting in dynamical population trapping. Tuning the non-Markovian interactions with respect to each other allows for the formation of inter-reservoir correlations, dynamically preservering feedback-induced coherence in the system. These findings have implications for the fields of quantum thermodynamics and nonequilibrium physics, as well as dynamical quantum phase transitions with ergodic, entropic or negentropic information exchange, where taking account of such dissipative structures may unravel new phenomena. Future works will aim to advance our architecture to a full 2D representation via projected entangled pair states with combined memory bins, potentially allowing for simulations of multi-level systems at improved time resolutions. Fig. 2) and (c) the combined scenario (see Fig. 6). (d) shows a convergence analysis with respect to the order of the series expansion performed in the Liouvillian L(t), with the inset showing a zoomin on the long term dynamics.
Appendix A: Convergence analysis
Here we provide a detailed analysis on the numerical convergence of the presented results. In light of the memory limitations imposed by an intensive scaling of entanglement when describing two non-Markovian reservoirs simultaneously, we first confirm convergence with respect to the memory depth of the continuous harmonic reservoir n c , i.e., the validity of the finite memory approximation for the considered scenario. Fig. 7(a) shows the population dynamics corresponding to Fig. 6 at ϕ = 1.17, T = 4 K and increasing memory depths n c of the contin-uous reservoir. The left inset shows a zoom-in on the long term dynamics, while the right inset depicts the resulting dimension of the link index connecting the two tensor networks, demonstrating the here occurring exponential growth of inter-reservoir entanglement with increasing memory depths. While the results are clearly not convergent at n c = 2 (dark blue line), the cases n c = 3 and n c = 4 (light blue and green lines) already are in very good agreement. Between the cases n c = 4 and n c = 5 (dashed orange line), no difference can be seen even at close range (see left inset), hence justifying the choice of n c = 4 employed in our calculations. Fig. 7(b) shows the same dynamics calculated at increasing Schmidt value cutoff precisions d cut applied during the singular value decomposition [46]. Here, no differences between the cutoffs d cut = 10 −8 (dark blue line) and d cut = 10 −14 (dashed orange line) can be seen even at close range (see inset), underlining the convergence of our results with respect to the employed cutoff precision d cut = 10 −12 (green line). Hence, no crucial entanglement information has been truncated during the time evolution.
In addition, we investigate the numerical convergence of our results with respect to the time evolution step size ∆t to ensure that no errors occur during the Trotter decomposition. We first calculate the dynamics of both reservoirs independently for decreasing step sizes. Fig. 8(a) shows the time evolution dynamics resulting from the time-discrete reservoir at ϕ = 1.17 for decreasing step sizes ∆t (see Fig. 5). Aside from minor differences during the initial time steps, the long term dynamics show very good agreement for all employed ∆t, underlining the convergence of the feedback algorithm at the chosen step size ∆t = 0.3 ps (light blue line). Fig. 8(b) shows the polarization dynamics imposed by the continuous reservoir [see Fig. 2(b)], calculated at T = 4 K and decreasing time evolution step sizes ∆t. Again, the resulting long term dynamics at ∆t = {0.3 ps, 0.15 ps, 0.075 ps} are in good agreement. Since both processes converge individually with respect to the time discretization, the combined setup can be expected to converge as well, since the involved time scales remain the same. Fig. 8(c) shows the dynamics of the combined system coupled to both structured reservoirs for time discretizations ∆t = {0.3 ps, 0.24 ps}, corresponding to n d = {4, 5} feedback bins. Due to limited computational resources imposing restrictions on the memory depths of both reservoirs and thus the maximum achievable time discretization at a given feedback time τ , the presented calculations are performed until t = 10 ps, where they exhibit good agreement with each other. Lastly, in Fig. 8(d) we verify the convergence of the numerical implementation with respect to the order of the series expansion performed for the Liouvillian L(t) [see Eq. (13)]. While the resulting dynamics show minor variations between the eighth and ninth order series expansions on a close scale (blue and green lines, see inset), no difference can be observed when comparing the dynamics resulting from ninth and tenth order expansions (dashed orange line). In conclusion, the employed tenth order series expansion of L(t) ensures convergent results as well. | 7,676.2 | 2020-11-10T00:00:00.000 | [
"Physics"
] |
Small ruminants as a pathway to reduce urban poverty: An empirical analysis of Sudan
Aim: This study aimed to measure the energetic incidence of poverty and determines the main factors that cause urban poverty. Moreover, the study examines the key role of the livestock sector in poverty reduction in urban regions and develops an analytical tool to aid in urban area poverty mitigation through goats and sheep ownership. Materials and Methods: The study mainly depends on primary data assembled through structured and unstructured questionnaires, which were distributed among the targeted groups in the urban area in Sudan. Poverty line and poverty indices were calculated and measured using various well-known methods. The causes of poverty were estimated using logistic regression, and the effect of small ruminants in poverty alleviation was estimated using multivariate regression analysis. Results: The study findings indicate that both food and income poverty lines are less than the standard poverty line. In addition, the results imply that rural migration and crime predictors are among the most important factors in increasing urban poverty in the study area. Furthermore, livestock ownership has a significant impact on poverty reduction. Conclusion: The study concludes that small ruminants are playing a key role in reducing urban poverty. Thus, the study urges planners and policy-makers to support policies that promote livestock sector development as a strategy to alleviate poverty in Sudan.
Introduction
Sudan is an agrarian developing country classified as low income with a per capita income of <$1,045 (USD) per year (2019). The Gross Domestic Product at constant prices in 2016 was 41.3 [1]. The last National Household Budget and Poverty Survey in Sudan conducted by the Central Bureau of Statistics (CBS), during 2014/2015 confirmed that 36.1% of the population in Sudan was living in absolute poverty while urban poverty accounted for 39.9%. The per capita/year is 2966 Sudanese Pound (SDG) in urban areas, which displays higher average consumption levels than rural areas (SDG 2698). Poverty assessment in Sudan has been limited, but studies have provided evidence of high-income poverty. A study conducted by Faki [2] analyzed poverty with wide coverage by state and based on aggregate consumption of five main components (food, non-food, durable goods, housing, and energy). It puts North Sudan at an overall poverty level of 46.5%. Rural poverty was way above urban poverty (27%). An authorized study performed by CBS [1] conducted all over Sudan regions reported that the poor population in urban areas is slightly higher than in rural areas.
Livestock production is a dynamic sector for the National Economy in Sudan. Sudan total domestic livestock population in 2017 was about 104 million heads. Goat populations constituted more than 31 million heads, while sheep population constituted nearly 40 million heads [3]. At the national level, the livestock sector is characterized as the most active sector in national income. The livestock shares are estimated to be 18-25%, and it shares foreign currencies through the export of the various livestock products in the form of live animals, meat, and leathers. In addition, it represents a livelihood activity for about 60% of the population and provides employment for about 40% of the population [4]. At household level, livestock is the main source of food, employment, income, transportation, prosperity, and enhancing crop production by providing organic manure fertilizers and draught power. Shrestha et al. [5] indicated that when crops are not sufficient to ensure food security, livestock can be used as a source of food.
Nowadays, in Sudan, no adequate studies have investigated urban poverty reduction, whereas a Available at www.veterinaryworld.org/Vol.12/December-2019/19.pdf majority of international, national, and regional research has focused on rural poverty reduction [6][7][8][9]. The lack of updates to the poverty line and studies of economic policy in urban areas have made it difficult to make comparisons across time, particularly in recent years, through which the country's economy has deteriorated and marginally fallen. Moreover, no recent research has been accomplished toward poverty reduction in Sudan. Likewise, civil wars and political problems led to successive migration from unsecured areas, particularly from rural areas, to urban areas. The unplanned migration created various and serious problems faced by the migrators. Hence, most of the migrators have no access to the basic needs of life and struggle to live in a new environment. Furthermore, the abrupt financial fall of the economy successively led to increased poverty among the rural migrators; those migrators reverse their initial wealth, which are mainly seasonal crops and rearing of livestock.
This study aimed to analyze and measure the poverty indices in the urban areas and to identify the main factors beyond urban poverty causes. Furthermore, the study attempts to construct an identical strategy approach of urban poverty reduction by practicing small ruminants in the indoor household.
Ethical approval
For this study, the collection of the questionnaire was approved by the Faculty of Animal Production, University of Gezira, Sudan. The oral permission was obtained from the target's household head in the study area.
Methodology and data description
Bahri region is selected to act as the study area (located in the Eastern part of Khartoum National Capital). It was selected because the majority of livestock is distributed in this region compared to other regions around Sudan capital. The study mainly depends on primary data assembled through structured and unstructured questionnaires which were distributed among the targeted groups in the urban area in Sudan. About 300 questionnaires were distributed randomly for the household heads in the Bahri region (year 2017/2018). As the study mainly concentrated on extreme poverty (food and non-food expenditures) in the region, the primary data collected to cover the information of monetary (economics data) and nonmonetary indicators (demographic and socioeconomic). The major data include household head ages, jobs, family size, income, foodstuff (types, purchasing prices, consumption, etc.), health, education, housing conditions, water aspects, and its problems, social participation, political conflicts, crimes, and various features of livestock issues.
Poverty line and measurement method
The poverty line is calculated based on dietary intake kilocalories per equivalent, which is later termed as food poverty, and the situation of household income and expenditures is termed as money or income poverty.
Poverty is a complex and multifaceted problem, and varieties of methodologies are used for poverty calculation and examination. According to objectives and the nature of the data collected, poverty was calculated using the Foster-Greer-Thorbecke (FGT) method suggested by Foster et al. [10]. The investigation follows the cost of basic needs method to poverty analysis, which was performed in many African countries. The three sound methods of poverty measures (headcount ration, poverty gap, and poverty square) were analyzed by succeeding FGT method. In addition, a fourth measure (Watts's index) was calculated. Watts's index takes into account both income and the number of people in poverty [11].
Poverty line
Assuming that the Y i denotes household income and Z denotes the poverty line, both income and poverty line were calculated by comparing the total expenditure of the individual household by the international poverty line, which was originally set to be roughly <$2/day [12]. In addition, poverty line was calculated by summation of the kilocalories of Sudanese food staff to the standard kilocalories requirement according to the World Health Organization. The minimum level of Sudanese household's expenditures should enable them to buy 2100 kcal of food per person per day [13]. If the individual consumes less than the required kilocalories (2100), it was considered poor; otherwise, it was nonpoor. Poor = 1 if C ≤ 2100 and non-poor = 0 if C > 2100, whereas the C = total consumption of kilocalories per day. For estimating the number of poor and non-poor according to their daily expenditure in USD, we assign the value 1 when income ( y i ) falls below the poverty line (z) and 0 if income fall above the poverty line. Thus, poor = 1 if y i ≤ z and non-poor = 0 if y i > z.
Poverty indices
The general FGT equation of poverty indices was measured as below:
The headcount
The headcount was calculated by matching the Y i of each household to Z. The headcount index (H i ) was the sample average of the variables weighted by the number of people in each household n i . The parameter α determines the measure sensitivity to the degree of deprivation for those below the poverty line [10]. When α equals zero the measure reduces to the below formula:
Poverty gap (depth)
The poverty gap of urban people represents the depth of poverty. It means the distance separating the people from the poverty line. It was generated when the parameter α equals s one, which can be defined as follows: Whereas P ug is the urban poverty gap.
Poverty square (severity)
Poverty square or severity reflects the inequality between the poor household. This measure is illustrated by the below equation: The study tries to calculate Watts' index, which is a function of an individual income and poverty line. It is the first distribution-sensitive poverty measure that was proposed by Watts [11]. Watts' index is calculated by dividing the poverty line by income of the urban sample of the adult equivalent [14]. Taking logs, (which involves the computation of the logarithm of each income) and taking the sum over the poor and the household size variable n i replaced with the adult equivalent size a i . According to World Bank [15], the equation of Watts index is represented below: Where s is the total number of the household in the sample.
Logistic method of urban poverty causes
The study follows various researchers [16][17][18][19][20] to estimate poverty causes using binary logistic regression. Binary regression is used to predict the relationship between dependent variables (predicted variable), which is dichotomous and represents poor versus non-poor households. The independent variables (predictors) represent the demographic features of the sampled population and the types of livestock.
To identify key determinants or causes of poverty, a dichotomous variable indicating whether the household is poor or not is computed [19]. Estimated probability of being poor, given the values of explanatory variables, in this study is a single categorical variable: π = is predicted probability, P r = probability, and P is a binary variable of poor household. P i = 1 if the household is poor in observation i, otherwise P i = 0, if the household is not poor in observation i.
The X s values are a set of predictors, which can be discrete, for instance, number of animals, family size, etc., also, it can be continuous, such as age and income.
Later the poverty binary model could be: β β ε + = + (9) β 0 = a cluster fixed or random effect and β is a vector of parameters.
x i is a vector of household demographic features or others explanatory variables, which include a set of individual characteristics X (e.g., gender, education, age of the household head, and livestock ownership). Furthermore, some political variables are included in the model, for example, rural migration due to crimes and/or conflicts.
Poverty reduction model
One of the key issues of various international institutions [21,22] is to suggest and adopt the various types of policies and strategies for poverty reduction in the world in general and poverty reduction in developing countries in Africa, specifically. Hence, widespread malnutrition and most African nations are agrarian and depend mainly on crop production as the main source of income generation. The strategy used for urban poverty reduction in this study is through livestock practicing. The study selected two livestock, which is goats and sheep. Thus, these two types of livestock are much cheaper and simply can be kept inside homes.
Poverty line elasticity model
Numerous researchers investigate poverty reduction, taking into account the different forms and formula of poverty elasticity with respect to growth, income, inequality, and standard of living [10,[23][24][25][26].
Our study tries to analyze the poverty lines' (income and food poverty) elasticities to display the responsiveness, or how poverty lines change in response to a livestock production change.
To explain the role of livestock in poverty eradication in urban areas, the study uses a regression model to evaluate the incorporation of the livestock in the household to reveal the importance of practicing livestock activities in the household. Multiple regression procedures are constructed to explain the relative response of per capita income to increases in goats and sheep heads, and hence the poverty alleviation. The equation can be written as follows: Equation (11) is a log-linear model aiming to estimate the poverty elasticity with respect to numbers of sheep and goats owned by the household as follows: Where Y p = poor income and the predictor variables are goat and sheep heads, respectively. Extra equations of the multiple regressions are performed to show the effect of livestock production in food poverty reduction. In this equation, the dependent variable used is food poverty line and the independent's variables used are consumption of both milk and meat in the households, as shown by Equation (12). (13) is a log-linear model to estimate poverty elasticity with respect to milk and meat consumption by the household as follows: Whereas Z = food poverty line, MC = milk consumption, and MTC = meat consumption. The regression models were executed using the capabilities of SPSS, SAS, and Excel software programs.
Results
The descriptive results show that 65% of respondents are males while 35% are females. The majority of females migrated from civil wars or conflict areas in Western or Eastern regions or escaped from natural hazards in Northern regions.
The average age of the surveyed household headed is 52 years. The average family size is found to be approximately eight people, and this is a normal phenomenon in Sudan.
Considerable numbers of the surveyed samples (35%) are unschooled and almost illiterate, and practicing the informal jobs ranged between street sellers to hired labors with marginal employers. Only 4.3% received university education and those were the only segment of the sampled survey who own houses. According to the food poverty line (household member that consumed (<2100 kcal), the numbers of the poor constituted nearly 222 (74%) while the non-poor are 78 (26%). Most of the poor households are females (85%) and the remainder are males (15%).
Urban food consumption and poverty line
From Table-1 [27], it is clear that household actual consumption is 1888.09 cal, which is less than the recommended calories by 9.6%. It is notable that coffee comprises a higher food poverty line (SDG 107.96) followed by cereal food (SDG 62.35). In addition, the study result confirms that the food poverty line is 8.48/household and equal to $1.04/person (average family size eight people). The non-food expenditure is estimated to be 132.5 SDG. Furthermore, the extreme poverty line is estimated to equal nearly 11.80 SDG/ household and $ 1 1.47/person (Tables-1 and 2).
As shown in Table-2, the urban poverty measurements reveal that the proportion of those who live below the poverty line is 74%, which indicates that more than half the surveyed sample is poor. Likewise, the poverty depth and severity are estimated to be 68.28% and 45.26%, respectively. Furthermore, it implies that 68.28% of the poor are slightly far from the poverty line (1.04). In addition, the mean proportionate poverty gap of the urban poor is high, as shown by Watts' index (55.25%).
Urban poverty causes
The connection between risk factors and the incidence of urban poverty is shown in Table-3.
The goodness of fit of the model is shown to be significant (χ 2 = 39.58). The logistic regression result shows 1 One USD = 32.50 SDG at the time of survey. that there is an increase in the likelihood of being poor with job types, family size, number of males, crimes, diseases, affection, and migration (the odd ratios >1). Male household head age, education level, and livestock ownership have odds ratios <1, which indicate that the likelihood of being poor is decreased for these variables. Education level has a low value of odds ratio (0.460); this indicates that the poor living in urban regions have more access to education. Moreover, the findings show that households with access to clean water were significantly less likely to be in poverty compared to households without access to clean water. However, residents living in extreme poverty suffer from the risks of other factors, such as no access to health care and disease infections (including malnutrition). Migration and crime predictors were among the most important factors in increasing urban poverty in this study.
Role of small ruminants in urban poverty reduction
Ram [26], using the panel data of Vietnam, suggested that livestock production contributes to poverty reduction. The multiple regression models were run to estimate the impact of livestock on poverty reduction (Tables-4 and 5). The overall model is significant (F=59.64, R 2 =0.570).
Based on the multiple regression model results, the study shows that the increase in goat numbers lead to increase in the per capita income of the urban poor by 33.79 SDG. In addition, when the sheep numbers increase by one head, the per capita income increases by 41.57 SDG (t=5.613). Furthermore, a 1% increase in goat and sheep heads increases per capita income by 0.54% and 0.30%, respectively.
On the other hand, the results in Table-5 show a direct connection between poverty reduction and consumption of livestock products (meat). If meat consumption increases by 1 kg, the kilocalories of the household increase by 126 kcal. However, milk consumption had an insignificant effect on households' kilocalories.
Based on the elasticities estimation, all results were inelastic, as shown in Tables-4 and 5. Table-5 shows that a 1% increase in meat consumption results in a 0.957% increase in the urban poor kilocalories.
Discussion
Urban poverty becomes core attention of the developing countries' governments within which the people suffer from refugees, civil wars, and resource conflict. Adequately, studies are performed worldwide using different methods of urban poverty measurements and roughly compare rural poverty to urban poverty.
In this study, the results show that there is an increase in the likelihood of being poor with family (Table-3) and this result confirmed by Kabir and Maitrot [28] who stated that the larger the size of the beneficiary's household, the more negative the effect on economic growth and well-being. Furthermore, this study reveals that migration is constantly contributing to increase poverty incidence in the urban area; however, Christiaensen [29] showed that in Tanzania, the migration to secondary towns helps in poverty reduction.
In addition, De Janvry and Sadoulet [30] analyzed the change in the relative number of rural and urban poor using aggregate analysis. Their findings showed that the incidence of rural poverty was declining relative to the incidence of urban poverty, and the population was rapidly leaving in the rural sector. In addition to the universal increase in commodity prices, food prices steered to increase urban poverty incidences. The study results indicate that food price is high, which is influenced by urban poverty and this result agrees with the findings argued by Meng et al. [31] that the poverty line is linked positively with an increase in relative food price. Study shows a deep level of deprivation and higher incidence of poverty for urban people who are under the poverty line in the study area, while De Janvry and Sadoulet [30] found that urban poverty dominates aggregate poverty; the urban poor captures more than half of the aggregate increase in real income. Furthermore, De Janvry and Sadoulet [30] indicated that the incidence of rural poverty in Latin American countries is considerably higher and deeper than the incidence of urban poverty. The study result displays that the poverty income is sensitive in the urban area; however, De Janvry and Sadoulet [30] stated that the overall rural poverty is less sensitive to aggregate income growth compared to urban poverty and Yamada [32] used the quantile regression model and argued that the coefficients in urban areas are larger than rural area. Moreover, the regression coefficients have been decreasing slightly as time passes and do not have constant changes across the deciles. Employment tends to affect food poverty dynamics differently in urban and rural areas. The results of logistic regression used in this study reveal that there is an increase in the likelihood of being poor with job types. Eigbiremolen and Ogbuabor [33] used the dynamic food poverty continuous model in Nigeria to examine the impact of selected covariates on (log) growth of food consumption expenditure and the outcome showed that urban households which have a household head that is employed, increases his/her per capita spending on food over time in relation to households whose heads are unemployed. Thus, the authors conclude that employment reduces food poverty. Cho [34] indicated that households in urban areas were found to be multi-dimensionally poor. Some researchers went beyond child labor poverty in urban areas, such as Dayioğlu [35], who investigated the determinants of child labor in urban Turkey with low household income or in poverty using a probit model. The results showed that child employment goes down with household income. However, the effect of household income on child employment is not great, and the likelihood of children's employment is highly significant.
Poverty reduction has been a political priority from most developing nations and received great attention to international organization agendas [26,36]. The strategy used for urban poverty reduction in this study is through livestock practicing. Accordingly, different strategies and policies of poverty reduction are executed worldwide. Alwang et al. [37] examined the poverty reduction through innovations to improve staple crop germplasm and found that the main difficulties to greater poverty reduction include limited access to credit, services and markets, and small landholding sizes of poor farmers. Furthermore, the authors confirm that landholding size is an important barrier to poverty reduction. In addition, farm's technologies that enhance farm productivity can be a sustainable pathway to improve household food security and to enable households to climb out of poverty [38]. However, van Noordwijk [39] examined integrated natural resources management as a pathway to poverty reduction, while Benfica et al. [40] argued that providing extension services to smallholders is most effective at raising growth and reducing poverty. An increase in the output would increase farmers' income and reduce poverty in rural areas [41]. The farmers adopting more adaptation practices have higher food security and a lower level of poverty [42]. Martin [43] argued that the focus should be on the sustainability of local socio-technical systems, even if the options chosen are less efficient in the short term. Hansen [44] argued that climate-risk management interventions could play in efforts to reduce rural poverty. There are no studies and/or reviews that empirically analyze the relationship between poverty reduction and women's human rights [45].
The study results confirmed that urban poverty could be alleviated and reduced by the consumption of livestock products (Table-5). By viewing the alleviation and reduction of poverty through practicing livestock, remarkable investigations have performed in this matter. Do et al. [46] evaluated the impact of livestock production on poverty reduction and estimated the average treatment effect on the treated using the matching-difference-in-difference method. They indicated that owning a large livestock size meaningfully reduces the depth of poverty, and access to dairy livestock assets can provide important benefits for women [36]. However, Ali [47] argued that despite climate risk, decreased livestock production increased household income and lowered poverty levels.
Conclusion
Most developing countries face both rural and urban poverty during recent years. Based on the case study in urban Sudan, this article measures the incidence of poverty and determines the main factors that cause urban poverty. Furthermore, the study examines the vital share of the livestock sector in urban poverty reduction and develops an analytical tool helping to mitigate poverty in urban areas through goats and sheep ownership. The study adopted the standard tool of poverty measurements and estimated the urban poverty line and poverty indices (headcount, depth, severity, and Watts index). Furthermore, the study used logistic regression to reveal the likelihood of poverty determinants. Multiple regression models were also used to show the role of livestock in poverty reduction. In addition, livestock elasticities were estimated to display the responsiveness of poverty lines to changes in livestock production.
The model incorporated small ruminants as a strategy for poverty alleviation. The results indicate that both food and income poverty lines are less than the standard poverty line. In addition, the results implied that rural migration and crimes are among the most important factors contributing to the increase in urban poverty in the study area. The study concluded that small ruminants are playing a key role in urban poverty reduction. The study suggests that planners and policy-makers should encourage policies that promote livestock development and increase financial credit of livestock production. The purpose is to escalate the urban poor beyond poverty and helps them to obtain the basic needs of life, especially for the migrators who escaped from vulnerable regions. | 5,811.4 | 2019-12-01T00:00:00.000 | [
"Economics"
] |
Mobile Device ECG Classification using quantized Neural Networks
In this paper, a novel method for classifying electrocardiogram signals in mobile devices is proposed, which classifies different arrhythmias according to the Association for the Advancement of Medical Instrumentation standard EC57. A convolutional neural network has been constructed, trained and validated with the MIT-BIH Arrhythmia Dataset, which has 5 different classes: normal beat, supraventricular premature beat, premature ventricular contraction, fusion of ventricular and normal beat, unclassifiable beat. Once trained and validated, the model is subjected to a post-training quantization stage using the TensorFlow Lite conversion method. The obtained results were satisfactory, before and after the quantization, the convolutional neural network obtained an accuracy of 98.5%. With the quantization technique it was possible to obtain a significant reduction in model size, thus enabling the development of the mobile application, this reduction was approximately 90% compared to the original model size.
Introduction
Cardiovascular diseases are one of the leading causes of death worldwide. This type of pathology affects the cardiovascular system, specifically the blood vessels and the heart. There is a growing rate of the population suffering from cardiovascular diseases, including arrhythmia [1]. Sudden death caused by cardiac arrhythmia is a major public health problem worldwide, accounting for 15% to 20% of all deaths. It is estimated that 180,000 to 300,000 sudden cardiac deaths occur in the US annually [2].
In the clinical diagnosis of heart disease, arrhythmia indicates a severe change in heartbeat function and may cause a stroke or sudden cardiac death if left untreated [3]. For detection of cardiac anomalies, the use of electrocardiogram (ECG) is a fundamental diagnosis. This clinical analysis can perform the monitoring of cardiovascular diseases, where this procedure is done manually. However, there are several problems with manually analyzing ECG signals, such as: similarity to other time series data, difficulty in detecting and categorizing different waveform and signal morphology. For a human being, this task is time-consuming and error-prone [4], [5].
To solve problems with manual ECG signal analysis, many studies in the literature use machine learning techniques to accurately detect signal anomalies [6], [7], [8]. However, in these approaches a final application to the learning models is not proposed, a final application is important because the death rate caused by these anomalies is high, as already mentioned.
Technological advances and increased computational power make possible to use techniques that can assist in clinical diagnosis, such as artificial bioinspired Neural Networks in human neurons. As an example, Convolucional Neural Networks (CNNs) [9], [10]. But CNNs require a high computational cost in end applications, which makes the development of these applications a challenging task.
This paper presents a mobile ECG signal classification system capable of classifying different arrhythmia according to the Advancement of Medical Instrumentation (AAMI) EC57 standard [11], [12]: normal beat, supra-ventricular premature beat, premature ventricular contraction, fusion of ventricular and normal beat, unclassifiable beat.
The proposed method consists in the implementation of a CNN for the arrhythmia classification task, the model was trained and tested using the MIT-BIH Arrhythmia Dataset [13], containing 109,449 ECG signal samples. After training and validation, the model was submitted to a quantization stage, where the post-training quantization technique was used, using the TensorFlow Lite (TFL) [14] conversion method. This technique is to drastically reduces model size, power consumption and processing, enabling the development of a mobile application using a CNN.
This article is divided into seven sections. In section 2, it consists of a summary of the literature. Section 3 describes the data set as well as its analysis, while section 4 presents the methodology and its steps. In section 5 describe the evaluation metrics of the neural network, 6 in turn, is about the results and discussions, and finally present the conclusion in section 7.
Related Work
Analyzing the state of the art, it is possible to notice the increasing number of cardiac signal analyzes, as an example of Wang et al. [15], which aims to propose a method of heart rate ECG identification, suitable for short-term signals. The method aims to completely preserve the original temporal and morphological information of the QRS complex, solving the problem of T-wave displacement. They use the Principal Component Analysis Network.
Berkaya et al. published an ECG survey, which consists of a literature review related to ECG analysis [16], considering the following aspects: preprocessing, resource extraction, resource selection, resource transformation, classification, application fields, databases and success measures, is also mentioning the following: most used classifiers in the literature.
Lu et al. proposed a classication system with generalization capacity [17], where a resource extraction and balancing through the Random Sampler Algorithm was used, as a classifier was used the Random Forest, thus obtaining precision results above 99.0%.
ECG signal processing techniques for real time analysis are implemented in Raj at al. [18] and Varatharajan et al. [19], They use the Support Vector Machine algorithm for pattern recognition. These methods can be used for screening and pathological classifications, as well as a weighted kernel to identify Q, R, and S waves at the ECG signal input to classify the pulse level.
In [20], Zihlmann et al. explains the use of two Deep Neural Networks (DNNs) architectures for ECG classification by assessing the atrial fibrillation classification data set provided by PhysioNet/CinC Challenge 2017. The first architecture is a CNN with mean-to-length feature aggregation. over time and second is a convolutional recurrent neural network that combines a 24-layer CNN with a long-shortterm-memory 3-layer network for temporal aggregation of features In [21], Hannui et al. demonstrated the implementation of a DNN to classify 12 heart rate classes using 91,232 single-lead ECGs from 53,549 patients who used a single electrode ambulatory ECG monitoring device. DNN reached an average area under the characteristic operating curve of 0.97.
In [3], Yang et al. proposes a new method of arrhythmia classification through ECG with Statement on Standards for Attestation Engagements and a softmax regression model. The algorithm is employed to hierarchically extract high level resources from the huge amount of ECG data.
The article of Xia et al. [22], features an automatic wearable ECG classification and monitoring with a stack denoization autoencoder system. Using a wireless sensor device to retrieve ECG data and send that data to a Bluetooth 4.2 computer, where softmax regression is used to rate the ECG beats.
The work of Shaker et al. [23], proposes a new data augmentation technique using Generative Adversary Networks (GANs), to balance the classes of the MIT-BIH Arrythmia Dataset. They use two deep learning approaches: a) An end-to-end approach and a two-stage hierarchical approach, b) Based on CNNs. The results show that, with the increase in the data generated through the proposed technique, it is possible to observe an effective improvement in the performance of ECG classification in relation to the classification of the original dataset.
The article of Gao et al. [24], implements a Long Short Term Memory (LSTM) neural network to use the timing features in ECG signals, Focal Loss (FL) is used to resolve the imbalance of the MIT-BIH arrhythmia database. The results show that the LSTM network with FL obtains an accuracy of 99.26%.
In Kumar et al. [25], it presents a method for ECG classification, which employs a generalized signal pre-processing technique and uses a Multi-layered Perceptron network for the arrhythmia classification task, according to the AAMI EC57 standard. The method is trained and evaluated using PhysioNet MIT-BIH data set, obtaining an average accuracy of 98.72%.
Dataset Description and Analysis
The data collected consists of samples of ECG signals that had a recording time of 1400 milliseconds. The number of samples is divided into 5 classes, namely: class 0 with 90,589, class 1 with 2,779, class 2 with 7,236, class 3 with 803 and class 4 with 8,039 samples that can be observed in figure 1. Given the numerical representation of the quantity of each record of the above classes, we have: • Class 0: -Normal beat: That would be a sign of ECG from a person with their heart health under normal and healthy conditions. • Class 1: -Supraventricular premature beat: It represent premature activation of the atria from a different site from the sinus node and may originate from the atria or atrioventricular node (called premature junctional beats) [27]. • Class 2: -Premature ventricular contraction: It is an event where the heartbeat is initiated by the Purkinje fibers in the ventricles rather than the sinoatrial node, the normal heartbeat initiator [28]. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 • Class 3: -Fusion of ventricular and normal beat: A fusion beat occurs when electrical impulses from different sources act on the same region of the heart at the same time [29]. • Class 4: -Unclassifiable beat: Beats that have not been associated with any rating.
Each class presented consists of signals of different types of spectra, ie different graphical representation of the intensity distribution of a wavelength from a given signal. In figure 2, each class of ECG signal is represented, containing ten samples for each label. In signal analysis, there are some ways that can alter and change how the signal is viewed, aiding in observations and interpretations. One of the ways that is widely used is the signal spectrum, which is the representation of the components in a graph that shows their amplitude versus frequency.
Where ω k is the frequency of the sample and t n is the instant of time in seconds, being n ≥ 0. X(ω k ) is the spectrum of x at frequency ω k and x(t n ) is defined as the input of the signal amplitude over time t n .
Methodology
This section presents the methods used for mobile application development, and describes the metrics used to evaluate the performance of the implemented neural network. The diagram illustrated in Figure 3 shows the main constituent parts of the method. It consists of four main modules: a) CNN training; b) Model quantization using TensorFlow Lite [30], which aims to optimize the model for mobile application and d) Development of Android application for final classification of ECG signals.
Neural network training and architecture
According to [31], to train a machine learning model, it is necessary to divide the data into two sets (training and testing). The training dataset is the data sample used to fit the model, where the model sees and learns from this data [32]. The test data set, however, is the data sample used to provide an unbiased evaluation of the model in the training data set after adjusting the model hyperparameters [32].
To perform the neural network training, the data set was divided into training and testing, with 70% of each data class used for training and 30% used for testing. furthermore, the amount of samples for training was 76,614 while for testing is 32,834 samples.
As mentioned in section 3, not all signs are the same length, to treat this problem, samples are cut, reduced or filled with zeros, if necessary, for a fixed dimension of 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 187 values. After this pre-processing, all signals go through a transformation section with the DFT. Thus, input the neural network are vectors of a dimension with a length of 187, after they have passed through a DFT. Figure 4 illustrates the proposed network architecture for the arrhythmia classification task. This architecture is based on the architecture proposed by [12]. All convolution layers apply 1D convolution and each has 32 kernels of size 5. Max pooling with size 5 and strides 2 are also used on all pooling layers. The predictor network consists of five residual blocks followed by three fully connected layers with 512, 256, 5 neurons respectively and one softmax layer to predict the output class. The loss function is used with Cross Entropy, to map the network output the Softmax probabilistic function is used.
In order to train the model, the Adam Optimizer is used [33], with a learning rate of 0.0001, the Batch Size of 200 samples, the number of Epochs used is 100 and the Training Time of the network was 13 minutes.
For the implementation of the neural network, the computer library TensorFlow is used [30]. Processing was performed using a Geforce GTX 1060 graphics card with 1280 CUDA cores (processors), 6 GB of dedicated memory, 12 GB of RAM and a fourth generation Core i5 processor.
Quantized Neural Networks
Quantized Neural Networks use low precision weights and activations. These networks are trained from zero to arbitrary fixed point accuracy. Where in precision, QNNs that use fewer bits require deeper and wider network architectures than networks that use more precise operators, while requiring less complex arithmetic and fewer bits by weight [34].
A method has been introduced to train extremely low precision QNNs with weights and activations (eg 1 bit) at run time. In train-time, quantized weights and activations are used to calculate parameter gradients. During the next steps, QNNs dramatically reduce memory size and access by replacing most arithmetic operations with bit-by-bit operations [35]. This is to say that the quantization scheme is an integer mapping q to real numbers r, that is, as follows: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 This scheme consists in the multiplication of two square arrays N x N of real numbers, r 1 , e r 2 , with its product represented by r 3 = r 1 r 2 . We denote the entries of each of these matrices r α (α = 1, 2, or3) as r (i,j) α for i < i, j < N , and the quantization parameters with which they are quantified as (S α , Z α ). We denote the inputs quantized by q (i,j) α . Then, equation 2 becomes: From the definition of matrix multiplication we have: which can be rewritten as: For this application post-training quantization [30] was performed, thereby reducing the model size while improving CPU latency, with little degradation in [30] model accuracy. These techniques can be performed on a trained TensorFlow model and applied during TFL conversion.
There are several post-training quantization options [30], as can be seen from Table 1, the methods chosen and implemented are Weights Hybrid quantization and Full Integer quantization of weights and activations. Weight quantization is the simplest method of Post Training quantization, only floating point weights are quantized to 8 bits precision (also called "hybrid" quantization) [30]. This technique is performed with the model to TFL converter. The full integer quantization of weights and activations results in a fully quantized model, but the model still uses floating point input and output [30].
Mobile application development
The method chosen by this paper uses the TFL Java API [30], in figure 5, where you can develop applications for Android and IOS. TFL is TensorFlow's solution with lightweight mobile and embedded models. The application was developed on the Android platform, aiming at the classification of ECG signals. Data entry consists of loading a text file containing the signal to be analyzed and, after the signal is loaded, it will pass through the neural 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 network where it will be classified. The output will be given, making available its class in which it was classified by the network with its accuracy. The application interface can be observed according to the figures: 6, 7.
Metrics of the evaluation
The final accuracy of the model is estimated by the equation, where Ac f is the sum of the differences between the actual value y i and the expected valueŷ i with this it is possible to infer the generalization of the network.
As a statistical tool, we have the confusion matrix that provides the basis for describing classification accuracy and characterizing errors, helping to refine accuracy. The confusion matrix is formed by an array of squares of numbers arranged in rows and columns that express the number of sample units of a given category, inferred by a decision rule, compared to the current category [36].
The measurements derived from the confusion matrix are: total precision, which was chosen by the present work, individual class precision, producer precision, user precision, Kappa index, among others.
The total precision is calculated by dividing the sum of the main diagonal of the error matrix, that is, the total sum of the correct predictions, by the total number of samples collected.
As a statistical tool to evaluate the performance of the model, precision and recall are also used. Where precision is given by dividing the numbers of true positives, by the sum of true positives and false positives. The recall is calculated by dividing the true positives, by the sum of the true positives and false negatives.
F1 Score is a simple metric that takes Precision and Recall into account. This is simply the harmonic medium of precision and recall [37].
Results and Discussion
This section discusses and presents the results obtained at each stage of the development of this article. In Table 2, a comparison is made with the works Kachuee et al. [12], Acharya et al. [39]. It is noteworthy that the performance of each quantization method used is demonstrated, together with the result of the developed mobile application. Figure 8, is the confusion matrix before the quantization section and the training progress can be analyzed in figure 9. ECG classification results are compared with three other studies using the same MIT-BIH dataset, Table 2. In [12], it is proposed a CNN method for heart rate classification, this study had an average accuracy of 93.4% in the arrhythmia classi -1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 fication. In addition, a method for transferring the acquired knowledge to the task of classifying myocardial infarction is suggested.
In reference [39], a 9-layer CNN is developed to automatically identify 5 different categories of ECG heartbeat with an accuracy of 93.5%. In Table 2, it is possible to see that the performance of the method proposed by the present work surpasses the accuracy of classification of Kachuee et al. [12] and Acharya et al. [39]. It is also noteworthy that both comparative works do not implement a final application for the proposed CNNs. Already this article develops a mobile application to assist the diagnosis of different arrhythmias. On the figure 8, training progress is shown, as it is possible to analyze network performance over each season. It is also noted that test accuracy is similar to training accuracy.
As mentioned in section 4.2, two methods of quantization were implemented: Hybrid quantization of Weights and the Integer quantization of Weights and acti -1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 vations. The size of the model before the quantization section was 10.2 MB, after the quantization the model started to have the size 3.4 MB with the Hybrid quantization, for the Integer total quantization the model stayed with the size of 862.0 KB. With this it is possible to notice the reduction in the model size after quantization. This significant reduction in the size of the model is crucial for the development of the proposed mobile application, as it also allows a reduction in the computational cost required for the application to work on a mobile device.
On the Table 4, the accuracy of the neural network after quantization is displayed, so it is possible to analyze the accuracy of each quantization method. To evaluate the efficiency of the model, it is tested with different samples quantities, it can be noticed that the variation is small in both the quantization methods. Moreover, these preliminary results made it possible to think of an efficient Android application, with a simple and intuitive user interface, capable of performing the classification 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 Table 4 Accuracy after quantization with different sample quantities, using the two quantization methods (Hybrid and Integer). of 5 different arrhythmias. Our goal is to facilitate the use, mobility and low cost classification of different arrhythmias. As can be seen from figure 1, the dataset is unbalanced with more than 60% of signal samples belonging to class 0 (Normal beat). To analyze the model efficiency against the classification of each class, the model evaluation metrics for each class are used. In Tables 3, 5, it is possible to analyze the results of the metrics for each class both Hybrid and Integer quantization methods, as it can be observed that the model performed well even with the unbalanced dataset.
Quantity of samples Accuracy
According to the Figures 6, 7, can be observed that the interface and home screen of the application. In the beginning, it has a graphic where will be visualized the signal, which was loaded by the user. The x-axis of the graph represents the time at which the analyzed signal was recorded. Y axis, in turn, consists of the amplitude of the signal. The application has a simple and intuitive interface which works as follows: to load a signal, the user must click on the LOAD button, which will be directed to the Android file explorer, to select the desired signal and that it should be in csv text format. With the signal loaded, the application is ready for classification, which will be given after clicking the CLASSIFY button. The result will be displayed below the signal graph, which will have the class to which the signal was classified and below will have the prediction percentage.
In the execution of the application, some steps occur in which it is of great importance for the progress of the same. Thus, for comparison purposes, the two methods of quantization of the neural network were used. Where the first test is Hybrid Quantization, which had a runtime for application loading of 38.49 ms, using 26.7 MB RAM and 12% CPU, for classification, got a time response time 36.46 ms. In the second method consisting of Total Quantization, a time of 7.97 ms for application loading was achieved using 17.6 MB ram and 9.5% CPU. The smartphone used for testing was the Zenfone 4, which has a 4x 2.2 GHz + 4x 1.8 GHz processor and 4 GB of RAM.
Conclusions
In this article, we proposed a novel mobile application capable of classifying 5 different types of heartbeat, where the first class is normal and the others arrhythmias. This method yielded quite satisfactory results, with more than 98% accuracy. A comparison was also made with two other works in the literature, where the present article obtained superior results in relation to them. | 5,811.4 | 2020-07-16T00:00:00.000 | [
"Computer Science",
"Medicine"
] |
The Permutation Test as an Ancillary Procedure for Comparing Zero-Inflated Continuous Distributions
Empirical estimates of power and Type I error can be misleading if a statistical test does not perform at the stated rejection level under the null hypothesis. We employed the permutation test to control the empirical type I errors for zero-inflated exponential distributions. The simulation results indicated that the permutation test can be used effectively to control the type I errors near the nominal level even when the sample sizes are small based on four statistical tests. Our results attest to the permutation test being a valuable adjunct to the current statistical methods for comparing distributions with underlying zero-inflated data structures.
Introduction
Statistical analysts sometimes encounter data that have an excessive number of zeros and these data often present analytical difficulties because traditional methods rely on assumptions that may be unrealistic and plausible transformations may not be found.Many studies have reported on statistical methods for analyzing count data with excessive zeros [1][2][3][4][5][6].Some zero inflated data may be viewed as having a mixed distribution where zeros have a point distribution and the distribution of non-zero observations is positive and continuous.This distribution has not been investigated adequately and statistical methods with favorable Type I and Type II errors for comparing these non-traditional distributions are desired.
Testing equivalence of zero-inflated populations in the context of underlying mixed distributions is equivalent to testing equality of the probabilities of zeros and simultaneously equality of the parameters of the non-zero observations [7].The likelihood ratio (LR) [8] and Wald [9] tests are two widely used methods.These two methods typically perform well if the probability density function that applies under the null hypothesis is known.Recently, Monte Carlo simulations were employed to compare several approaches including the LR, Wald, central limit theorem (CLT), modified central limit theorem (MCLT) tests with respect to their empirical Type I errors and testing powers for three zero-inflated continuous distributions [7].The LR, Wald, and MCLT tests were found to be preferable to the tests based on central limit theory.
There are two important issues when several populations with zero-inflated data structure are compared.First, the underlying distribution is usually unknown and, therefore, the assumptions of specific distributions can be easily violated by using assumption-constrained methods.Second, empirical Type I errors and testing powers are difficult to determine because the relevant parameters are almost always unknown even if the assumed distribution is correct.Moreover, a small sample size may contribute to higher Type I and Type II errors.Thus, a test that controls the empirical Type I errors and yields valid estimates of testing powers is helpful.
Permutation tests are advocated for data analysis when assumptions required to validate parametric procedures are violated [10][11][12][13].Unlike parametric tests, permutation tests can generate probabilities by repeatedly "resampling" the data and evaluating the obtained results with reference to an empirically derived distribution [14,15].Permutation tests have two major advantages: 1) they can be used to adjust the empirical Type I errors and the testing powers and, 2) they can be used when some assumptions required to justify parametric tests are violated.
Hence, their use may lead to more appropriate statistical conclusions.
The purpose of this study was to investigate the issues raised above pertaining to the use of ancillary permutation tests to compare several populations when the random variable of interest has either a known or unknown zero-inflated continuous distribution.Four statistical tests were compared with respect to both their empirical type I errors and testing powers.First we assumed the data followed a zero-inflated exponential distribution as reported by Zhang et al. [7].Empirical Type I errors and testing powers for these tests were compared with and without adjunct permutation tests by empirical estimates obtained using Monte Carlo simulations.Section 2 describes a general permutation test that generates an empirical probability for each test.Simulated results for four carefully selected parameter configurations are presented in Section 3. Finally, Section 4 demonstrates the results with the permutation test for a data set reported by Koopmans [16].
Four Testing Methods
Performances of four tests including the likelihood ratio (LR) [8], Wald [9], central limit theorem (CLT), and modified central limit theorem (MCLT) tests [7] were evaluated.The CLT test considers only the population means calculated over all zero and non-zero observations while the MCLT test considers both the probability of zeros and simultaneously the mean of non-zero observations.The first two tests are distribution-based while the other two are distribution-free based.Maximum likelihood (ML) estimators [17] are required for both the LR and Wald tests.For the CLT and MCLT tests, the Wald test was incorporated to derive the probability for each test [7].These methods were detailed in one of our previous papers [7] and were not repeated in this study.
Permutation Test
The procedures of using the permutation test in zeroinflated data are: Step 1: Calculate the p-value using each of the above mentioned four tests (e.g.LR) to analyze the original data; Step 2: Reshuffle the original data and randomly assign the data to different populations without replacement; Step 3: Calculate the p-values by the same method used in Step 1 for the reshuffled data obtained in Step 2; Step 4: Repeat Steps 2 and 3 "N times"; Step 5: Construct the sampling distribution of p-values obtained in Steps 2 through 4; Step 6: Locate the p-value in this distribution that corresponds to each p-value calculated in Step 1.If the pvalue from the original data is in the main body of the distribution (α/2 to (1 − α/2)), then there is no significant difference at probability level α among populations.Otherwise, there is evidence that the difference between (among) populations is significant.
The above procedures from Steps 1 to 6 were applied to all four tests in this study.
Simulation Procedure
In our empirical investigation we assumed interest was in testing the hypothesis that three zero-inflated distributions had identical means.We simulated data from three zero-inflated distributions with sample sizes ranging from 25 to 300 and performed each of four tests repeatedly using the replicate samples to test the null hypothesis.We tabulated the number of rejections of the hypothesis under each known scenario to estimate Type I errors and powers.Twelve sample sizes (n = 25 × s, where s = 1, 2, •••, 12) were considered and the nominal probability level was set at 0.05 throughout.Although different configurations were considered only one was listed for the null distributions and three for alternative distributions as described in Table 1.The first configuration in Table 1 was designed to estimate the empirical Type I errors and the remaining three configurations were designed to estimate the empirical testing powers.Each set of simulated data was analyzed by the four tests with and without employment of the permutation test.Repetitions of 1000 simulated samples were used for each case.All simulations were conducted by a C++ program written by the authors of this paper.
Simulation Results
First the number of permutations sufficient for statistical tests at a given probability level is determined.The Type I errors and testing powers from 100 to 2000 different permutations for configurations 1 and 2 with sample size 200 are summarized in Figures 1 and 2, respectively.These figures clearly demonstrate that both empirical Type I errors and testing powers became reasonably stable after the sample size surpassed 100 permutations.Results from additional simulations for various different sample sizes and configurations showed similar trends.Thus samples of 500 permutations were chosen for all the remaining simulations.
The empirical Type I errors of the four tests with and without permutation tests are summarized in Table 1 for the case of a zero-inflated exponential distribution.The differences between observed Type I errors and the nominal 0.05 level tend to be smaller as the sample size increases for all four tests without permutation tests, indicating that all these tests tend to perform better as the sample size increases.However, with the permutation tests, the empirical Type I errors are close to the nominal 0.05 level for different methods and various sample sizes including small sample sizes (Table 2).The results indicate that the permutation tests can reduce the high Type I errors that are prevalent with small sample sizes.When the sample sizes are large, i.e., at least 100, the empirical Type I errors for the four statistical methods are almost identical irrespective of using the permutation tests.Tables 3-5 present the empirical powers of the four tests for three parameter configurations as defined in Table 1.As expected, the testing power increased for all four tests as the sample size increased.The testing powers obtained without permutation tests were typically lower than those obtained with permutation tests for all methods when the sample size is small (100 and below).However, as the sample size increases, the testing powers were similar irrespective of using permutation tests.As for parameter configuration 2 described in Table 1, the CLT test and the other three tests have similar testing powers because only means for the non-zero observations contributed the differences (Table 3).As for designs 3 and 4, the CLT test has an extremely low testing power compared with other three tests (Tables 4 and 5).The increase or decrease of both zero probability level and the non-zero mean made the differences among populations hard to detect with the CLT method, while the other three tests are sensitive and maintain desirable testing powers.This indicates that the LR, Wald, and the MCLT tests are better than the CLT test in general.When the zero probability levels among populations are similar, the CLT test is still a good option.
In many situations, the distribution for a given zeroinflated data set is unknown.It will be interesting to reveal the empirical Type I errors and testing powers obtained using these methods by assuming the following distributions.In this study, we generated 1000 simulated data sets based on different parameter configurations as described in Table 1 with the zero-inflated exponential distribution.Then the LR and Wald methods were applied to test the differences among three populations by assuming the data follow zero-inflated gamma and lognormal distributions.Although simulations for various sample sizes were conducted only the results for configurations 1 and 2 with sample size of 200 were reported (Table 1) because the similar patterns were observed for different configurations with different sample sizes (data not shown).Given zero-inflated exponential data, both the LR and Wald tests resulted in unfavorably high Type I errors if no permutation tests were applied; however, these type I errors were adjusted substantially to be close to the nominal level on using the permutation test.On the other hand, the testing powers obtained by the LR and Wald tests were lower when the lognormal distribution was assumed.For the gamma distribution, both the LR and Wald tests have similar and desirable testing powers when the permutation tests are applied (Table 6).The results suggested that the tests could have caused either higher Type I errors or lower testing powers when an inappropriate distribution was assumed.However, with the permutation tests, the chance to make Type I errors can be greatly decreased, yet the testing powers can be desirable in many cases.
Application
Koopmans [16] reported results of a study of seasonal activity patterns of field mice.Data consisted of the average distances traveled between captures by field mice at least twice in a given month.The distances were rounded to the nearest meter.A large number of zero distances were observed in addition to non-zero distances resulting in data with a zero-inflated distribution.The exact distribution of the non-zero observations is unknown.Various LR tests were used to identify which parameter(s) were associated with the seasonal differences by assuming the data followed a mixture of zero-inflated logsnormal distributions [18].To illustrate our approach, we analyzed the data by four tests alone and by the permutation test with 1000 repetitions assuming the underlying distribution was a zero-inflated exponential distribution ( employment of the permutation tests, indicated that the mice distances differed significantly among the three seasons.
Discussion
It is desired that a statistical method sustains a preset nominal Type I error and a high testing power.Many methods are based on the appropriate statistical assumptions and require a large sample size.In some situations, the sample size may be very small and test statistics may yield unfavorable Type I errors and testing powers.In addition, the real distribution is often unknown so desirable testing properties cannot be expected on employing distribution-based tests.In this study, we investigated statistical properties of the permutation tests integrated with four distribution-based tests to compare populations with zero-inflated data structures.Based on the results from the simulated zero-inflated exponential data, several conclusions can be made on use of the permutation test: 1) high Type I error caused by the appropriate statistical tests without the permutation test for small sample sizes can be adjusted to the preset nominal level when the permutation test is used; 2) high Type I errors caused by the inappropriate assumptions can be adjusted to the preset nominal level; and 3) for a large data set, both the type I errors and testing powers are similar regardless the use of the permutation test for appropriate distribution assumptions.The same conclusions applied for the other two types of zero-inflated continuous distributions including gamma and lognormal distributions (results not shown).
As reported by Zhang et al. [7] and in results of this study, the LR and Wald tests hold similar type I errors and testing powers but they are distribution dependent.If an inappropriate distribution is assumed, both inflated Type I errors and low testing powers can occur (Table 6).The CLT test is data structure dependent because it considers only the population mean including zeros.When the population means are similar (because the populations have similar probabilities of observations equal to zero) and their non-zero observations have similar distributions, then the CLT test may have statistical properties similar to the other three tests.The MCLT test considers two parameters: the zero probability and non-zero mean and thus is better than the CLT test and robust for most cases.In addition, high Type I errors caused by the MCLT test can be adjusted by the permutation test for small sample size.Therefore, the MCLT test can be recommended for general use regardless whether the data distribution is known or unknown.Numerical investigation on other types of distributions should help gain more information regarding the MCLT method.
Even though the permutation test showed several major advantages, the LR and Wald test still sustain desirable Type I errors and testing powers and are not as computationally intensive when the distribution for a large data set is known or the assumed distributions are appropriate.Nevertheless, the permutation test could be a valuable addition to the current statistical tests especially when a data set is small or the distribution is unknown.
for null hypothesis and 2 to 4 are alternative hypotheses.† : δ j and β j are zero probability level and mean of exponential distribution for j th population.
Figure 1 .
Figure 1.Empirical type I errors obtained by 20 different numbers of permutations.(LR = likelihood ratio, CLT = central limit theorem, and MCLT = modified central limit theorem).
Table 2 . Empirical Type I errors for zero-inflated exponential distribution based on 1000 simulations.
† : LR = likelihood ratio, CLT = central limit theorem, and MCLT = modified central limit theorem; ‡ : 500 permutations were used.
Table 3 . Empirical testing power for zero-inflated exponential distribution based on 1000 simulations for configuration 2.
‡ : 500 permutations were used.
Table 7 )
. The results for all four tests, with and without
Table 6 . Empirical type I errors and testing powers esti- mated by the LR and Wald tests by assuming three differ- ent distributions (exponential, Exp, Gamma, and log nor- mal, LogN) as zero-inflated exponential distribution with and without permutation tests based on 1000 simulations for population size 200.
‡ : Based on design 1 and design 2 in Table1, respectively; ∆ : 500 permutations were used. | 3,677.8 | 2012-06-27T00:00:00.000 | [
"Mathematics"
] |
The Application of Baum-Welch Algorithm in Multistep Attack
The biggest difficulty of hidden Markov model applied to multistep attack is the determination of observations. Now the research of the determination of observations is still lacking, and it shows a certain degree of subjectivity. In this regard, we integrate the attack intentions and hidden Markov model (HMM) and support a method to forecasting multistep attack based on hidden Markov model. Firstly, we train the existing hidden Markov model(s) by the Baum-Welch algorithm of HMM. Then we recognize the alert belonging to attack scenarios with the Forward algorithm of HMM. Finally, we forecast the next possible attack sequence with the Viterbi algorithm of HMM. The results of simulation experiments show that the hidden Markov models which have been trained are better than the untrained in recognition and prediction.
Introduction
Currently, the network security situation is increasingly sophisticated and the multistep network attack has become the mainstream of network attack. 2012 Chinese Internet network security reports released by the National Computer Network Emergency Response Technical Team Coordination Center of China (CNCERT/CC) show that the two typical multistep attacks: warms and distributed denial of service (DDOS) [1] account for 60% of overall network attacks. Multistep attack [2] means that the attacks apply multiple attack steps to attack the security holes of the target itself and achieve the devastating blow to the target. There are three features of attack steps of multistep attack. (1) In the multistep attack, there is a casual relationship between multiple attack steps. (2) The attack steps of multistep attack have the property of time sequence [3]. (3) The attack steps of multistep attack have the characteristics of uncertainty [4].
Multistep attack is one of the main forms of network attack behaviors, recognizing and predicting multistep attack that laid the foundation of active defense, which is still one of the hot spots nowadays. Literature (application of hidden Markov models to detect multistep network attacks) proposed a method to recognize multistep attack based on hidden Markov model.
Markov model literature (improving the quality of alerts and predicting intruder's next goal with hidden colored Petrinet) introduced the concept of attack "observation, " but both stayed in the specific attack behaviors, which have some limitations. Current research on the approaches to forecast multistep attack behaviors mainly includes four types: (1) the approach to forecasting multistep attack based on the antecedents and consequences of the attack [5]. It applies the precursor subsequent relationship of the event, to forecast the attacker wants to implement attacks in the near future. Because of the complexity and the diversity of the attack behaviors, this approach is difficult to achieve. (2) The approach to forecasting multistep attack based on hierarchical colored Petri-nets (HCPN) applies the raw alerts by Petrinets and considers that the attack intention is inferred by raw alerts [4]. But this approach focuses on the intrusion detection of multistep attack behaviors. (3) The approach to forecasting multistep attack based on Bayes game theory could forecast the probability that the attackers choose to 2 The Scientific World Journal attack and the probability that the defenders choose to defend in the next stage rationally [6,7]. However, in current study, only two-person game model is established, so this approach has some limitations. (4) The approach to forecasting multistep attack based on attack intention [3,8]
Hidden Markov Model
Hidden Markov model was first proposed by Baum and Petrie in 1966. It is a statistical model, which is used to describe a Markov process which contains a hidden parameter [9]. The research object of this model is a data sequence; each value of this data sequence is called an observation. Hidden Markov model assumes that there still exists another sequence which hides behind this data sequence; the other sequence consists of a series of states. Each observation occurs in a state, the state cannot be observed directly, and the features of the state can only be inferred from the observations.
A complete hidden Markov model (HMM) is usually represented by a triple = ( , , ), which includes the following five elements: (1) a finite state, which is represented by the set , where = { 1 , 2 , . . . , } and, at time , the state is denoted by ; (2) the set of observations, which is represented by the set , where = { 1 , 2 , . . . , }; The model of recognizing and forecasting multistep attack based on hidden Markov model is shown in Figure 1.
There are three problems which can be solved by hidden Markov model well. Step 2. Iterative calculation. for = 1, 2, . . ., Step 3. Termination. We can obtain the parameters of hidden Markov model. Correspondence between the problems and algorithms of hidden Markov model are shown in Figure 2.
Hidden Markov model is usually used to deal with the problems related to the time sequence and it has been widely used in speech recognition, signal processing, bioinformation, and other fields. Based on the characteristics of the attack steps of hidden Markov model and the problems that hidden Markov model can be solved, we apply the hidden Markov model to the field of recognizing and forecasting multistep attack. Firstly, the improved Baum-Welch algorithm is used to train the hidden Markov model , and we get a new hidden Markov model . Then we recognize the alert belonging to attack scenarios with the Forward algorithm of hidden Markov model. Finally, we forecast the next possible attack sequence with the Viterbi algorithm of hidden Markov model.
The Approach to Recognizing and Forecasting Multistep Attack
The steps of the approach to recognizing and forecasting multistep attack are as follows.
Step 1. Obtain the initial state matrix (old), state transition matrix (old), and observation matrix (old) of HMM ( ). Step 2. Use the improved Baum-Welch algorithm to train the initial state matrix (old) and observation matrix (old), and we get an initial state matrix (new), observation matrix (new), and a new HMM ( ).
Step 3. Recognize the alert belonging to attack scenarios with the Forward algorithm.
Step 4. Forecast the next possible attack sequence with the Viterbi algorithm.
The flow chart is shown in Figure 3.
The Introduction of Baum-Welch Algorithm.
If we want to apply the hidden Markov model to the multistep attack, the biggest problem is to determine the observations of HMM. A better parameter can improve the efficiency of 0.000 0.000 0.000 0.000 0.998 0.001 0.001 Recognizing multistep attack is mainly based on the alert sequence. First, we calculate the probability of alert sequence generated by the given HMM(s). Then we decide that the attack which has the maximum is likely to be the ongoing attack. The structure of recognizing multistep attack with Forward algorithm is shown in Figure 4.
Viterbi Algorithm. The pseudocode of Viterbi algorithm is as in Algorithm 3.
Predicting the behavior of multistep attack is mainly to determine the intentions that the attackers have been completed and forecast the next possible attack intentions. The structure of forecasting multistep attack with Viterbi algorithm is shown in Figure 5
Baum-Welch Algorithm: Train the Given HMM(s).
Based on the literature (approach to forecast multistep attack based on fuzzy hidden Markov model), we can obtain the initial state matrix, state transition matrix, and observation of DDoS HMM, as is shown from Tables 1, 2, and 3. The data set which is used in the simulation experiment is an attack scenario testing data set LLDOS1.0 (inside) provided by DARPA (Defense Advanced Research Projects Agency) in 2000. We extract two kinds of multistep attack from it; they are DDoS multistep attack and FTP Bounce multistep attack. While the calculation of the state transition matrix is completely the statistical calculations on data, we only train the initial state matrix and observation matrix of HMM. We can see that there are a large number of zeros in observation matrix clearly and the observation matrix is the sparse matrix. So we train the matrix(s) by block. We suppose that the number of observation sequences is S and the length of S is 32, where S multiplied by 32 equals the number of training data. And there is no corresponding sequence of state. In this regard, we can obtain the initial state matrix (new) and the observation matrix (new) of the DDoS HMM ( ), as is shown in Tables 4 and 5 Tables 6 and 7, respectively.
When the alerts "Alert 1 " and "Alert 3 " were received, according to the Forward algorithm of hidden Markov model, We can see from the above results, p(alerts | DDoS HMM) > p(alerts | FTP Bounce). That is to say, the ongoing multistep attack behavior is likely to be DDoS HMM.
Viterbi Algorithm: Forecast the Next Possible Attack
Sequence. When the alert sequence {Alert 1 , Alert 3 , Alert 7 , Alert 8 , Alert 10 } was received by the console, we can obtain the completed intent sequence {State 1 , State 2 , State 3 , State 4 }. That is to say, now completed intentions are the previous four attack intentions; the next intention will be state 5 .
Comparison of Results.
We compare the results between the untrained HMM(s) and the trained HMM(s) by Baum-Welch algorithm; the comparison of results are shown in Table 8.
Conclusion
The biggest difficulty of hidden Markov model applied in multistep attack is the determination of observations. Now the research of the determination of observations is still lacking, and it shows a certain degree of subjectivity. In this regard, we train the existing hidden Markov model(s) by the Baum-Welch algorithm of HMM based on several groups of observation sequence. And we can obtain a new hidden Markov model which is more objectively. Simulation experiments results show that the hidden Markov models which have been trained are better than the untrained in recognition and prediction. | 2,315.6 | 2014-05-28T00:00:00.000 | [
"Computer Science"
] |
Towards precise PICO extraction from abstracts of randomized controlled trials using a section-specific learning approach
Abstract Motivation Automated extraction of population, intervention, comparison/control, and outcome (PICO) from the randomized controlled trial (RCT) abstracts is important for evidence synthesis. Previous studies have demonstrated the feasibility of applying natural language processing (NLP) for PICO extraction. However, the performance is not optimal due to the complexity of PICO information in RCT abstracts and the challenges involved in their annotation. Results We propose a two-step NLP pipeline to extract PICO elements from RCT abstracts: (i) sentence classification using a prompt-based learning model and (ii) PICO extraction using a named entity recognition (NER) model. First, the sentences in abstracts were categorized into four sections namely background, methods, results, and conclusions. Next, the NER model was applied to extract the PICO elements from the sentences within the title and methods sections that include >96% of PICO information. We evaluated our proposed NLP pipeline on three datasets, the EBM-NLPmod dataset, a randomly selected and re-annotated dataset of 500 RCT abstracts from the EBM-NLP corpus, a dataset of 150 Coronavirus Disease 2019 (COVID-19) RCT abstracts, and a dataset of 150 Alzheimer’s disease (AD) RCT abstracts. The end-to-end evaluation reveals that our proposed approach achieved an overall micro F1 score of 0.833 on the EBM-NLPmod dataset, 0.928 on the COVID-19 dataset, and 0.899 on the AD dataset when measured at the token-level and an overall micro F1 score of 0.712 on EBM-NLPmod dataset, 0.850 on the COVID-19 dataset, and 0.805 on the AD dataset when measured at the entity-level. Availability and implementation Our codes and datasets are publicly available at https://github.com/BIDS-Xu-Lab/section_specific_annotation_of_PICO.
Introduction
Healthcare providers rely on the evidence from published biomedical literature for assessing the effectiveness of a new treatment or intervention for diseases.With the evidence and insights gained from well-designed clinical studies such as randomized control trials (RCTs) in PubMed (Frost et al. 2020, Abani et al. 2021, Emani et al. 2021), it is possible to explore the best treatment options for diseases such as Coronavirus Disease 2019 (COVID-19) (Emani et al. 2021).However, the number of RCTs that keep growing every day has become a challenge for clinicians to keep themselves up to date with the new evidence.In 2020, a total of 29 256 RCT abstracts (80 RCT abstracts per day) were added to PubMed.This increased to 29 983 RCT abstracts (82 RCT abstracts per day) in 2021 and 28 482 RCT abstracts (78 RCT abstracts per day) in 2022.These estimates show a continuously growing number of publications related to RCT and, it has become extremely difficult for clinicians to gain knowledge from these publications to provide the best care for their patients.Currently, more than 595 000 RCT abstracts are in PubMed.
In 1995, the PICO (P: Population, I: Intervention, C: Comparison/Control, O: Outcomes) framework (Richardson et al. 1995) was introduced to formulate a well-defined research question and facilitate the literature search for evidence-based medicine (EBM).Manual synthesis of PICO-based evidence from the published literature is both time-consuming and costly.Hence, automated approaches were developed for extracting PICO elements.Earlier works were based on rule-based approaches and machine learning (ML) approaches (e.g.Support Vector Machine, Random Forest, Conditional Random Field) and rely heavily on hand-crafted features (Huang et al. 2006, Demner-Fushman and Lin 2007, Boudin et al. 2010, Chabou and Iglewski 2015, 2018).Recently, deep learningbased (DL) approaches are applied to overcome the tedious feature engineering task and boost performance (Jin andSzolovits 2018b, Zhang et al. 2020).Both ML/DL approaches rely heavily on the annotated corpus to achieve high performance.The time and cost involved in the manual annotation by domain experts, the distribution of PICO elements within RCT abstracts, and the variations in PubMed abstracts (e.g.structured and unstructured) are the major reasons for the lack of a large, publicly available annotated corpus for PICO extraction.Nye et al. (2018) released the EBM-NLP corpus with 4993 RCT abstracts annotated (using crowdsourcing) with PICO elements at two levels of granularity.The first level is the named span-level where the phrases containing Population, Intervention, and Outcome (PIO) information are annotated.The second level further distinguishes the named span-level with more fine-grained labels (e.g.distinguish P according to gender, age, condition, etc.).Though the EBM-NLP corpus has revived the PICO extraction task, it has certain limitations.The overall named entity recognition (NER) model performance is not satisfying (a token-level F1 score of 68% on the span-level and 48% on the hierarchical labels).There is no entity-level evaluation or information retrieval evaluation at the abstract level.Based on our observation, probable reasons for low performance are: (i) Information regarding PICO elements is scattered throughout the abstract with interventions and outcomes often presented multiple times with different lexical variants.(ii) PICO elements in RCTs can be complex, making the annotation task complex.For example, all interventions may not refer to the clinical trial under consideration (e.g., mentions of interventions of prior clinical trials), and identifying distinct outcomes of a particular intervention in multi-arm trials is often difficult.(iii) PICO elements can vary widely across different diseases hence requiring large datasets to learn meaningful patterns.For example, intervention may be a pharmacologic substance (e.g.doxorubicin) for diseases such as cancer or a music-based therapy for certain neurological disorders.(iv) The EBM-NLP corpus utilized a crowdsourcing approach for the annotation process.The expert annotation is limited to only 200 abstracts.The remaining abstracts in the corpus were annotated by the annotators recruited by Amazon Mechanical Turk (AMT), who were not well-trained.The inter-annotator agreement (IAA) is low (0.50 for P, 0.59 for I, and 0.51 for O) even among the medical experts.The annotation schema is also complex, and it is difficult for the annotators to comprehend all the details.This results in annotation inconsistencies.
We hypothesized that selecting only the specific sections of an RCT abstract and title that contain relevant PICO information for annotation can significantly reduce the annotation complexity, effort, and the issues mentioned above.In the current study, we analyzed the distribution of PICO elements in different sections of RCT abstracts.Our findings suggested that the Title and Methods section covers most of the PICO elements.Consequently, we developed a two-step natural language processing (NLP) pipeline to identify and extract the PICO elements.To assess its generalizability, we further evaluated it on three different datasets namely EBM-NLP mod, dataset, Alzheimer's disease (AD) dataset, and COVID-19 dataset.The major contributions of this work are: 1) A novel two-step NLP pipeline that classifies the sentences from a PubMed abstract into different sections, backgrounds, methods, results, and conclusions and extracts the PICO elements.
2) EBM-NLP mod dataset derived from the EBM-NLP corpus.The dataset includes 500 randomly selected RCT abstracts.The PICO elements were re-annotated to overcome the limitations of the EBM-NLP corpus.Jin and Szolovits (2018a,b) treated PICO detection as a sequential sentence classification task rather than a single sentence classification task.They utilized neural network architectures such as long short-term memory (LSTM) (Jin and Szolovits 2018b) and bi-directional long short-term memory (Bi-LSTM) (Jin and Szolovits 2018a) to encode the contextual content from the preceding and succeeding sentences to improve the prediction of the current sentence.Recently, several deep learning approaches that utilize pre-trained language models such as SciBERT and BERT are applied to improve the performance of sentence classification (Cohan et
Recognition of PICO elements
NER is used to identify the PICO elements (Nguyen et al. 2017, Nye et al. 2018, Kang et al. 2019, 2021, Zhang et al. 2020, Dhrangadhariya et al. 2021, Liu et al. 2021b).Nye et al. (2018) presented two baseline models, the linear CRF model and the LSTM-CRF model for identifying PICO elements in the EBM-NLP corpus.A recent study showed improved performance on PICO extraction when the NER model was first pre-trained on the EBM-NLP corpus, and further fine-tuned with the additional data annotated by themselves (Kang et al. 2019).Zhang et al. (2020) proposed an approach by combining sentence classification, disease entity recognition, and disease mapping using various deep learning models (convolutional neural network, Bi-LSTM, etc.) for extracting P and O elements.To alleviate the reliance on time-consuming manual annotation by experts, a span detection approach for PICO extraction that uses only lowquality, crowd-sourced, sentence-level annotations as inputs, was proposed by Liu et al. (2021b).The authors applied a masked span prediction task in which input spans were replaced with predefined mask tokens and a pre-trained neural language model [BLUE (Peng et al. 2019)] was used to infer which spans contribute most to the PICO sentence classification results using the EBM-NLP corpus.A multi-task learning approach that learns and recognizes both coarse-grained descriptions (e.g.40 children aged 7-11 with autism spectrum disorder) and constituent finer semantic units (e.g., "40" shows "sample size", "7-11" shows "age" and "autism spectrum disorder" shows "condition") was explored by Dhrangadhariya et al. (2021).In that study, the EBM-NLP corpus was utilized as it provided a multi-level annotation: the span-level (level 1) annotation corresponds to the coarse-grained descriptions and other levels of annotation focus on specific semantic units.Recently, the Easy Data Augmentation (Wei and Zou 2019) technique incorporated with the Unified Medical Language System (UMLS) knowledge (including synonym replacement, random insertion, random swap, and random deletion) was evaluated on PICO extraction (Kang et al. 2021).
Materials and methods
The overview of our NLP pipeline is illustrated in Fig. 1.First, the sentences in both structured and unstructured RCT abstracts were classified into background, methods, results, and conclusions using our recent work on sentence classification (Hu et al. 2022).Next, P, I, C, and O elements were extracted using a NER model.
Distribution of PICO elements in RCT abstracts
We conducted a preliminary evaluation of the distribution of PICO elements across different sections of the RCT abstract.We randomly selected 30 RCT abstracts and reviewed them manually to identify the unique mentions of PICO elements in different sections of the abstracts.The purpose was to identify the specific sections with high coverage of PICO elements.We hypothesize that annotating only these sections may reduce the redundancy, ambiguity, and time and effort involved in the manual annotation.We also hypothesize that our proposed approach is liable only for a minimum loss of information.Titles of RCT abstracts usually include P and I elements.They provide a precise, and accurate description of the study, and are easy to annotate.Our analysis shows that the Methods section achieved the highest coverage of 95.2% among all the sections.The Title and Methods sections together achieve coverage of 96.7% (Table 1).The PICO elements in the results section are often duplicates of the methods section.Thus, we considered only the Title and Methods sections for annotation.The annotation can easily be extended to other sections with our sentence classification model.(Hu et al. 2022) shows high performance in predicting methods, results, and conclusion sections, the classifier has great difficulty distinguishing the background and objective sections.The difference between the sentences from the background and objective sections is less obvious when compared to other sections.We relabeled the sentences from the objective section as background and redefined the section labels as background, methods, results, and conclusions.
Evaluation data
We created three datasets to evaluate and report the performance of the sentence classification model.In our dataset splits, we have meticulously ensured that the test set used for the NER task does not overlap with the training data for sentence classification.Our approach eliminated potential data leakage and ensured an unbiased evaluation of both tasks.
Section-specific annotation
Annotating a high-quality dataset is a labor-intensive task and identifying the domain experts is challenging.Though the crowdsourcing approaches have shown some promising results on corpus generation, the IAA (e.g.EBM-NLP corpus) is relatively low.This results in a suboptimal performance of a NER model.
Initially, we followed the annotation guidelines from EBM-NLP to annotate the PICO elements.The IAA reported using Cohen's Kappa coefficient was only 0.3 (Supplementary Materials S1.2 for the equation for Cohen's Kappa coefficient).Further investigation revealed three major reasons for achieving a low kappa coefficient: (i) the original annotation guidelines from EBM-NLP are complex and complicated.It first necessitates annotating the Participant, Intervention, and Outcome elements at the span-level, and further annotates the specific details at a granular level.For instance, Participant is annotated at the span-level and the specific details about the participant (i.e.age, gender, and condition) are annotated at the granular level.Likewise, Intervention is annotated at the span-level and the specific details about the intervention (i.e.physical, non-physical, and Control) are annotated at the granular level.Note that Control is annotated as a subtype of Intervention.(ii) The original annotation guidelines lack specific rules for defining entity boundaries for the PICO elements.In addition to defining the PICO elements with examples of inclusion and exclusion, the guidelines only mention -"mark the longest contiguous text that includes such a description."While differences in span boundaries in annotations by different annotators are a major reason behind low IAA, our experience in annotating the documents for other studies has shown that the specific rules regarding modifiers, articles, prepositions, and overlapping entities improve the IAA.(iii) The repeated mentions of interventions and outcomes across different sections of the abstract lead to ambiguity and missing annotations, especially in complex interventions and multi-arm trials.For example, consider a study comparing a multicomponent community health program versus usual care on several health outcomes.As we move along different sections of the abstract, we may come across repeated mentions of these interventions including a detailed mention, a specific component mention, abbreviation of the program, and even some generic reference such as "the We resolve the issues observed in the original annotation guidelines, by: (i) annotating the PICO elements at a single level; (ii) enriching the annotation guidelines with a set of linguistic rules to define the boundaries (see Supplementary Materials S3), and (iii) retaining the sentences only from the Title and Methods sections (see Supplementary Materials S3).This is based on our preliminary experiments as shown in Table 1.Annotating the PICO elements mentioned in the Title and Methods sections took only 60 s per abstract.This is significantly lower than the time required to annotate the PICO elements mentioned in all the sections of a PubMed abstract (i.e.146 s).Switching from the hierarchical annotation to a single-level annotation was based on several limitations and challenges that we noticed in the original multi-level annotation.Firstly, the hierarchical-level annotation significantly increases the time, effort, and cost of the annotation process.Secondly, prior research using the EBM-NLP corpus has reported issues with the fine-grained classification of PICO elements utilized for the hierarchical annotation.One such issue was concerned with the "Intervention" element.According to Dhrangadhariya et al. (2021), even the human annotators find it difficult to classify certain interventions as education or psychological.Their experiments showed the least performance of 0.31 F1-score on the "Intervention" class.From the error analysis, the authors suspect that the ambiguities arising from the split of coarse-grained PICO into fine-grained PICO classes were one of the issues for such low performance.
Using the revised annotation guidelines, two annotators with medical background re-annotated the EBM-NLP mod dataset and annotated two additional datasets related to COVID-19 and AD.The IAA was calculated for each PICO element and the entire dataset using Cohen's Kappa coefficient.The approach achieved the Cohen's Kappa coefficient of 0.714, 0.808, 0.701, and 0.790 for the P, I, C, and O components, respectively, and 0.746 for all PICO elements.
Prompt-based learning for sentence classification
Recent works show the feasibility of using natural language prompts for tuning the pre-trained language models for specific downstream tasks (Petroni et al. 2019, Ding et al. 2021, Liu et al. 2021a).In our prior research (Hu et al. 2022), we applied prompt-based learning to classify sentences from RCT abstracts.In brief, our approach classifies sentences by predicting the mask position in RCT abstracts using promptbased learning.Other existing approaches use traditional machine learning and deep learning to classify sentences.The performance of our sentence classification approach surpasses the performance of the previous state-of-the-art approach using Hierarchical Sequential Labeling Network (HSLN) (Jin and Szolovits 2018a).A more detailed description of our method is in Supplementary Materials S1.1.We applied the model from our previous work to classify the sentences in the RCT abstracts from the EBM-NLP mod dataset, COVID-19 dataset, and AD dataset.The parameters used in our promptbased learning approach were set as follows: dropout ¼ 0.5, batch size ¼ 8, learning rate ¼ 6eÀ6, optimizer ¼ AdamW, and learning rate decay ¼ 0.01.We evaluated our prompt-learning model and compared our model with that of the HSLN architecture on the EBM-NLP mod dataset, COVID-19 dataset, and AD dataset independently.We used the standard evaluation metrics, precision (P), Recall (R), and F1 scores.
NER for PICO extraction
A recent work on identifying P, I, and O elements using LSTM-CRF within BERT achieved a 0.68 F1-score on the EBM-NLP corpus when evaluated at the token-level.Another work by Gu et al. (2022) used the pre-trained model from PubMedBERT and achieved a 0.73 F1 score on the same corpus at the same token-level.We used the same experimental setting (i.e.data, training parameters) and evaluation script provided by the EBM-NLP corpus and reported the performance at the token-level using micro-averaged precision, recall, and F1-score (Supplementary Results Table S3).In addition, we also reported the performance at entity-level by matching the exact spans.
We trained the NER models for the EBM-NLP mod dataset, COVID-19 dataset, and AD dataset using PubMedBERT for five epochs with a learning rate of 1eÀ5 and a batch size of 32.We also experimented with the NER models with other pre-trained models including BERT (Devlin et al. 2018), BioBERT (Lee et al. 2020), BioM-ALBERT (Alrowili and Vijay-Shanker 2021), and BioM-ELECTRA (Gu et al. 2022).The PubMedBERT was better among all pre-trained models in our previous study.
NER evaluation
We evaluated the performance of the NER models on two levels: (i) token-level; and (ii) entity-level.For the token-level evaluation, we used the original evaluation script from EBM-NLP corpus.The script excludes all the "Outside" labels in Inside, Outside, Beginning (IOB) tagging for a fair comparison.The token-level evaluation may not be the best approach because many biomedical named entities include multiple tokens, and the goal is to identify the whole entity.For example, the Participant element, "86 hospitalized COVID-19 patients," was partially identified as "86 hospitalized" and "patients" in the token-level evaluation, The condition, "COVID-19" was omitted.This results in an incomplete representation of the Participant element.For the entity-level evaluation, the entire entity span is viewed as the Participant element.The evaluation preserves comprehensive information.The entity-level evaluation is more reliable than the token-level evaluation.For both types of evaluations, we calculated precision (P), recall (R), and F1 scores for each PICO element, as well as the micro-averaged overall scores for P, R, and F1.These micro-averaged overall scores are computed using the sum of True Positives (TP), False Positives (FP), and False Negatives (FN) from each PICO element.The formula for these scores is provided in Supplementary Materials S1.3.
End-to-end evaluation
In addition to evaluating the sentence classification model and the NER system individually, we also performed an end-toend evaluation to assess the combined performance of the sentence classification model and NER.For a more direct comparison with the standalone NER module, we implemented our two-step pipeline on the identical dataset used to evaluate the standalone NER module.We maintained a consistent evaluation metric used for the standalone NER module.
Performance of the sentence classification module
Our approach achieved an overall F1 score of 0.953 on the EBM-NLP mod dataset, 0.931 on the COVID-19 dataset and 0.962 on the AD dataset.The F1 score for the Methods section alone is 0.949 for the EBM-NLP mod dataset, 0.923 for the COVID-19 dataset, and 0.955 for the AD dataset (Table 3).A comparison between our sentence classification approach and the existing state-of-the-art approach using HSLN architecture validates the superior performance of our approach.A detailed comparison of both approaches and their performance is in Supplementary Results Table S4.
Performance of standalone NER module
Table 4 presents the number of P, I, C, and O elements in three datasets.Table 5 and Supplementary Results Table S5 show the performance of PubMedBERT on PICO extraction (i.e.token-level and entity-level respectively) on all three datasets.We observed that our pipeline achieves higher performance on AD and COVID-19 datasets than the EBM-NLP mod dataset.Our approach looks promising with F1 score >0.8 for token-level evaluation and >0.68 for entity-level evaluation on all three datasets.
End-to-end evaluation of PICO extraction system
The token-level F1 score for the EBM-NLP mod dataset, COVID-19 dataset, and AD dataset, was 0.833, 0.928, and 0.899.Similarly, the entity-level F1 score for these datasets was 0.712, 0.850, and 0.805.The performance of our twostep NLP pipeline system is better across all datasets when compared to the standalone NER module.
Discussion and future work
Our section-specific annotation schema aimed to reduce annotation inconsistencies by classifying sentences before NER, decreasing annotation complexity and time.Although focused on the methods section, it can be extended to other sections if needed.Our method balances minimizing complexity and information loss, covering 95.2% of all entities in the Methods section and improving inter-annotator agreement.Our system has significantly enhanced PICO extraction performance, but with a modest impact on the COVID-19 dataset, possibly due to its specificity and contemporary nature.
However, there is further scope for improvement.To perform error analysis, we evaluated our NER results on the entity-level by partial match.The models achieved 0.848, 0.924, and 0.899 for EBM-NLP mod , COVID-19, and AD datasets, respectively.Detailed performance by different entity types and the confusion matrix are shown in Supplementary Results Table S7 and Supplementary Results Fig. S2.The F1 score for the Control element is lower when compared to other elements across all three datasets (Tables 5 and 6).The confusion matrix shows that 17% of controls are confused with the Intervention element of PICO.In several studies, Control and Intervention elements are the same and it is difficult to distinguish them without a proper context.In most cases, a proper understanding of whether an intervention was applied to a "study group" or a "control group" is required to distinguish between the Control and Intervention elements.The sentence, "Patients with severe COVID-19 were randomly divided into two groups: the standard treatment group and the standard treatment plus hUC-MSC infusion group," says that the standard treatment was given to both groups.The first mention was annotated as the control and the second mention was annotated as the Intervention element.This may be confusing to the model because of the lack of context to learn the difference between the Intervention and Control elements.To ameliorate this confusion between Intervention and Control elements, we need to strategically incorporate more contextual information.This might be achieved by performing NER on all the sentences from each abstract together, rather than processing each sentence in isolation.This approach has the potential to provide the model with richer contextual information.We also observed that numerous Control elements labeled as "Placebo," lead to model overfitting.For instance, in the sentence, "either 3 weeks of taper and 5 weeks of placebo only or continuing use of risperidone."The Control is "5 weeks of placebo" as per the annotation guidelines.However, the model only partially identified "placebo" as the Control.Such issues can be resolved by incorporating more specific annotation rules, such as, "do not include any prepositional phrases preceding a Control element." Nearly 19% of the "Inside" label was misclassified as the "Outside" label in the Participant element.But the "Beginning" label of the Participant element is well-identified.This suggests that the NER model struggles with discerning the ending position of the Participant element.The issue may arise from our annotation guidelines, which include the rule for annotating the longest noun phrase for the Participant element.Clear boundary rules might address this.For example, we could stipulate that "only one prepositional phrase in a Participant element should be included." The overall performance of our PICO extraction system also depends on the performance of sentence classification model.While our sentence classification model categorizes the sentences into general categories based on rhetorical roles, this could be extended to classifying sentences directly into P, I, C, and O categories.The NICTA-PIBOSO corpus incorporated P, I, and O categories in addition to the background and study design, with every sentence belonging to a single category.However, it is very common that a sentence may include multiple PICO elements.Hence, developing a multilabel sentence classification model for identifying PICO categories will be more beneficial when compared to the existing classification model.
The broader applicability of our approach is yet to be established.Testing models trained on specific datasets like AD and COVID-19 on other diseases may offer insights.Utilizing transfer learning could save further annotation time.
Many existing studies (Nye et al. 2018, Zhang et al. 2020) including our work use only abstracts to identify the PICO elements from PubMed articles.However, certain interventions and outcomes are reported only in the full-text of the article.Our approach will miss extracting the interventions and outcomes mentioned only in the full-text articles.Developing a system that can extract PICO elements from both PubMed abstract and full-text articles or applying our pipeline to the full-text might be useful.
In the future, we plan to develop an interface that can take a PubMed identifier or an abstract as input and returns a list of extracted PICO elements as output.The Cochrane Review currently provides a PICO search engine associated with their reviews.However, those PICO elements are manually curated from a limited number of articles.
Conclusions
We presented a new two-step extraction approach to extract PICO elements from the RCT abstracts.We modified annotation guidelines to improve the annotation quality and IAA and reduce annotation complexity for PICO extraction.By annotating the method section and title alone, we not only reduce annotation complexity for PICO extraction but were able to achieve a much higher performance on retrieving unique PICO elements without much loss of information from a subset of RCT abstracts from the EBM-NLP corpus.We verified the usability and reliability of our system by applying and evaluating it on an unseen dataset.
Figure 1 .
Figure 1.Overview of PICO extraction system.(a) The two-step PICO extraction system that includes a sentence classification and NER of PICO elements.(b) Training and test datasets for sentence classification and NER
Table 1 .
Distribution of PICO entities in different sections of the abstract for a random selection of 30 abstracts.
Table 2 .
Number of sentences in each class for the sentence classification evaluation dataset.
Table 3 .
Performance of the prompt-based sentence classification model on the evaluation dataset of COVID-19, AD, and EBM-NLP mod .
Table 4 .
Statistics of PICO elements in all three corpora.
Table 5 .
Entity-level evaluation of the standalone NER model on all three corpora by exact match.
Table 6 .
An end-to-end entity-level evaluation on all three datasets. | 6,101.2 | 2023-09-01T00:00:00.000 | [
"Computer Science",
"Medicine"
] |
Poisson manifolds of strong compact type over 2-tori
In arXiv1312.7267, the first non-trivial example of a Poisson manifold of strong compact type is given. The construction uses the theory of K3 surfaces and results in a Poisson manifold with leaf space $S^1$. We modify the construction to obtain a new class of examples. Specifically, we obtain for each strongly integral affine 2-torus a Poisson manifold of strong compact type with said torus as leaf space.
Introduction
Like symplectic geometry, Poisson geometry started from the mathematical formalisation of classical mechanics.Roughly speaking, a Poisson manifold is a smooth manifold equipped with a Poisson bracket on its space of smooth functions, which allows one to formulate Hamiltonian dynamics.Examples of Poisson manifolds include symplectic manifolds and duals of Lie algebras, an early glimpse into the deep connection with symplectic geometry and Lie theory.Unlike symplectic manifolds, Poisson manifolds are very flexible in nature.For instance, every manifold admits a Poisson structure and there is no local classification of Poisson structures.For this reason it is common to restrict one's attention to specific classes of Poisson manifolds, where one can formulate deep results about their geometry.In this paper we are concerned with Poisson manifolds of compact type (PMCTs).PMCTs are the "compact objects" in Poisson geometry.They were first introduced in [Crainic and Fernandes 2005] and their role in the theory is analogous to the one played by compact Lie algebras in Lie theory.Just as there is the special class of compact semisimple Lie algebras among compact Lie algebras, there is an important distinguished class among PMCTs, namely that of Poisson manifolds of strong compact type (PMSCTs).A simple class of examples of PMSCTs is given by compact symplectic manifolds with finite fundamental group, but it is difficult to construct examples that are not symplectic.The first such example was given in [Martínez Torres 2014], building on work of [Kotschick 2006].There a regular PMSCT is constructed whose symplectic leaves are all diffeomorphic to a K3 surface and whose leaf space is diffeomorphic to a circle.One can form new PMSCTs by taking products of the aforementioned examples, but apart from these no other examples are known.In this paper we use the construction of [Martínez Torres 2014] to obtain new examples of PMSCTs.It is known that the leaf space of a PMSCT must be a compact integral affine orbifold and in the example of [Martínez Torres 2014] this is the "standard" integral affine structure on the circle.In this work we show that all strongly integral affine circles and two-dimensional tori can appear as the leaf space of a PMSCT.
In order to explain our main result, recall that a Poisson structure on a manifold M is a Lie bracket on C ∞ (M) which is a derivation in each entry.Equivalently, a Poisson structure is a bivector π ∈ X 2 (M) satisfying [π, π] = 0.This is the definition we work with in this paper.Every Poisson manifold has a partition into symplectic manifolds.This symplectic foliation can be viewed as a singular foliation integrating the (singular) distribution π # (T * M) ⊂ T M. If π has constant rank, this is actually a regular foliation.In this case the Poisson manifold is called regular.
The "global" objects in Poisson geometry are the so-called symplectic groupoids.A symplectic groupoid is a Lie groupoid G ⇒ M carrying a multiplicative symplectic form ∈ 2 (G).A Poisson manifold (M, π ) is called integrable if there exists some symplectic groupoid (G ⇒ M, ) for which the target map t : (G, ) → (M, π ) is a Poisson map (see [Crainic et al. 2021]).PMCTs are defined as those Poisson manifolds that are integrated by a source connected, Hausdorff symplectic groupoid having a certain compactness property.Contrary to the case of Lie groups and Lie algebras, there are multiple notions of compactness for Lie groupoids, namely a Lie groupoid G ⇒ M is called • source proper, or s-proper, if the source map is proper; • compact if the space of arrows G is compact.
Accordingly, we say that (M, π ) is of proper/source proper/compact type if it admits a source connected, Hausdorff symplectic groupoid of proper/source proper/compact type, respectively.
The types just defined depend on the choice of integration of (M, π ).However, just like for Lie groups, there is a unique "largest" integration, namely the one with 1-connected source fibres.This is often called the Weinstein groupoid.We say that an integrable Poisson manifold has strong proper/source proper/compact type if its Weinstein groupoid is Hausdorff and has the corresponding type.As mentioned above, we will focus here on Poisson manifolds of strong compact type.
Unlike general Poisson manifolds, PMCTs have a rich geometry transverse to their associated symplectic foliation.For example, the leaf space of a regular PMCT inherits the structure of an integral affine orbifold.Roughly speaking this means that the leaf space has an orbifold atlas where the transitions are integral affine maps.The precise statement can be found in [Crainic et al. 2019a;2019b], where many other properties of PMCTs are discussed.
As mentioned above, the first example of a PMSCT that is not symplectic was given in [Martínez Torres 2014].The construction there is inspired by [Kotschick 2006], where nontrivial results on the geometry of K3 surfaces are used to construct a free symplectic circle action with contractible orbits.The orbit space of such an action is a PMSCT with smooth leaf space a circle endowed with its standard integral affine structure (that is, the one it inherits as a quotient of ޒ by ޚ acting by translations).In general, it is not known whether any compact integral affine orbifold can appear as the leaf space of a PMSCT.On the one hand constructing strong PMCTs is a difficult problem on its own, and on the other not much is known about the classification of compact integral affine manifolds in dimension greater than two.The integral affine structures on a circle are easily classified, and the classification of integral affine structures on compact 2-dimensional manifolds was obtained in [Mishachev 1996;Sepe 2010].The main result of this paper is the following.
Main theorem.Any strongly integral affine circle or two-dimensional torus can be realised as the leaf space of a PMSCT.
Here by a strongly integral affine structure we mean an integral affine structure with integral translational part (see [Sepe 2013, Remark 5.10] and Remark 4.2).
Our strategy to prove this result is as follows.Using the geometry of K3 surfaces one constructs a universal family of marked Kähler K3 surfaces (see Section 3) to which one can apply a general method from [Crainic et al. 2019b] to obtain PMSCTs.Using this construction together with the classification of integral affine 2-tori from [Mishachev 1996], one obtains examples of PMSCTs for all isomorphism classes of strongly integral affine 2-tori.
This paper is organised as follows.In Section 2, we provide some background on PMCTs and we recall the general method of constructing regular PMSCTs from [Crainic et al. 2019b].In Section 3 we recall the relevant results on K3 surfaces that are needed for our construction.The resulting examples of PMSCTs have symplectic foliation a fibration over S 1 or ޔ 2 with typical fibre the smooth manifold underlying a K3 surface.The symplectic structures on the fibres vary in a controlled fashion which ensures that the Weinstein groupoid is a compact symplectic groupoid.Finally, Section 4 is dedicated to the actual constructions, which includes some lengthy computations.We treat the circle case first and this includes the original example from [Martínez Torres 2014].Lastly, we construct the PMSCTs with leaf space the strongly integral affine 2-tori.
Background & general construction of PMSCTs
The construction we give below is based on two results on PMCTs: (a) The leaf space carries an integral affine orbifold structure (see [Crainic et al. 2019b, Section 3]).
We briefly recall these results before giving the general construction.Here we only need to consider the case of 1-connected leaves.In this case the leaf space is smooth, since this assumption implies that the monodromy groupoid of the symplectic foliation is proper and has trivial isotropy groups.Then both (a) and (b) above simplify significantly.
2A.The integral affine structure on the leaf space.Recall that an integral affine structure on a manifold B is given by an atlas whose transition functions are integral affine maps.Equivalently, it is specified by a lattice ⊂ T * B locally spanned by closed 1-forms.
Consider a regular, s-connected, proper symplectic groupoid (G, ) ⇒ (M, π ).As mentioned above, we assume that the associated symplectic foliation F π has 1-connected leaves so that the leaf space B is a smooth manifold.We obtain a lattice ⊂ ν * (F π ) as follows: (1) For each x ∈ M, the kernel of the exponential map g x → G x gives a lattice in g x .
(2) The isomorphism g x ∼ = ν * x (F π ) induced by allows us to transport it to the conormal space.This lattice descends to an integral affine structure ⊂ T * B on B.
2B.The linear variation theorem.We assume now in addition that (G, ) ⇒ (M, π ) is source proper.Denoting the symplectic leaf corresponding to b ∈ B by (S b , ω b ), we form the vector bundle and the lattice inside it.Associated to this we have the Gauss-Manin connection ∇ on H 2 , uniquely determined by requiring the sections of H 2 ޚ to be parallel.Note that π gives us a section The Gauss-Manin connection allows us to study the variation of ϖ : parallel transport makes H 2 into a 1 (B)-representation and we define the variation map var ϖ : 1 (B) → H 2 to be On the other hand, we also have the linear variation map var lin ϖ : TB → H 2 given by v → ∇ v ϖ and the affine variation map var aff ϖ := ϖ + var lin ϖ .The linear variation theorem relates the variation and affine variation maps by means of the developing map associated to the integral affine manifold (B, ).Associated to the lattice * ⊂ TB we have a canonical flat connection on TB (not to be confused with ∇ above).This makes TB into a TB-representation, and since the connection is torsion-free the identity map TB → TB is an algebroid cocycle.The developing map is defined to be the groupoid cocycle dev : 1 (B) → TB integrating it.
Remark 2.1.One can show that after fixing b ∈ B and a basis of b this boils down to the classical notion of developing map defined on the universal covering space (see [Crainic et al. 2019b, Section 4.2]): We can now state the linear variation theorem as follows.
where c i ∈ H 2 (S b 0 , )ޚ are the Chern classes of the torus bundle s −1 (x) → S b 0 , where x ∈ S b 0 (see [Crainic et al. 2019b, Corollary 4.4.4]).This local formulation is reminiscent of the linear variation theorem from [Duistermaat and Heckman 1982].In other words, Theorem 2.2 can be viewed as a global formulation and generalisation of the classical Duistermaat-Heckman theorem.
2C.The construction.The construction we describe in this section yields a PMSCT with 1-connected symplectic leaves, whose leaf space is a complete integral affine manifold.This means that the leaf space is a quotient of ޒ q by a free and proper action of a discrete group of integral affine transformations.Note that if the Markus conjecture holds true, then in fact every compact integral affine manifold is of this type (see [Goldman 2022, Section 8.6]).This allows us to give an explicit formulation of the linear variation, similar to the discussion following Theorem 2.2.The setup is as follows.
Let E → ޒ q be a fibre bundle with typical fibre S, a compact 1-connected manifold, and assume that E admits a Poisson structure π E whose symplectic leaves are precisely the fibres of this bundle.As in Section 2B we have (i) the vector bundle H 2 → ޒ q whose fibres are the degree two cohomology groups of the symplectic leaves, (ii) the lattice H 2 ޚ ⊂ H 2 of integral cohomology, (iii) the associated Gauss-Manin connection ∇ and } be a discrete group of integral affine transformations acting freely and properly on ޒ q , and assume that there is a Poisson action of on (E, π E ) making the projection E → ޒ q equivariant.Then setting M := E/ and B := ޒ q / , we get a (smooth) fibre bundle p : M → B, again with typical fibre S, and a Poisson structure π on M whose leaves are the fibres of p.In other words, (M, π ) is a regular Poisson manifold with leaf space B. Note also that B, being a quotient ޒ q / , naturally inherits an integral affine structure.
We can now state the general method of constructing PMSCTs.It is a reformulation of [Crainic et al. 2019b, Proposition 4.4.6].
Proposition 2.3.Let (M = E/ , π ) be constructed as above.Assume that there exists a ∇-flat section s ∈ (H 2 ) and linearly independent sections c 1 , . . ., where pr i : ޒ q → ޒ denotes projection onto the i-th coordinate.Then (M, π ) is of strong s-proper type and the induced integral affine structure on B agrees with the one coming from the quotient ޒ q / .In particular, if B is compact then (M, π ) is a PMSCT.
Proof.Pulling back the integral affine structure on B along p : M → B yields a transverse integral affine structure on the symplectic foliation F π , i.e., a lattice in its conormal bundle.We denote this lattice by ⊂ ν * (F π ).The main point is that for all x ∈ M, the monodromy group N x (M, π ) is equal to the lattice x .In fact, using the description of the monodromy groups for regular Poisson manifolds as the variation of symplectic areas (see [Crainic and Fernandes 2004, Section 6]) this follows directly from (2-1).The integrability criteria for Poisson manifolds then imply that (M, π) is integrable.Furthermore, since S has trivial fundamental group, the isotropy groups of the Weinstein groupoid (M, π ) fit into the exact sequence where ∂ x is the monodromy map at x. Therefore, from our previous discussion, it follows that x (M, π) ≃ ν * x (F π )/ x , i.e., that the isotropy group at x is compact.Since this holds for all x ∈ M and since S is also compact, this shows that the Weinstein groupoid is s-proper.We start by listing some definitions and results concerning K3 surfaces, after which we describe the moduli spaces and universal families for K3 surfaces.These results can be found in [Barth et al. 1984].Finally, following [Martínez Torres 2014], we use the Calabi-Yau theorem to turn the universal family into a Poisson manifold and the strong Torelli theorem to establish a Poisson action on it, setting us up to apply our construction.
Definition 3.1.A K3 surface is a compact, 1-connected complex surface with trivial canonical bundle.
Every K3 surface is Kähler (see [Siu 1983]).All K3 surfaces have the same underlying smooth manifold S (see [Barth et al. 1984, Corollary VIII.8.6]); this will be the model fibre used in Proposition 2.3.The intersection form on H 2 (S, )ޚ turns it into a lattice and this lattice is isomorphic to the aptly named K3 lattice, which we denote by (L , ( •, • )).It is the unique even, unimodular lattice of signature (3, 19) (see [Barth et al. 1984, Proposition VIII.3.2(ii)]).Explicitly, we have , where U = ޚ ⊕2 with form given by 0 1 1 0 and E 8 = ޚ ⊕8 with form given by the Cartan matrix of E 8 ; it is important for us that this form is positive definite.We also set L ޒ := L ⊗ ޒ and L ރ := L ⊗ ;ރ note that these are models for the real and complex cohomology, respectively.
A Hodge isometry is called effective if its -ޒlinear extension maps some Kähler class of X ′ to one of X .
Effectiveness of a Hodge isometry is equivalent to requiring it to map the Kähler cone of X ′ to that of X (see [Barth et al. 1984, Proposition VIII.3.10]).Barth et al. 1984, Corollary VIII.11.4]).Let X, X ′ be K3 surfaces.Then for any effective Hodge isometry ϕ : H 2 (X ′ , )ޚ → H 2 (X, )ޚ there exists a unique biholomorphism f : X → X ′ such that f * = ϕ.This result is ultimately used to obtain the action in Proposition 2.3.
3B. Moduli spaces and universal families.
There are two moduli spaces and corresponding families for K3 surfaces: one takes into account the Kähler structure and the other only considers the complex structure.We start now with the latter.Definition 3.4.A marked K3 surface is a pair (X, ϕ) consisting of a K3 surface X and a marking ϕ, i.e., an isometry ϕ : H 2 (X, )ޚ → L. Two marked K3 surfaces are equivalent if there is a bihomolorphism between them intertwining the markings.The moduli space of marked K3 surfaces is the set of equivalence classes: It follows immediately from the definition that any K3 surface admits, up to scalar multiplication, a unique nowhere vanishing holomorphic 2-form.In fact, one can show that, again up to scalar multiplication, there is a bijection between complex structures on S and closed, complex 2-forms σ ∈ 2 (S, )ރ satisfying σ ∧ σ = 0 and σ ∧ σ > 0. This motivates the following definitions.We will use the same letter to denote a marking ϕ : H 2 (X, )ޚ → L and the induced maps ϕ : H 2 (X, )ޒ → L ޒ and ϕ : H 2 (X, )ރ → L ރ .
Definition 3.5.The period domain is given by We define the period map τ where σ X is a nowhere vanishing holomorphic 2-form on X .
Theorem 3.6 [Barth et al. 1984, Theorem VIII.12.1].The moduli space M 1 admits the structure of a 20-dimensional complex manifold such that the period map τ 1 : M 1 → becomes a surjective local biholomorphism.Furthermore, there exists a universal family U → M 1 of marked K3 surfaces.
Remark 3.7.Recall that a family is universal if any other family is locally the pullback of it by a unique map (see [Barth et al. 1984, Section I.10]).The fibre of the universal family U → M 1 over any t ∈ M 1 is a marked K3 surface (X t , ϕ t ) such that [(X t , ϕ t )] = t.Furthermore, these markings vary smoothly in the sense that they induce local trivialisations of the bundle t∈M 1 H 2 (X t , .)ޒThere are still some inconveniences present here.It can be shown that M 1 is not Hausdorff, and that the period map τ 1 is not injective (see [Barth et al. 1984, Remark VIII.12.2]).These problems disappear when taking into account the Kähler structure.
Definition 3.8.We define M 2 to be the subset of the bundle It can be shown that M 2 is a real-analytic manifold of dimension 60 (see [Barth et al. 1984, Lemma VIII.9.3] and its proof).One should think of a point in M 2 as an equivalence class of marked K3 surfaces together with a specified Kähler class.Note that there is a projection map pr : Inspired by some analysis of the Kähler cone of K3 surfaces (see [Barth et al. 1984, Sections VIII.3 and VIII.9]) one makes the following definitions.Definition 3.9.Set The refined period domain is then given by The refined period map τ 2 : M 2 → K 0 is defined as Theorem 3.10 [ Barth et al. 1984, Theorems VIII.12.3 and VIII.14.1].The refined period map is a diffeomorphism.
We set K U := (pr•τ −1 2 ) * U.This is a real-analytic family (i.e., fibre bundle) over K 0 with extra data attached: the fibre over (k, [σ ]) is a triple (X, ϕ, ω) consisting of a K3 surface X , a marking ϕ : H 2 (X, )ޚ → L and a Kähler class ω ∈ H 2 (X, )ޒ such that ϕ(ω) = k.These markings vary smoothly in the same sense as before, and hence so do the Kähler classes.
The family K U → K 0 is universal for real-analytic "marked Kähler K3 families", i.e., real-analytic families of K3 surfaces equipped with smoothly varying markings and Kähler classes.
Theorem 3.11.Let X be a compact complex manifold with vanishing first Chern class.Then for any Kähler class ω ∈ H 2 (X, )ޒ there exists a unique Ricci flat Kähler metric whose Kähler form belongs to ω.
This theorem applies in particular to K3 surfaces, and thus we can use it to endow the fibres of K U → K 0 with smoothly varying Kähler forms, turning it into a Poisson manifold (see also [Martínez Torres 2014, Section 2.1.3]).
Corollary 3.12.The family K U admits a regular Poisson structure π K U whose symplectic leaves are the fibres of K U → K 0 .Moreover the symplectic form on the fibre X over (k, [σ ]) with marking ϕ is the Kähler form associated to the unique Ricci flat Kähler metric on X with Kähler class ϕ −1 (k).
3D.The action.We will construct an action on K U by the group O(L) of isometries of the K3 lattice.Note that there is an obvious induced action of O(L) on K 0 .Proposition 3.13.There is a Poisson action of O(L) on (K U, π K U ) with respect to which the projection K U → K 0 is equivariant.
Proof.Fix γ ∈ O(L) and p ∈ K 0 .Using the notation from above, denote the triple over p by (X p , ϕ p , ω p ) and similarly for γ ( p).It is easy to see that is an effective Hodge isometry, so that by Theorem 3.3 we obtain a biholomorphism f p γ : X p → X γ ( p) .The universality of the family then gives neighbourhoods U and V of p and of γ ( p) respectively and an isomorphism ( , ψ) : K U| U → K U| V extending f p γ : through the biholomorphism f p γ , K U becomes a deformation of X p at two basepoints, p and γ ( p).Since K U is universal, these two deformations are locally isomorphic.Writing q : X q → X ψ(q) for the fiberwise maps, it then follows that for all q ∈ U we have that * This implies first of all that ψ = γ | U , from which it follows that q = f q γ , since biholomorphisms of K3 surfaces are uniquely determined by their induced maps on degree 2 integral cohomology (see [Barth et al. 1984, Proposition VIII.11.3]).Thus these fibrewise biholomorphisms f p γ , p ∈ K 0 , together form an automorphism F γ : K U → K U. It is immediate from the above construction that F id = id, and from the uniqueness part of Theorem 3.3 it follows that , meaning that we have an action of O(L) on K U.This action makes K U → K 0 equivariant by construction.Finally, from the uniqueness part of the Calabi-Yau theorem it follows that each f p γ preserves the symplectic forms on the fibres, meaning that the action is by Poisson maps.□
The examples
From our work in Section 3 we have a Poisson manifold (K U, π K U ) with leaf space K 0 such that (i) the cohomology classes of the symplectic forms on the leaves are described in terms of the leaf space K 0 (Corollary 3.12); (ii) the natural action of O(L) on K 0 lifts to a Poisson action on (K U, π K U ) (Proposition 3.13).
In order to apply the construction described in Section 2, we need to find a suitable embedding ޒ q → K 0 and a suitable subgroup ⊂ O(L).We rephrase Proposition 2.3 in the current setting in order to make this more precise.For a different version of this result see also [Martínez Torres 2014, Theorem 1].
Corollary 4.1.Assume that we have an embedding f : ޒ q → K 0 and a subgroup ⊂ O(L) such that (i) there exist a ∈ L ޒ and linearly independent a 1 , . . ., a q ∈ L such that the L ޒ -component of f has the form (x 1 , . . ., x q ) → a + q i=1 x i a i ; (ii) the action of on K 0 preserves the image of f ; (iii) the induced action on ޒ q is free, proper and by integral affine maps.
Then M := f * K U/ with the Poisson structure induced from π K U is a Poisson manifold of strong s-proper type with leaf space B := ޒ q / .If B is compact, M is a PMSCT.
Remark 4.2.We can now explain why our construction leads to PMSCTs with strongly integral affine leaf spaces.On the one hand, because of Theorem 2.2, we are forced to consider embeddings with integral variation, i.e., the a i must lie in the integral lattice L. On the other hand, to apply Theorem 3.3 we need to consider isometries of integral cohomology, i.e., we need to act by elements of O(L).These two technical limitations together only allow for strongly integral affine leaf spaces in the examples.
Remark 4.3.At the level of the symplectic groupoid, one can see that the leaf space being strongly integral affine implies that the restriction of the symplectic form to the identity component of the isotropy (a torus bundle) lies in the integral cohomology.See [Sepe 2013, Remark 5.10].
We now recall the classification of strongly integral affine structures for S 1 and ޔ 2 .
Theorem 4.4.The strongly integral affine circles are, up to isomorphism, the quotients ޚ/ޒ where the -ޚaction is generated by x → x + p, for a fixed p ∈ ޚ ≥1 .
Proof.It is easy to see that all integral affine circles are complete.Hence, it suffices to classify, up to conjugation, embeddings ޚ → Aff ޚ )ޒ( inducing free and proper actions.These are precisely the actions generated by x → x + a with a > 0.
Proof.The classification of all integral affine structures on 2-tori is given in [Mishachev 1996, Theorem A].Restricting to strongly integral affine structures and using the Smith normal form for matrices with integer entries to simplify the possibilities from type (I) yields the above classification.□ Remark 4.6.The integral affine 2-tori of type (I) are (isomorphic to) products of integral affine circles.Thus to find examples of PMSCTs with leaf space of this type one can simply take products of PMSCTs with leaf space S 1 , constructed in Section 4A.This yields Poisson manifolds of dimension 10 whose leaves are products of K3 surfaces.However, the examples we construct in Section 4B are six-dimensional Poisson manifolds with K3 surfaces as symplectic leaves and thus result in "smaller" examples.
Remark 4.7.Continuing the previous remark, note that by taking products we can also realise some higher-dimensional integral affine tori as the leaf space of a PMSCT, namely those that are isomorphic to a product of some of the integral affine circles and 2-tori classified above.
Before we move on to the examples, we establish some notation.Recall that L = U ⊕3 ⊕ (−E 8 ) ⊕2 .We denote the standard bases of the three copies of U by {u, v}, {x, y} and {z, t}, so that (u, v) = (x, y) = (z, t) = 1 with all other combinations yielding zero.Recall also that −E 8 is even and negative definite.Finally, let {e 1 , . . ., e 8 } be a set of real numbers such that the set {1, e 1 , . . ., e 8 , e 2 1 , e 1 e 2 , . . ., e 2 7 , e 7 e 8 , e 2 8 } consisting of 1, e 1 , . . ., e 8 and their pairwise products is linearly independent over the integers, or equivalently the rationals.The existence of such a set is guaranteed by [Mordell 1953].We then set e := (e 1 , . . ., e 8 ) ∈ (−E 8 ) ޒ , scaling if necessary such that |(e, e)| ≤ 1 2 , and we set a := (0, e), b := (e, 0) ∈ (−E 8 ) ⊕2 ޒ ⊂ L ޒ .Let us outline the strategy for the examples below.In each case, we start by defining f and .It is fairly straightforward to check items (ii) and (iii) from Corollary 4.1 and that the image of f is contained in K .It then remains to show that it is actually contained in K 0 .This is the more involved part of the computations.
4A.The PMSCTs with leaf space the circle.We will construct a PMSCT whose leaf space is a strongly integral affine circle, i.e., we want the action of ޚ on ޒ generated by x → x + p with p ∈ ޚ ≥1 .The case p = 1 is the one treated in [Martínez Torres 2014] and the computations carried out below for general p are an obvious generalisation of the computations there.
Consider the map f : ޒ → L ޒ × (ސL ރ ) defined by and the map ϕ : L → L defined by u → u, v → v + py, x → x − pu, y → y on the first two copies of U and as the identity on the other summands of L. It is easily checked that ϕ is an isometry and that This implies that the image of f is invariant under the action of := ⟨ϕ⟩, and also that the induced action on ޒ is the one we need.
To show that the image of We need to find a contradiction.Let us write with A, . . ., F ∈ ޚ and d i in the i-th copy of −E 8 .Since E 8 is even and positive definite, we can write (d i , d i ) = −2n i , for n i ∈ ޚ ≥0 .The above conditions then translate into three equations: This is where the seemingly strange choice of e comes in.There exist k 1 , . . ., k 8 ∈ ޚ such that (d 1 , e) = i k i e i and since {1, e 1 , . . ., e 8 } is linearly independent over the integers by choice of e, it follows from (4-4) that we must have Since the bilinear form on −E 8 is nondegenerate, it follows that d 1 = 0 and thus that n 1 = 0.
Case C ̸ = 0: From (4-2) we get and substituting this into (4-3) yields Combining this with (4-1) gives From the properties of e we get Cd 2 = 0, implying that d 2 = 0 and thus also that n 2 = 0, so that we are left with The PMSCTs with leaf space a torus of type (I).Here we construct a PMSCT with leaf space the torus ޔ 2 with an integral affine structure of type (I).This means that we want the action of ޚ 2 on ޒ 2 generated by (x, y) → (x + p, y) and (x, y) → (x, y + q), with p, q ∈ ޚ ≥1 .Consider the map f : ޒ 2 → L ޒ × (ސL ރ ) defined by the map ϕ : L → L as in the previous example and the map ψ : L → L defined by u → u, v → v + qt, x → x, y → y, z → z − qu, t → t on two copies of U and as the identity on the other summands of L. It is easily checked that these are isometries and that This implies that the image of f is invariant under the action of := ⟨ϕ, ψ⟩, and also that the induced action on ޒ 2 is as desired.
To show that the image of f is contained in K , let f 1 , f 2 , f 3 be the three "components" of f , as before, and let (s, r ) ∈ ޒ 2 .We compute It remains to check that f (s, r ) ∈ K 0 for all (s, r ) ∈ ޒ 2 .
Case B ̸ = 0: From (4-7) and (4-8) we get Substituting this into (4-6) gives and combining this with (4-5) we obtain The assumptions on e imply that Cd 2 + Ed 1 = 0, so that this becomes This is impossible under the assumption B ̸ = 0, since n i ∈ ޚ ≥0 .□ 4C.The PMSCTs with leaf space a torus of type (II).In this example we will construct a PMSCT whose leaf space is a torus with an induced integral affine structure of type (II), namely one induced by the action of ޚ 2 on ޒ 2 generated by (x, y) → (x + p, y) and (x, y) → (x + ny, y + q), where n, p, q ∈ ޚ ≥1 .Consider the map f : the map ϕ : L → L defined as before and the map ψ : L → L defined by u → u, v → v + qt, x → x − nz + qnu, y → y, z → z − qu, t → t + ny on the copies of U and the identity on the other summands of L. It is easily checked that these are isometries and that This implies that the image of f is invariant under the action of := ⟨ϕ, ψ⟩, and also that the induced action on ޒ 2 is the desired one.To show that the image of f is contained in K , denote once more by f 1 , f 2 , f 3 the "components" of f , and let (s, r ) ∈ ޒ 2 .Since Like before we write and we set (d i , d i ) = −2n i with n i ∈ ޚ ≥0 .The goal is to find a contradiction.The main equations are now Since C, n, q ̸ = 0 and r ∈ ,ޒ we must have that F 2 ≥ 2Cq 2Anq 2 + C(8n 2 q 3 + 2q) + Dq + (d 2 , e) .
Since C ̸ = 0, this is certainly impossible when C and (d 2 , e) have the same parity.So let us assume that they have opposite parity, so that the equation becomes But this is impossible, since B − 2Cnq ̸ = 0, B ̸ = 0 and n 1 , n 2 ≥ 0, giving us the desired contradiction.□
ϖ
Theorem 2.2 [Crainic et al. 2019b, Theorem 4.4.2].One has a commutative diagram 1 This rather abstract formulation can locally be made explicit.Let b 0 ∈ B and choose an integral affine chart (U, ϕ) centered at b 0 such that ϕ(U ) is convex and such that M → B trivialises over U .This induces a trivialisation : H 2 | U ∼ = U × H 2 (S b 0 , .)ޒThe chart induces an identification T b 0 B ∼ = ޒ q and allows us to consider "straight line" paths from b ∈ U to b 0 .Restricting to such paths the above diagram becomes Finally, since ⊂ ν * (F π ) is closed, Hausdorffness of the Weinstein groupoid follows from [Alcalde-Cuesta and Hector 1995, Theorem 1.1].□ 3. Background on K3 surfaces and the Poisson structure on the universal family | 8,419.6 | 2021-03-06T00:00:00.000 | [
"Mathematics"
] |
Structure and Sequence Aligned Code Summarization with Prefix and Suffix Balanced Strategy
Source code summarization focuses on generating qualified natural language descriptions of a code snippet (e.g., functionality, usage and version). In an actual development environment, descriptions of the code are missing or not consistent with the code due to human factors, which makes it difficult for developers to comprehend and conduct subsequent maintenance. Some existing methods generate summaries from the sequence information of code without considering the structural information. Recently, researchers have adopted the Graph Neural Networks (GNNs) to capture the structural information with modified Abstract Syntax Trees (ASTs) to comprehensively represent a source code, but the alignment method of the two information encoder is hard to decide. In this paper, we propose a source code summarization model named SSCS, a unified transformer-based encoder–decoder architecture, for capturing structural and sequence information. SSCS is designed upon a structure-induced transformer with three main novel improvements. SSCS captures the structural information in a multi-scale aspect with an adapted fusion strategy and adopts a hierarchical encoding strategy to capture the textual information from the perspective of the document. Moreover, SSCS utilizes a bidirectional decoder which generates a summary from opposite direction to balance the generation performance between prefix and suffix. We conduct experiments on two public Java and Python datasets to evaluate our method and the result show that SSCS outperforms the state-of-art code summarization methods.
Introduction
Source code summarization is a popular research task in the code comprehension field which aims to generate natural language descriptions of code for developers to rapidly comprehend the functionality or usage. With the increasing volume of software code, nearly 90% of the development cost is spent on software maintenance, (e.g., version iteration, program comprehension and bug fixing) [1]. High-quality code summaries undoubtedly can effectively reduce the cost on program comprehension. Nowadays, in a practical developing environment, most of the code summaries are likely to be missing or lacking, or the summaries do not match the code due to a series of human mistakes or the large volume of code. Developers' effort in writing qualified summaries determines whether code can be effectively comprehended by other developers. Recently, researchers are devoted to source code summarization tasks to generate high-quality code summaries automatically instead of hand-writing, and it is still challenging.
Early studies on code summarization tasks mainly generated a summary from templates [2] or retrieved a summary from a similar code snippet based on information retrieval (IR) techniques [3]. Whether the template-based approaches or the IR-based approaches, the quality of the summaries is far from satisfactory.
Recently, due to the remarkable achievement in neural machine translation tasks utilizing deep learning techniques, studies on automatic code summarization based on a sequence-to-sequence architecture that contains an encoder for code representation and a decoder for the summary generation attracts researchers devoted in this area. Researchers [4][5][6] only consider the code as plain text data, which is called sequential information (e.g., token sequence, API sequence) and generate summaries based on it with an RNN (Recurrent Neural Network) or CNN (Convolutional Neural Network) sequenceto-sequence model. Further, transformer [7] architecture shows great potential in the Natural Language Processing (NLP) area providing a breakthrough for all language-related tasks. As a result, works such as [8,9] on using transformers to generate summaries, achieved better results than the RNN-based or CNN-based methods. Ahmad et al. [8] first proposed a transformer-based method on the code summarization task, which achieved excellent performance and leads the code summarization area into the transformer-based model stage. Because of the popularization and performance of transformers, almost all recent works [9][10][11][12] are conducted based on the transformer architecture and achieve high scores in each evaluation metric. However, only considering sequence information without considering the structure of code leads to a incomplete representation of code. Thus, researchers utilized the structural information of code (e.g., Abstract Syntax Tree (AST), Control Flow Graph (CFG), Data Flow Graph (DFG), Program Dependence Graph (PDG)) to represent code more comprehensively. Specifically, a part of code summarization research [13][14][15] used ASTs as the input to generate code summaries. Shido et al. [16] proposed an LSTM-based (Long Short-term Memory Network) tree structure model to directly capture AST features that enable a traditional NLP model to apply to tree-structural data. Alon et al. [15] extracted multiple tree paths from root node to terminal node based on ASTs as the code representations as structural information for summary generation. Currently, Graph Neural Networks (GNNs) achieve great performance on graph data (e.g., protein prediction, knowledge graph), so researchers [17,18] consider constructing AST into a graph by adding additional edge relations and utilizing GNNs to capture structural information achieving great results.
Generally, to comprehensively represent source code and achieve great generation performance, it is undoubtedly that we must consider both structure and sequence information, thus leading to the two motivations for our work. The first motivation of this paper is to capture the sequence information in a hierarchical view while almost all the works consider code sequence as plain text, despite the fact that codes are a kind of document data containing hierarchical structure. Code is not only a composition of tokens but also a composition of statement sentences, which also means that code contains document property. Thus, to better capture the sequence information, we propose a hierarchical sequence encoder adopting a hierarchical encoding strategy from token-level to sentence-level. For the second motivation, code ASTs or graphs often contain large scale nodes and edges for which the GNNs do not have the parallel computation ability unless we combine batches of code ASTs or graphs into a larger graph, leading to low computational efficiency. Thus, we utilize the high parallel computation ability of transformers to encode a batch of code graphs with a special mask technique. Furthermore, for natural language generation (NLG) tasks, the decoder is an auto-regressive module which generates sentences from left to right, suffering from an exposure bias issue that error will be accumulated and passed to the following step. In conclusion, this leads to an unbalanced performance between prefix and suffix. The prefix denotes the first m tokens in a generated sentence and the suffix denotes the last n tokens. Thus, we adopt a bidirectional decoding strategy to balance the performance between prefix and suffix. We will detail the above designs in Section 3. Contributions: • We propose a structure and sequence-aligned code summarization method called SSCS, which can not only capture both types of information in a way superior to the previous works but also balances the generating performance between prefix and suffix. • We design a transformer-based AST encoder to explicitly encode the structural information in a multi-perspective format and aggregate it with an adapted fusion strategy.
• A hierarchical sequence encoder is proposed to capture the document property of code, which enables the model to understand sequence information from fine grain (token-level) to coarse grain (sentence-level). • We conduct experiments on two public datasets on Java and Python to evaluate the effectiveness of our proposed SSCS. Furthermore, the results show that our approach outperforms exiting state-of-the-art code summarization methods on BLEU, METEOR and ROUGE metrics.
Related Work
In early periods of code summarization task, researchers generated natural language descriptions of code based on the Software Word Usage Model (SWUM) by analyzing the signatures of Java methods [2]. Subsequently, Information Retrieval (IR) techniques emerged and were widely used in code summarization tasks. Haiduc et al. [3] proposed IRbased techniques for automatic code summarization, which searches similar code snippets or keywords in the code database and extracts the summary from similar ones as the summary of the original code. These methods generate code summaries with low flexibility. Recently, researchers utilized deep-learning techniques to represent code and adopt the encoder-decoder framework to generate a more flexible summary. To our best knowledge, recent methods can be broadly divided into three categories: sequence-only methods, structure-aligned methods and pre-trained methods.
As the encoder-decoder framework achieves great success in Neural Machine Translation (NMT), which is able to generate sentences of arbitrary length, researchers consider code summarization as a translation task that translates program language (PL) into natural language (NL). Thus, studies generate summaries with sequence information of the code (e.g., code plain text, API sequence). Iyer et al. [4] were the first to propose CODE-NN, an RNN-based encoder-decoder framework equipped with an attention mechanism, generating a summary from plain text. Allamanis et al. [5] utilized a Convolutional Attention Neural Network (CNN) to generate a function-name-level short summary for specific code. Hu et al. [6] designed a method which utilizes API sequences as complementary information for the code sequence, adding an extra API sequences encoder to represent the API sequence. Wei et al. [19], who regarded code summarization and code generation as complementary tasks, proposed a dual learning model to learn both code and summaries simultaneously, improving the performance of each task. The emergence of transformers [7] brought huge progress in the natural language generation (NLG) field, becoming the mainstream architecture for the subsequent works. Ahmad et al. [8] first proposed a transformer-based method to generate code summaries which add a copy mechanism and relative positional encoding strategy to strengthen the absolute position encoding strategy. Moreover, Wan et al. [1] utilized reinforcement learning strategy, adding an extract actor-critic module to conduct the generation performance. The above methods generated summaries with sequence information only and demonstrated the potential of deep-learning techniques in the code summarization field.
Only considering the source code as a sequence ignores the structural property of the code and leads to an incomplete representation of the code, thus the performance of the code summarization task encounters a bottleneck. To overcome the bottleneck caused by incomplete representation, researchers constructed a more comprehensive representation by adding structure information from ASTs. Hu et al. [13] proposed a transformation algorithm that transforms tree-structure data from AST into sequence data SBT (structurebased traversal) and adopted the RNN-based encoder-decoder framework to generate the summary. LeClair et al. [14] utilized both AST sequence and code sequence with two separate encoders to capture the different kinds of information. Liang and Zhu [20] proposed a tree-based recursive neural network to directly capture AST data instead of transforming the AST into a sequence. Shido et al. [16] proposed tree-LSTM to capture the structural information. Fernandes et al. [21] constructed sequence graphs of code and used Gated Graph Neural Networks (GGNN) [22] as the encoder that can capture the distance relationship in the sequence graphs. Alon et al. [15] extracted multiple paths from the AST, constructing a path-based sequence to represent the structure of code. LeClair et al. [23] proposed a graph-based neural network which uses an RNN-based network to capture the sequence information and GCN for the structural information. Wang et al. [24] generated code summary by aggregating source code sequence information, ASTs and a control-flow graph with reinforce learning strategy. In general, by integrating sequence information and structural information, code summarization has made great progress but there is still a barrier to satisfy the needs of industry.
Recently, due to great improvement provided by pre-trained models (e.g., BERT [25], T5 [26]), some researchers designed pre-trained models specifically for code representation. Feng et al. [27] proposed pre-trained model CodeBert which considers the representation of the data flow of code, achieving promising results in the code summarization task. Furthermore, Wang et al. [28] proposed CodeT5 based on T5 [26] considering the token type of code and utilizing a denoising sequence-to-sequence pre-training strategy. The above code pre-trained models generally understand deeper relationship within codes by feeding large enough data, and the results are pretty impressive.
Overview
The overview of our proposed method(SSCS) is shown in Figure 1. Figure 1. Overall process of SSCS. A model with two unique encoders for sequence information and structure information is followed by a bidirectional decoding module for left-to-right decoding and right-to-left decoding.
To detail our proposed method, we separate the whole process into two stages, training stage and inference stage. We can clearly see that SSCS contains three main components encoder, fusion module and decoder, which will be detailed in the following section.
As shown in Figure 1, code and ground-truth summary are both utilized in the training stage, while only code will be utilized in the inference stage. First the code will be processed into AST and lines of code sequence for the AST encoder and hierarchical sequence encoder, respectively, followed by a fusion module. Later the fusion output will be used by the decoder. The main difference between the training stage and inference stage is the process of the decoder. In the training stage, the ground-truth summary will be fed into the decoder and computes the loss with the decoder output for backward propagation, while the input is the generated token from the last step in the inference stage.
Data Preprocessing
The dataset we use for our work contains large <code, summary> pairs. We simply tokenize the summary into a list of tokens. Furthermore, to reduce out-of-vocabulary issues caused by a large scale of unique tokens, we tokenize any CamelCase or snake_case defined by developers. The process of acquiring sequence information for code is same as the summary but we maintain the document property of code. Thus, the code sequence information consists of multiple tokenized lines of code sub-sequence.
Knowing that SSCS is a structure and sequence aligned model, we also need to parse code into an AST to obtain structural information. In this paper, we evaluate our method on two public Java and Python datasets. We generate Abstract Syntax Trees (ASTs) with opensource tool tree-sitter (https://tree-sitter.github.io/ (accessed on 15 July 2022)) for Java code and ast (https://github.com/python/cpython/blob/master/Lib/ast.py (accessed on 26 July 2022)) module for Python code. Moreover, we also add extra data flow information to strengthen the structural information. As shown in Figure 1, the variable sum is computed from addition of a and b, and we can see in the Code AST that the leaf node sum has data relationships with variable a and variable b, respectively. Finally, we obtain the node sequence by pre-order traversal and its adjacency as the AST encoder inputs.
Token and Node Embedding
Before sending the node and token into the model, we need to vectorize the token and node. We first create a dictionary to calculate the total number of tokens and nodes separately, and use num token and num node to represent them. Thus, we are able to vectorize the token or node as one-hot vector, which is a num token or num node size vector consisting of a unique "1", and "0"s for the rest of the positions. To accelerate the computation, we usually embed the discrete one-hot vector into dense vector of size dim. The most simple way is to multiply a matrix with a size of numxdim, and we are able to transform the vector space R num into dense vector space R dim .
AST Encoder
Wu et al. [9] proposed the SiT model which used transformer architecture to directly capture the structural information instead of using Graph Neural Networks that inspired our work. Inspired by SiT, we designed a transformer-based AST encoder capturing the structural information in multi-view. From the last section, we know that two inputs will be imported into the AST encoder, so we first define each input. Given an AST with L nodes N = {n 1 , n 2 , n 3 , . . . , n l }, where n j ∈ R dim denotes each node vector and dim denotes the dimension of node vector in vector space R. A denotes the adjacency matrix in the shape of LXL, the computation process of the AST encoder can be split into three blocks: multi-view attention computation, adapted weight fusion and feed-forward network. Figure 2 shows the overview of an AST encoder.
Global Self-Attention The computation of the global self-attention is based on the vanilla self-attention in transformers [7]. We treat the AST of the source code as an undirected complete graph, which means a node ni can learn the relation from the whole tree without any blocking. Therefore, we are able to capture the global AST representation. The global self-attention mechanism is denoted as follows: where N = {n 1 , n 2 , . . . , n l } denotes the input sequence of nodes, l denotes the node sequence length and d k is the dimension of K. W Q , W K and W V , W Q , W K , W V ∈ R dim×dim are three learnable matrices using as projection to transform the vector space into a different vector space. NW Q , NW K and NW V represent matrix multiply operation (matrix N dot matrix W).
Structure-induced Self-Attention
We follow the previous work by Wu et al. [9] to represent the structure information using transformer architecture equipped with a special attention mechanism. The structure-induced self-attention network (Si-SAN) is able to capture the structural information instead of using Graph Neural Networks (GNNs). The computation of the Si-San is to multiply the adjacency matrix by key-query pairs: where A denotes the adjacency matrix of the code. edge(i, j) denotes there is an edge between n i and n j . The attention score between n i and n j will be dropped out when Local Self-Attention To further capture the structural information, we also adopt a local attention network to capture the local information. By adding a window mask, we can initialize a window which can slide through the whole tree to learn the local relation.
where M win denotes the window matrix for constraining the computation of node pairs in window distance. Adaptive Weight Fusion Layer In SiT [9], the process of the encoder module is a global self-attention network followed by a structure-induced self-attention network, where the global information will be diluted by the Si-SAN. Thus, we adopt a superior fusion strategy for the different views of information by using an adaptive weight fusion layer, which is shown in Figure 2.
For vector G , G obtains the relation weights by summing up the vectors after matrix multiplication with the rest of the two dense vectors and S and L repeat the same process. To simplify the computation, we join G , S and L and compute the dot value between the joint matrix and its transpose matrix. We sum up the relation weights and normalize as the adaptive weight for G . S and L repeat the same process to obtain the adaptive weight for themselves. The computation process is shown below.
The final AST encoder output is the weighted sum of G, S and L followed by a Feed-Forward network.
where x denotes the output from the adaptive weight fusion layer, W 1 , W 2 are two learnable matrices. For activation function ReLU, ReLU(x) = max(0, x). During attention computation stage, the QK T operation creates a square matrix, so we can utilize mask strategy to control the reception field. To better understand why the mask matrices are able to control the receptive field allowing the AST encoder to capture multi-view information, we visualize three kinds of mask matrix, two are human-defined (global mask and window mask) and one (structure-induced mask) is from the AST in Figure 3. The global mask (Figure 3a) is a matrix filled with "1", allowing the node sequence to construct a fully connected graph to capture global information (can be omitted). Each node is able to study from the rest of them. For the structure-induced mask (Figure 3b), each node studies according to the adjacency extracted from the AST and only studies from the node with an edge connection. The window mask is a special mask simulating the sliding window. As shown in Figure 3c, we take a window with size 2 as example. As the window is sliding forward, we are able to capture local information at each window.
Hierarchical Sequence Encoder
A code snippet is some kind of document which consists of several statement sequences. To maintain the document property instead of treating a code snippet as a single sequence, we adopt a hierarchical encoding strategy which captures the code sequence information from token-level to sentence-level.
Given lines of code sub-sequences S = {s 1 , s 2 , s 3 , . . . , s n }, s i ∈ R m×dim , where m denotes the max sequence length between all the sub-sequences and n denotes the line number, we are able to capture the hierarchical information using a hierarchical sequence encoder. The procedure is shown in Figure 3. First, we use the self-attention mechanism to capture the relation between the tokens in each sub-sequence. The first step output o i can be formulated as follows: where s i = {s 1 , s 2 , ..., s m }, s j ∈ R dim is the token representation. Second, to obtain the sentence-level representation, we send the first step output into a long short-term memory network (LSTM), which is able to condense the token-level representation into the sentence-level representation and capture the position information. The final layer of the hidden state in each sub-sequence computation stage is used as the sentence-level representation for each sub-sequence.
where Rep i token ∈ R m×dim , h i ∈ R 1×dim is the final layer hidden state in the mth time step generated by LSTM.
Then, in the same way as step one, we adopt the self-attention mechanism to capture the relation between sub-sequences.
Finally, the second LSTM network has the same effect as the first LSTM network. We use all the final layer hidden states generated in each time step as the code sequence representation.
In Figure 4, we also represent the change of the vector shape in each step to better represent the hierarchical encoding process. At first, the shape of input lines sequence vector is 3D, which has three dimensions ([Line, Length, Dim]), and after the word-level self-attention and LSTM, the vector is compressed into 2D ([Line, Dim]). The final shape remains 2D after the sentence-level self-attention and LSTM modules. Thus, we transform 3D lines of sequence vector into a 2D vector for later combination.
Encoder Output Fusion
In above section, we utilized two encoders for capturing structural information and sequence information. We define the output from the AST encoder as Encoder AST (N, A) and Encoder S (S) from the hierarchical sequence encoder.
The SSCS encoder output is obtained by jointing the outputs from the AST encoder and hierarchical code sequence encoder. The computation is as follows: where Encoder AST denotes the AST encoder and Encoder S denotes the hierarchical code sequence encoder. N refers to the input node vector and A is adjacency. S refers to the lines of code sequence.
The encoder output will be utilized in the decoding stage for generating the summary.
Bidirectional Decoder
Liu et al. [29] found that the quality of the prefixes of translation hypotheses is much higher than that of the suffixes in machine translation tasks. Furthermore, in order to produce more balanced translations, Liu et al. adopted a simple strategy for joint training the forward decoder and the backward decoder.
Inspired by the previous work of Liu et al., we adopt the same strategy to produce more balanced predictions. We simultaneously generate sequence in the Left-to-Right (L2R) direction and Right-to-Left (R2L) direction. Both directions guide each other by optimizing the shared parameters which are learnable. We define the input vector for the L2R decoder as .., ← − y n } for the R2L decoder. The L2R decoder and R2L decoder can be regarded as two sub-tasks. By optimizing the shared parameters, one task can guide the other one. The details are presented below.
The L2R decoder and R2L decoder share the same Embedding Layer, which converts the one-hot vector of the token into a dense vector: where Embedder(x) = xW, W is a learnable matrix, a converting discrete vector into a continuous vector. PE denotes position encoding operation, used for capturing sequence order. We define Query, Key, Value in different directions as ( . The two directions' multi-head attention outputs can be computed as below: After the attention computation, the outputs will be sent into a Feed-Forward network, which contains two linear transformations and a ReLU activation function.
By using the So f tmax function, we can obtain the probability of the generated token.
In the training stage, the bidirectional guiding decoder generates two directions' outputs. Our goal is to find the parameter θ which can maximize the likelihood of success. For a training data pair {x n , − → y n , ← − y n } N t=1 , x n is the encoder output, − → y n and ← − y n are the decoder inputs. The generation procedure is auto-regressive which means the generation for the next token is based on the previous all generated tokens. We compute the likelihood as follows: The final likelihood function is the joint of the likelihood functions of two sub-tasks: where λ is the weight to balance the guidance of the two sub-tasks. λ decides the training purpose to focus on which direction task more. j denotes the time step and J denotes the whole generating step length. Different from the training stage, we adopt a beam search strategy for the summary generation. Beam search strategy expands the searching area reserving the best top k token (k denotes the beam size) instead of shrinking the best one as in a greedy search. For each inference step, we generate beam-size candidate sequences and reserve k-best at last. The final sequence is chosen between the two outputs from the L2R decoder and the R2L decoder. Furthermore, we can regard greedy search as a beam search strategy when beam size is 1.
Datasets
Two widely used public datasets are taken into consideration to ensure the performance and generalization ability of SSCS. The Java dataset provided by Hu et al. [6] consists of 87,136 <code, summary> pairs, which is able to test the generation performance on Java language. Furthermore, we also test on a Python dataset provided by Wan et al. [1] containing 92,545 <code, summary> pairs. We follow the procedures of Hu et al. and Wan et al. and divide the datasets into training set, valid set and test set. Table 1 shows the statistics of both datasets including the size for each set and the average length of code and summary in each program language.
Metrics
We evaluate the performance of summarization with BLEU [30], METEOR [31] and ROUGE [32], which are widely used for testing the sentence generation performance in natural language generation (NLG) tasks.
BLEU metric is used to calculate the number of n-gram matches between the generated sequence and the reference sequence and calculate the average, n = 1, 2, ..., N. The calculation formula is as follows: where ∑ gram n ∈s match(gram n ) denotes the number of n-grams matches in the generated sequence and the reference sequence c. N(gram n ) denotes the total number of n-grams in the reference sequence, and ρ is the brevity penalty. ROUGE is a recall-based evaluation metric. We use ROUGE-LCS (Longest Common Sequence, ROUGE-L) as our ROUGE evaluation metric in our experiments. ROUGE-LCS takes the longest common sub-sequence between the generated sequence and the reference sequence as the starting point for calculation. The calculation formula is shown below: where LCS(g, c) denotes the longest common sub-sequence between the generated sequence g and the reference sequence c. The size of β determines whether to focus on the Recall rate or Precision rate. METEOR is proposed to solve some inherent defects of BLEU. It uses Word Net to compute specific sequence matches, synonyms, root words and affixes, and paraphrases to make them more relevant to manual judgment.
Hyper Parameters
For the hyper parameter settings, we follow the previous works by Ahmad et al. [8] and Wu et al. [9]. We set embedding size of source code and summary to 512. The layer for the AST encoder is set to 3 and 6 for the summary decoder. We initialize the learning rate as 1 × 10 −4 and use a 4000 step warm-up schedule. The maximum training epoch is set to 200 with an early stop mechanism. The maximum length for code is set to 300 and 100 for the summary. Adam optimizer is used for the optimization of the learning rate. We detail the parameter in Table 2. The experiments are conducted in a server with 4 Nvidia 2080ti GPUs and Ubuntu 18.04 OS (https://www.ubuntu.org.cn/, accessed on 29 January 2023).
Baselines
We compare SSCS with the recent code summarization models and the description of each model is shown as below.
• RL+HybridSeq [1] using a critic network with the BLEU score as the reward to conduct the learning of the model. • DeepCom [13] using SBTs traversed from ASTs and code sequence as inputs and using a hybrid attention mechanism to fuse these features. • API+CODE [6] utilizing the API sequence to enhance the representation of code that improved the performance of generating the summary. • Dual Learning [19] utilizing the duality between code generation task and code summarization task and training both tasks simultaneously. • Transformer [8] using a transformer with a copy mechanism and relative positional encoder to generate code summaries. • SiT [9] constructing a multi-view adjacent matrix to represent the relationships between the tokens in the code guiding the self-attention computation. • M2TS [10] constructing a multi-view AST feature at multiple local and global levels and proposing a fusion method to combine sequential information and structural information. • SCRIPT [11] introducing the structural relative positions between nodes of the AST to better capture the structural relative dependencies. • CodeScribe [12] introducing a novel triplet position for AST which is represented by GNNs and using a pointer-generator network to copy tokens from code tokens and tree nodes to summarize.
Results and Analysis
To represent the performance of SSCS, we compare it with eight state-of-the-art baselines. The baseline results are mainly from Ahmad et al. [8] and Wu et al. [9] and the others are from the original papers. The overall result is illustrated in Table 3. We split the baselines into two groups, one group is the RNN-based baselines, while the others are the transformer-based baselines. Compared with the RNN-based baselines (RL+Hybrid, Deepcom, API+CODE, Dual Model), SSCS is much more superior to them in all evaluation metrics on the Java and Python datasets. Although the recent transformerbased baselines have achieved excellent performance in the code summarization task, SSCS can still perform better than these approaches. Compared with M2TS, SSCS improves the performance of BLEU, METEOR and ROUGE-L by 2.94%, 2.19% and 2.47% on the Java dataset, respectively. Meanwhile, SSCS also exceeds M2TS by 3.64%, 1.56% and 4.43% on the Python dataset. We also conduct a comparison between SSCS and SCRIPT and the result also demonstrates the effectiveness of the SSCS. SSCS improves the performance of BLEU, METEOR and ROUGE-L by 2.89%, 2.64% and 3.65% on the Java dataset and 3.48%, 2.55% and 4.24% on the Python dataset.
Due to the different summary processing strategy between CodeScribe and the other baselines, we compare SSCS with CodeScribe in isolation. CodeScribe replaces all the numerical tokens with a unified symbol '< number >' and removes all the lexical forms (e.g., -s, -es, -ed). The result is shown in Table 4. In the same preprocessing method for the summary, we can see that SSCS still performs better than CodeScribe on the Java and Python datasets. SSCS improves results by 1.36%, 0.96% and 2.91% in BLEU, METEOR and ROUGE-L scores on the Java dataset and 2.20%, 1.22% and 3.39% on the Python dataset. Table 4. Comparison of our proposed approach with CodeScribe. * refers to experiment on the summary processed by Guo et al. [12]. Compared with SSCS, SSCS * takes a cleaner summary which removes all the lexical forms as decoder input. From the above experiment results, our approach outperforms the current state-ofthe-art methods. To better understand the main reasons of the improvement for SSCS, we conduct an ablation study to present the strength of each module we propose in the following section.
Ablation Study
To better understand why our proposed SSCS can achieve such a great performance, the ablation study is necessary for disclosing the main reasons for such improvement.
We first conduct two ablation experiments that remove the important components in SSCS. Thus, the full SSCS model degenerates into the extended Si-Transformer model. By studying the ablation of components, we can directly see the change of scores.
As shown in Table 5, we first remove the bidirectional decoder, which means we only generate the summary from a single direction.Based on this, the performance drops about 2%, 1.2% and 1.4% in BLEU, METEOR and ROUGE-L. Then we remove the sequence encoder and the generation of summary only depends on the structural information. The performance drops about 0.6%, 1% and 1.1% in BLEU, METEOR and ROUGE-L. The ablation study on components illustrates the effectiveness of each component. Furthermore, the bidirectional decoder contributes much in generating better summaries. The hierarchical encoder provides a view at document level, which helps to represent the code comprehensively. IO the left side of Table 5 "w/o" denotes "without" and also means we remove this component in our model. The top row is the full model, so we remove each component step by step. Moreover, we also conduct experiments of the different fusion strategies for the fusion of the global information, structural information and local information in the AST encoder. The fusion strategies are addition, element-wise dot-product, average and adaptive weight fusion. The result is shown in Table 6, the adaptive weight fusion achieves the best score compared with the other strategies. The element-wise dot-product achieves the lowest score, while performance of the addition strategy is close to the average strategy. It can be obviously seen from the results that the adaptive weight fusion strategy does improve the generation performance.
Validation Performance
To demonstrate the superiority of SSCS, we visualize the validation curve of BLEU and ROUGE scores compared with several baseline. Our approach is based on the transformer architecture; thus, we choose Si-transformer and vanilla transformer as contrasts. The results also contain the generation performance on different generating direction. L2R denotes generating the summary from left-to-right (normal writing order) and R2L denotes the opposite. It is obvious that SSCS achieves a higher score of both languages in the validation set from the first epoch to the last one, which also demonstrates the superiority of SSCS. Figure 5 shows the qualitative examples of SSCS, SiT, Transformer-L2R (Left-to-Right) and Transformer-R2L (Right-to-Left). It can be observed that Transformer-L2R can generate prefixes well but suffixes poorly, while Transformer-R2L achieves the opposite performance. By utilizing the potential of encoder and decoder, SSCS can generate summaries with balanced prefixes and suffixes. Compared with the SiT, our approach achieves better performance in both Java and Python languages. In general, SSCS is able to generate a more complete and accurate summary. We can clearly see from example 1 (upper left), suffering from exposure bias, L2R and R2L stop generating the summary almost in the middle of the summary, which means the stop symbol '<EOS>' comes up early. However, our proposed approach can overcome the generation hindrance, preventing the stop symbol coming up early, and allowing it to generate the whole summary. bytesEqual ( expected , actual ) ) ; } assertEquals ( ( byte ) _NUM , buf . getByte ( _NUM ) ) ; assertEquals ( ( byte ) _NUM , buf . getByte ( pos + capacity -_NUM ) ) ; } Reference: returns a new sector whose angles are the extremes of the this s ector and another . the new sector ' s minimum latitude and longitude will be the minimum of the two sectors . the new sector ' s maximum latitude and longitude will be the maximum of the two sectors . the sectors are ass umed to be normalized to + / -90 degrees latitude and + / -180 degrees longitude . the result of the operation is undefined if they are not .
Case Study
L2R:returns a new sector whose angles are the extremes of the this s ector and another . the new sector ' s minimum latitude and longitude will be the minimum of the two sectors . the new sector ' s maximum latitude and longitude will be the maximum of the two R2L:sector' ' s maximum latitude and longitude will be the maximum of the two sectors . the sectors are assumed to be normalized to + / -90 degrees latitude and + / -180 degrees longitude . the result of the operation is undefined if they are not .
SiT: returns a new sector whose angles are the extremes of the this s ector and another . the new sector ' s maximum latitude and longitude will be the maximum of the two EASCS: returns a new sector whose angles are the extremes of the this s ector and another . the new sector ' s minimum latitude and longitude will be the minimum of the two sectors . the new sector ' s maximum latitude and longitude will be the maximum of the two sectors . the sectors are ass umed to be normalized to + / -90 degrees latitude and + / -180 degrees longitude . the result of the operation is undefined if they are not .
Conclusions
In this paper, we propose a structure and sequence aligned code summarization model named SSCS which can achieve excellent performance compared with several code summarization baselines. SSCS extends the previous work and utilizes both structural information and sequence information. We design a multi-view mask strategy which enables transformer architecture to capture the AST more comprehensively. Inspired by the works on NMT, we first introduce bidirectional decoding into the code summarization task to release the exposure bias issue, which can generate better summary with both prefixes and suffixes. However, the performance of the automatic code summarization is far from satisfactory; it is still a tough mission to generate high-quality summaries. In future work, we will be devoted to exploring the potential of the encoder and the decoder for the code summarization task. For example, we plan to consider to transform SSCS into a large pre-trained language model. With the great performance achieved by the pre-training strategy, we believe the results will be more satisfying. Furthermore, there is still a long journey for automatically generating a high-quality summary of code.
Limitations
We have identified the following limitations to our work that may threaten the validity of our work: • Baselines Reproduction. Due to hardware limitations, we cannot reproduce all the baseline methods (e.g., CodeScribe, M2TS). Thus, we process the data using their released tools or use their processed data, and set most of the hyperparameters (e.g., max source length, max target length, max epoch) the same as theirs.
• Language Type. We only conduct our experiments on Java and Python, but it is necessary to experiment on other popular languages (e.g., C++, C#, SQL, Rust). We do not know whether the SSCS can achieve the same promotion as on Java and Python. • Evaluation Metrics There are no particular evaluation metrics for code summarization tasks. We follow the previous works on this task evaluating our approach using the metrics in the machine translation task or text summarization task. It is necessary to propose a metric particular for code summarization tasks. | 9,241.4 | 2023-03-26T00:00:00.000 | [
"Computer Science"
] |
Native RNA Purification Method for Small RNA Molecules Based on Asymmetrical Flow Field-Flow Fractionation
RNA molecules provide promising new possibilities for the prevention and treatment of viral infections and diseases. The rapid development of RNA biology and medicine requires advanced methods for the purification of RNA molecules, which allow fast and efficient RNA processing, preferably under non-denaturing conditions. Asymmetrical flow field-flow fractionation (AF4) enables gentle separation and purification of macromolecules based on their diffusion coefficients. The aim of the study was to develop an AF4 method for efficient purification of enzymatically produced antiviral small interfering (si)RNA molecules and to evaluate the overall potential of AF4 in the separation of short single-stranded (ss) and double-stranded (ds) RNA molecules. We show that AF4 separates monomeric ssRNA from dsRNA molecules of the same size and monomeric ssRNA from multimeric forms of the same ssRNA. The developed AF4 method enabled the separation of enzymatically produced 27-nt siRNAs from partially digested substrate dsRNA, which is potentially toxic for mammalian cells. The recovery of AF4-purified enzymatically produced siRNA molecules was about 70%, which is about 20% higher than obtained using anion-exchange chromatography. The AF4-purified siRNAs were not toxic for mammalian cells and fully retained their biological activity as confirmed by efficient inhibition of herpes simplex virus 1 replication in cell culture. Our work is the first to develop AF4 methods for the separation of short RNA molecules.
Introduction
Small RNA (sRNA) molecules (<200 nucleotides, nt) that do not code for any proteins, but rather possess regulatory functions, have been found in cellular organisms representing all domains of life, and are also encoded by some viruses [1][2][3][4][5]. Despite many sRNAs having already been discovered and characterized both structurally and functionally, new classes of sRNA, specific for a particular organism, tissue, or (patho)physiological conditions are still to be described [6]. Among the most abundant non-coding RNAs are ribosomal and transfer RNAs, which are engaged in protein synthesis; small nuclear RNAs, mostly involved in splicing of messenger (m)RNAs; and small nucleolar RNAs, which function in chemical modifications of various classes of RNAs [7]. MicroRNAs (miRNAs) and small interfering (si)RNAs are double-stranded (ds) non-coding RNAs with two-base 3 -terminal overhangs processed from longer dsRNA molecules by Dicer [8]. The size of siRNAs and miRNAs depends on the organism and is typically in the range of 21-27 nt [9][10][11]. These short dsRNAs have a significant role in cell physiology and multiple disease pathogenesis (see, for example, [12][13][14][15][16]). The mechanism of siRNAs and miRNAs action is based on their total processing time and reduces siRNA yields. Consequently, alternative advanced native RNA purification methods applicable to a wide range of ds-and ssRNA molecules are still in demand.
Field-flow fractionation (FFF) is a versatile separation technique, where separation occurs in a thin channel without a stationary phase. In asymmetrical flow FFF (AFFFF or AF4) a carrier liquid with analytes is pumped through a separation channel with a semi-permeable bottom part. The main incoming flow splits into longitudinal channel flow, which is going to the detector, and the cross-flow, which is perpendicular to the main channel flow, pushing the sample components towards the semi-permeable accumulation wall. Sample components diffuse against the cross-flow force depending on their diffusion coefficients and, hence, small sample components diffuse farther from the accumulation wall than do large particles with small diffusion coefficients. Therefore, smaller particles diffuse closer to the center of the separation channel, where the velocity of the channel flow is maximal, and elute first. In AF4, separation is based solely on the hydrodynamic sizes and is potentially achievable in any mobile phase. Previously, AF4 has been used to separate ss-and dsDNA molecules of different sizes [50][51][52] and free ssDNA from proteinbound ones [53]. The technique has also been used to separate and quantify Escherichia coli ribosomal and transfer RNAs [54,55]. Furthermore, AF4 has been used to characterize the size distribution of lipid nanoparticles loaded with siRNAs [56,57] and separate various miRNA carriers in the serum [58]. In these studies, however, the focus has been on the particles and the carriers and not on naked si-or miRNAs.
In the present work, we addressed the separation and purification of naked sRNA molecules using AF4. We optimized the AF4 conditions, including flow rates and buffer composition, to achieve separation and purification of enzymatically produced ss-and dsRNAs as well as siRNAs. The 88-nt-long ssRNAs and dsRNAs were separated well from each other with the developed AF4 method. Longer ssRNA molecules of 108 nt were prone to form intermolecular complexes, which could be separated from the monomeric form under non-denaturing conditions. Using AF4, enzymatically produced siRNAs were efficiently purified from partially digested dsRNA substrate. The quality of the AF4 purified enzymatically produced siRNA molecules was verified by evaluating the toxicity and biological activity of the resulting siRNAs in cell culture experiments. This demonstrated that siRNAs purified in buffered mobile phase containing a small amount of salt (20 mM NaCl) did not contain toxic, longer dsRNA, and fully retained biological activity.
Results
We studied five types of enzymatically produced sRNA molecules for their elution behavior and purification with AF4: 88-nt and 108-nt-long ssRNAs, 88-nt and 108-nt long dsRNAs, and 27-nt-long siRNAs derived from the UL29 gene of herpes simplex virus 1 (HSV-1), enhanced green fluorescent protein (eGFP) gene, or bacteriophage phi6 genome (Table 1; Figure 1; see Section 4.1.1 for details). DNAse-treated ssRNA and dsRNA synthesis reaction mixtures were directly used for AF4 injections, or the samples were pre-purified before the AF4 analysis ( Figure 1; see also Section 4.1.1). The short dsRNAs studied here are strong immunostimulators in human monocyte-derived dendritic cells, macrophages, glioma, and epithelial cells [31,59,60]. Furthermore, the UL29 siRNAs show antiviral activity in HSV-1 infected cell cultures and promising results have been obtained in mouse studies [31,[61][62][63].
The folding of ssRNA affects its hydrodynamic radius and, thus, elution behavior in AF4. The biochemical and physical properties of the produced sRNAs, including parameters related to the predicted ssRNA structures [ Figure S1 (88-nt RNA) and Figure S2 (108-nt RNA), Supplementary Materials], are described in Table 1. In our experiments, we used two AF4 instruments currently available on the market, AF2000 MT (Postnova Analytics) and Eclipse NEON (Wyatt Technology). The channel widths were 350 µm and 400 µm, respectively. The mobile phase was 20 mM NaCl, 50 mM Tris-HCl (pH 8.0), unless otherwise indicated. The AF4 methods developed for the RNA separation with AF2000 MT and Eclipse NEON are presented in Tables S1 and S2 (Supplementary Materials), respectively.
The membrane chemistry and pore size at the accumulation wall determine the dimensions of sample components that retain in the AF4 channel for separation. The regenerated cellulose (RC) membrane was a natural choice since its isoelectric point is around 3.4 and, hence, it is negatively charged at higher pH values [66] generating repulsive forces between the membrane and negatively charged RNA molecules in the mobile phase with pH 7.0-8.0.
Previous AF4 studies on protein-DNA aptamer interactions demonstrated that folded and non-folded poly(A) oligomers shorter than~30 nt were poorly retained when ultrafiltration membranes with molecular weight cut-off (MWCO) of 5 or 10 kDa were used at the accumulation wall [67]. Therefore, initially, we chose to use RC membrane with a 1 kDa MWCO to maximize the retention and thus recovery of sRNA molecules. We found that with the channel-flow rate of 0.5 mL/min, the maximal applicable cross-flow rate was approximately 2.3 mL/min, which was significantly lower than the cross-flow rate routinely achieved with a typical 10 kDa RC membrane (3.5 mL/min and larger). With the Eclipse NEON system, we used 10 kDa RC membrane allowing the use of 4 mL/mL cross-flow velocity (Table S2). With this system, we could record multi-angle light scattering (MALS) and concentration data [refractive index (RI) and UV] enabling deduction of molecular weight (M w ) for separated sample components.
Elution of 88-and 108-nt ssRNA Molecules
The purified 88-nt ssRNAs eluted as a single peak, and the agarose gel analysis of the peak fractions showed a uniformly migrating RNA species (Figure 2A,B). The average mass recovery of the 88-nt ssRNA in the peak fraction was 85 ± 2.5%. The homogeneity and monomeric nature of the eluting ssRNA were confirmed by collecting online light scattering data using the Eclipse NEON multidetector setup that showed M w of 35.0 kDa ± 0.7% throughout the peak ( Figure S3A, Supplementary Materials). The difference between the root mean square radii of gyration (R g ) for the two predicted conformers of the 88-nt ssRNA was too small (Table 1) for the molecules to be separated with AF4. The 108-nt ssRNA species eluted as two peaks ( Figures 2C, S3B, and S4A, Supplementary Materials), which suggested conformational and/or size heterogeneity in the sample. Secondary structure predictions indicated that the 108-nt ssRNA molecule could potentially form more folds than the 88-nt molecule ( Figures S1 and S2), and that the different isoforms have relatively large size differences (Rg) (Figures S1 and S2; Table 1). Sample components of the early-eluting fractions of the pre-purified 108-nt ssRNA sample migrated predominantly as a single band in the agarose gel ( Figure 2D). However, multiple bands were observed in the later eluting fractions. Based on the mobilities in agarose gel, we speculate that these larger RNA species are formed from intermolecular interactions between ssRNA molecules. Accordingly, the MALS data collected during AF4 separation showed that the Mw for the RNA in the second peak, 87.9 kDa ± 0.2%, was The 108-nt ssRNA species eluted as two peaks ( Figures 2C, S3B and S4A, Supplementary Materials), which suggested conformational and/or size heterogeneity in the sample. Secondary structure predictions indicated that the 108-nt ssRNA molecule could potentially form more folds than the 88-nt molecule ( Figures S1 and S2), and that the different isoforms have relatively large size differences (R g ) (Figures S1 and S2; Table 1). Sample components of the early-eluting fractions of the pre-purified 108-nt ssRNA sample migrated predominantly as a single band in the agarose gel ( Figure 2D). However, multiple bands were observed in the later eluting fractions. Based on the mobilities in agarose gel, we speculate that these larger RNA species are formed from intermolecular interactions between ssRNA molecules. Accordingly, the MALS data collected during AF4 separation showed that the M w for the RNA in the second peak, 87.9 kDa ± 0.2%, was comparable with that of the purified dsRNA molecule, 85.1 kDa ± 0.1% ( Figure S3B). This rules out the possibility that the observed second peak would represent different conformers of the monomeric ssRNA molecule. The single-stranded nature of the RNA was further verified by RNase sensitivity test ( Figure S4C) confirming that the molecules in the second peak are multimers of ssRNA molecules and not dsRNA. Thus, AF4 could be used to fractionate monomers from the different oligomeric forms of the 108-nt ssRNA. The average mass recovery of the pre-purified 108-nt ssRNA in the first peak fraction was 83 ± 1.6%.
An additional peak (peak 1; Figure 2A,C) was observed in the fractograms of the crude reaction mixture (T7 ssRNA synthesis reaction mixtures), but no RNA was observed in the corresponding fractions by agarose gel electrophoresis analysis ( Figure 2B,D). The position of this peak corresponds to the position in which NTPs elute (see Section 2.2.2 and Figure S6B). Importantly, no NTP peak was observed when ultrafiltration membrane with MWCO of 10 kDa was used ( Figure S4) since only a tiny fraction of NTP can be retained under these conditions (see Section 2.2.2). Otherwise, the fractograms of the purified and unpurified ssRNAs were similar ( Figure 2).
Mobile Phase Composition Affects ssRNA Elution
Cations present in the mobile phase have a strong impact on RNA folding, intermolecular interactions as well as on the interactions of RNAs with the ultrafiltration membrane and, thus, potentially affect the elution of ssRNA molecules [68,69]. Therefore, we compared behavior of the 108-nt ssRNA in two mobile phases: (i) in 20 mM NaCl, 50 mM Tris-HCl (pH 8.0) and (ii) in 10 mM Tris-HCl (pH 8.0). The lower ionic strength of the mobile phase accelerated the elution of ssRNA molecules. In addition, the peak was broader and had an extensive tail ( Figure S5A, Supplementary Materials). The intensity and number of the slowly migrating higher molecular weight bands in agarose gel were reduced when AF4 fractionation was carried out using mobile phase without NaCl compared to the standard conditions containing 20 mM NaCl ( Figure S5B,C). The negative charge of RNA in a buffer with 20 mM NaCl is at least partially shielded and, hence, repulsive interactions are reduced [69]. At the same time, short-range interactions, such as hydrogen bonding, Van-der-Waals, or hydrophobic forces between molecules and RC membrane at the accumulation wall are probably promoted, which translates into enhanced retention and formation of RNA intermolecular complexes. Base pairs formed on ssRNA are not stable at low ionic strength, and ssRNAs form flexible polymers with different conformations [70], which together with the enhanced interactions with RC membrane may explain the peak broadening for ssRNAs. In summary, proper optimization of the mobile phase composition can improve separation and increase the yield of ssRNAs.
AF4 Provides Good Separation between Small ssRNA and dsRNA Molecules in Native Conditions
AF4 fractionation of the 88-nt and 108-nt dsRNA species resulted in similar fractograms: one or two peaks were observed for the pre-purified molecules and three for the unpurified dsRNA samples. In the AF2000 MT instrument, the 88-nt dsRNA eluted between 11 and 15 min, and the 108-nt dsRNA between 12.5 and 16.5 min ( Figure 3A,C). In addition, the pre-purified dsRNA samples included a minor low-intensity peak, which most probably originated from ssRNA molecules incompletely precipitated with 2M LiCl, since the retention times correspond to those observed for ssRNA species (Figure 2A-D). These ssRNA contaminates were not observed by agarose gel electrophoresis analysis of the initial input sample which emphasizes the potential of AF4 for RNA quality control. The observation also indicates that AF4 fractionation can improve the separation of ss-and dsRNA molecules compared to the conventional LiCl precipitation method. The average mass recovery of AF4 purified 88-nt dsRNA was 66 ± 2.1% of the input pre-purified sam- ple. The average mass recovery of AF4 purified 108-nt dsRNA was 78 ± 3.8% of the input pre-purified sample, respectively. purified sample. The average mass recovery of AF4 purified 108-nt dsRNA was 78 ± 3.8% of the input pre-purified sample, respectively. The AF4 fractogram of the crude reaction mixture displayed a larger ssRNA peak than observed for the pre-purified sample and an early peak at ~2 min ( Figure 3A,C), at the position where NTPs elute (see Section 2.2.2; Figure S6B, Supplementary Materials). Agarose gel analysis revealed that the content of the dsRNA peak is homogenous in respect of RNA size. This was supported with the Mw distribution data from the MALS measurements ( Figure S3.) However, considering the elution behavior of the 108-nt ssRNA multimers ( Figure 2C,D), there is a possibility that the 108-nt dsRNA fraction ( Figure 3C) contains traces of ssRNA, although such ssRNA is barely detected by the agarose gel analysis ( Figure 3D). This observation underlines the importance of analyzing all the sub-components of the analyte separately. The AF4 fractogram of the crude reaction mixture displayed a larger ssRNA peak than observed for the pre-purified sample and an early peak at~2 min ( Figure 3A,C), at the position where NTPs elute (see Section 2.2.2; Figure S6B, Supplementary Materials). Agarose gel analysis revealed that the content of the dsRNA peak is homogenous in respect of RNA size. This was supported with the M w distribution data from the MALS measurements ( Figure S3.) However, considering the elution behavior of the 108-nt ssRNA multimers ( Figure 2C,D), there is a possibility that the 108-nt dsRNA fraction ( Figure 3C) contains traces of ssRNA, although such ssRNA is barely detected by the agarose gel Figure 3D). This observation underlines the importance of analyzing all the sub-components of the analyte separately.
AF4 Separates siRNAs and Partially Digested Dicer Reaction Products
SiRNA molecules pre-purified with AEX eluted as one major peak, which was clearly distinct from the void volume, thus enabling the proper separation of the siRNAs from the potential non-retained sample components ( Figure 4A,C). In contrast to the fractograms of the pre-purified siRNA molecules, the UV signal for the unpurified siRNA preparation produced by Dicer digestion from long dsRNA did not reach baseline level after the main peak eluted (compare Figure 4A-D). Agarose gel analysis revealed that these lateeluting fractions contained multimers of siRNA molecules ( Figure 4B,D). All samples had components that eluted with zero cross-flow force, indicating a large molecule size. Such components were more abundant in the non-purified samples and might include remnants of dsRNA substrate that were not completely processed ( Figure 4). SiRNA molecules pre-purified with AEX eluted as one major peak, which was clearly distinct from the void volume, thus enabling the proper separation of the siRNAs from the potential non-retained sample components ( Figure 4A,C). In contrast to the fractograms of the pre-purified siRNA molecules, the UV signal for the unpurified siRNA preparation produced by Dicer digestion from long dsRNA did not reach baseline level after the main peak eluted (compare Figure 4A-D). Agarose gel analysis revealed that these late-eluting fractions contained multimers of siRNA molecules ( Figure 4B,D). All samples had components that eluted with zero cross-flow force, indicating a large molecule size. Such components were more abundant in the non-purified samples and might include remnants of dsRNA substrate that were not completely processed ( Figure 4). In addition to the Postnova AF2000 instrument (Figure 4), the Wyatt Eclipse FFF NEON system coupled to the multi-detection platform was applied to purify siRNA molecules ( Figure 5). The MALS data suggested that the molecular weight of the purified siRNAs was 17.7 kDa ( Figure 5C), which was in good agreement with the theoretically calculated value of 17.3 kDa ( Table 1). The derived polydispersity index (M w /M n ) of 1.02 ± 2.6% confirmed the homogeneity of the sample.
To estimate the recovery of siRNAs with a 10 kDa RC membrane, we injected known amounts (7.5-15 µg) of siRNAs pre-purified with AEX to AF4 and collected siRNAcontaining fractions. After precipitation (see Section 4.1.4.), the mass recovery of the siRNA molecules was 72 ± 16%. A similar estimate for recovery, 70 ± 15%, was calculated by ASTRA 8.0 algorithm based on the collected signals from the concentration detectors (UV and RI). Thus, a 10 kDa RC membrane provided good recovery of the loaded siRNA together with the possibility to use higher flow rates without causing overpressure.
AF4 Fractionation Separates siRNAs from Contaminating Reaction Components
In addition to incompletely digested dsRNA molecules, which are separated from siRNAs with the developed AF4 method (Figures 4 and 5), reactions contain the Dicer enzyme and potential traces of bacteriophage polymerases and NTPs, although the last two are mostly removed from dsRNA sample during LiCl precipitation step [71]. To check that proteins and NTPs do not co-elute with the siRNA fractions during AF4 separation, we mixed pre-purified siRNA molecules with NTPs or a model protein, bovine serum albumin (BSA), and separated the mixtures using Eclipse NEON multidetector system (Wyatt) under the conditions described above (see Section 2.2.1, Figure 5). BSA with a molecular weight of 66.5 kDa was well separated from siRNA peak eluting substantially later than siRNAs ( Figure S6A). Bacteriophage polymerases and Giardia Dicer are all negatively charged at pH 8.0 and larger than BSA (phi6 polymerase-75 kDa, pI = 6.57; T7 polymerase-99 kDa, pI = 6.77; and Giardia Dicer-82 kDa, pI = 5.98), which should result in even longer retention in the channel and, hence, better separation from siRNAs compared to that observed with BSA. Most of the injected NTPs are removed with the cross-flow through a 10 kDa RC membrane. The fraction of NTPs (about 0.3%), which is retained in the channel, elute in the void volume, which is distinct from the siRNA peak, when 20 mM NaCl, 50 mM Tris-HCl (pH 8.0) is used as mobile phase ( Figure S6B).
To estimate the recovery of siRNAs with a 10 kDa RC membrane, we injected known amounts (7.5-15 μg) of siRNAs pre-purified with AEX to AF4 and collected siRNAcontaining fractions. After precipitation (see Section 4.1.4.), the mass recovery of the siRNA molecules was 72 ± 16%. A similar estimate for recovery, 70 ± 15%, was calculated by ASTRA 8.0 algorithm based on the collected signals from the concentration detectors (UV and RI). Thus, a 10 kDa RC membrane provided good recovery of the loaded siRNA together with the possibility to use higher flow rates without causing overpressure.
Mobile Phase Influences Elution and siRNA Purification by AF4
Preparation of siRNAs for therapeutic applications would benefit from the decreased number of processing steps and reduced hands-on time. Purification of siRNAs using water as a mobile phase during AF4 would eliminate the need for a desalting procedure and potentiate higher siRNA yield. We evaluated the separation of siRNAs in Tris buffer that did not contain NaCl as well as in ultrapure water ( Figure 6). A fractogram from AF4 experiments with the standard mobile phase of 20 mM NaCl, 50 mM Tris-HCl showed a single distinct siRNA peak that eluted after~6 min. When the analysis was performed in 10 mM Tris-HCl buffer lacking NaCl, the retention time slightly decreased and the peak was broader. When the fractionation was carried out in ultrapure water, the siRNAs eluted as high-intensity peaks at the beginning of elution in the void volume along with the non-retained small sample components. Furthermore, under these conditions, the separation between siRNAs and proteins used for their production might not be achieved, as the siRNA and BSA peaks are not fully resolved (compare Figures S6A and S7, Supplementary Materials).
10 mM Tris-HCl buffer lacking NaCl, the retention time slightly decreased and the peak was broader. When the fractionation was carried out in ultrapure water, the siRNAs eluted as high-intensity peaks at the beginning of elution in the void volume along with the non-retained small sample components. Furthermore, under these conditions, the separation between siRNAs and proteins used for their production might not be achieved, as the siRNA and BSA peaks are not fully resolved (compare Figures S6A and S7, Supplementary Materials). Mammalian cells are sensitive to the impurities that arise from the enzymatic production of siRNA preparation. For instance, dsRNA molecules that are longer than 30 base pairs (bp) activate the mammalian interferon system and lead to cell death via apoptosis or necroptosis [72,73]. Therefore, it is crucial that the purification procedure for siRNAs removes all undigested or partially digested dsRNA molecules. Other contaminants, such as ssRNAs, rNTPs, and enzymes used for siRNA production, should be also avoided. To assess the quality of siRNA purification with AF4, toxicity tests were performed in vitro using two cell lines, human nervous system-derived cells U-373 MG and human corneal epithelium (HCE) cells. The cell cultures were selected to represent tissues naturally susceptible to HSV-1 infection. Swarms of siRNAs purified with AEX ( Figure S8) were used as a reference since the AEX method for siRNA molecules is well established and extensively verified [31,49,62,63,74].
We evaluated the toxicity of siRNA batches purified with AF4 in either 20 mM NaCl, 50 mM Tris-HCl (pH 8.0), or in ultrapure water (see Section 2.2.3; Figure 5). To this end, the U-373 MG or HCE cells were transfected with the purified siRNAs on 96-well plates using 5, 10, or 20 pmols siRNA per well. Water was used for mock transfection. Transfection with 1 pmol of 88-nt dsRNA was included as a toxicity control since it leads to cell death [31]. As expected, the 88-nt dsRNAs caused about a 75% drop in the viability of U-373 MG and HCE cells, which verified the successful transfection procedure. In general, U-373 MG cells were more sensitive to transfection compared to HCE cells, and we observed about 75% and 90% cell viability in U-373 MG and HCE cells, respectively, after mock transfection (Figure 7). The viability of the cells did not differ significantly between the mock transfection control and transfection with AEX-purified siRNAs. SiRNAs separated with AF4 in a buffered mobile phase were also not toxic to the cells ( Figure 7A,B). Notably, the siRNAs are typically used in amounts below 10 pmols/well to induce efficient gene knockdown in mammalian cells, but even 2-fold higher concentrations of the siRNAs were well tolerated by the cells used in the experiments. However, siRNAs purified with AF4 using ultrapure water as a mobile phase caused a slight but statistically significant drop in cell viability for both U-373 MG and HCE cells ( Figure 7A,B), indicating compromised purity of these siRNA batches.
SiRNAs separated with AF4 in a buffered mobile phase were also not toxic to the cells ( Figure 7A,B). Notably, the siRNAs are typically used in amounts below 10 pmols/well to induce efficient gene knockdown in mammalian cells, but even 2-fold higher concentrations of the siRNAs were well tolerated by the cells used in the experiments. However, siRNAs purified with AF4 using ultrapure water as a mobile phase caused a slight but statistically significant drop in cell viability for both U-373 MG and HCE cells ( Figure 7A,B), indicating compromised purity of these siRNA batches. The cells were transfected with 5, 10 or 20 pmols of the purified eGFP or UL29 siRNAs per well, 1 pmol/well of 88-nt dsRNA or water. After 48 h post transfection, the cell viability was determined using CellTiter-Glo luminescent assay. The viability was calculated as a percentage from the viability of intact cells. Data are presented as the mean ± S.D. Asterisks represent a significant difference determined by Mann-Whitney U test either between a control (mock transfection) and a group of comparison or between two indicated groups; *-p < 0.05, and **-p < 0.01.
AF4 Purified siRNA Swarms Retain Their Antiviral Activity
We have previously demonstrated antiviral activity of AEX-purified UL29 siRNA swarms against HSV infection by measuring viral shedding with a plaque assay and UL29specific mRNA expression by quantitative reverse transcription PCR [31]. To confirm that AF4 purification of siRNA swarms does not affect their antiviral properties, we transfected U-373 MG and HCE cells with 5 pmol/well of siRNAs, and four hours later we infected the cells with 1000 plaque-forming units (PFU) of HSV-1 17+ strain. At the time point of 48 h post transfection (hpt), the supernatants were collected and titrated on Vero cells for a plaque assay. According to the assay, the HSV-1 had similar titers,~1.0 × 10 6 PFU/mL and 3.4 × 10 6 PFU/mL in U-373 MG and HCE cells, respectively. Mock transfection did not affect virus growth in cell cultures. The antiviral activities of the AF4-and AEX-purified UL29 siRNA swarms were similar ( Figure 8) and both reduced viral titers up to three orders of magnitude (p < 0.01) preventing the development of morphological changes induced by HSV-1 infection. The eGFP siRNA swarm also demonstrated some non-specific antiviral activity (Figure 8) related to the activation of innate immune responses [62,74], but this effect was not as prominent as that of HSV-specific siRNA swarm. transfection did not affect virus growth in cell cultures. The antiviral activities of the AF4and AEX-purified UL29 siRNA swarms were similar ( Figure 8) and both reduced viral titers up to three orders of magnitude (p < 0.01) preventing the development of morphological changes induced by HSV-1 infection. The eGFP siRNA swarm also demonstrated some non-specific antiviral activity (Figure 8) related to the activation of innate immune responses [62,74], but this effect was not as prominent as that of HSVspecific siRNA swarm.
Discussion
Small RNA molecules are becoming increasingly important therapeutic modalities, and new advanced methods for their purification, separation, and analysis are needed. Furthermore, newly discovered sRNA molecules need to be purified to facilitate their structural and functional studies. AF4 is a versatile separation technique, which allows gentle purification of biomolecules in their native state. We investigated the elution behavior of sRNA molecules, their potential purification and separation with the AF4 system. Furthermore, we applied AF4 coupled to a multidetector system (MALS, RI, UV, and fluorescence detector) to characterize molar mass, homogeneity, and recovery of Figure 8. Biological properties of AF4-purified siRNA swarms. To evaluate the antiviral activity of AEX-and AF4-purified siRNA swarms, the U-373 MG (A) and HCE (B) cells were transfected on 96-well plates with 5 pmols/well of either HSV-1-specific UL29 siRNAs or control non-specific eGFP siRNAs. After 4 h post transfection, the cells were infected with 1000 plaque forming units (PFU) of HSV-1 17+ strain. After 44 h post infection, the supernatant was collected for plaque assay. Data are presented the as mean ± S.D. Asterisks represent a significant difference either between a control (water transfection or water transfection + infection) and a group of comparison or a difference between two indicated groups as determined by Mann-Whitney U test; *-p < 0.05.
Discussion
Small RNA molecules are becoming increasingly important therapeutic modalities, and new advanced methods for their purification, separation, and analysis are needed. Furthermore, newly discovered sRNA molecules need to be purified to facilitate their structural and functional studies. AF4 is a versatile separation technique, which allows gentle purification of biomolecules in their native state. We investigated the elution behavior of sRNA molecules, their potential purification and separation with the AF4 system. Furthermore, we applied AF4 coupled to a multidetector system (MALS, RI, UV, and fluorescence detector) to characterize molar mass, homogeneity, and recovery of sRNA preparations. We used two types of ssRNA and dsRNA molecules, 88-and 108-nt-long, as well as 27-nt-long siRNAs ( Table 1).
The uniform hydrodynamic size of the short dsRNA molecules and their rod-shaped structure made them more feasible purification targets compared to single-stranded molecules, which are prone to adopt alternative conformations (Table 1, Figures S1 and S2) and form multimeric complexes ( Figure 2C,D, Figures S4 and S5). The pre-purified 88-and 108-nt dsRNAs eluted as single homogenous peaks ( Figure 3) and a sufficient separation was achieved between the monomeric ssRNA form and the dsRNA when the crude dsRNA synthesis reaction mixtures were analyzed (Figure 3).
The 108-nt ssRNA studied here appeared to have a tendency to form multimeric complexes, which resulted in its elution in two main peaks ( Figure 2C,D). Thus, our results suggest that AF4 could be used for the detection of alternative RNA forms and studies of RNA folding and multimerization. The formation of such multimers can be probably induced during the focusing step of the AF4 experiment, where the analyte is concentrated to a small volume (approximately 1-2 µL) potentiating intermolecular interactions [57]. Such observation suggests that AF4 could provide a means to promote the detection of alternative RNA forms and to study RNA folding and multimerization. However, if the formation of such structures compromises the purification process, the sample injection technique implemented in frit-inlet/dispersion channel, which omits the focusing step, could provide a solution [57]. The composition of a mobile phase affects the elution behavior of negatively charged RNA molecules [68]. Accordingly, the mobile phase with low ionic strength substantially decreased the intermolecular interactions observed between 108-nt ssRNAs and reduced the retention time of the analytes ( Figure S5).
SiRNA molecules were efficiently separated by AF4 from partially digested dsRNAs, NTPs, and enzymes used in their production (Figures 4 and 5). The method for purification of enzymatically created siRNA swarms was verified with two AF4 instruments currently available on the market; AF2000 MultiFlow FFF (Postnova Analytics) and Eclipse NEON (Wyatt). Both systems provided a similar level of siRNA purity. Despite our initial concern that 10 kDa RC membrane will not retain siRNA, the measured recovery using 10 kDa membrane was about 70% in the mobile phase containing 20 mM NaCl in 50 mM Tris-HCl (pH 8.0). This exceeds an average siRNA recovery obtained with AEX by 20% [31]. The AF4 operation using a membrane with 10 kDa MWCO instead of 1 kDa also allows higher flow rates reducing problems related to overpressure and resulting in a better separation and shorter analysis time.
We studied if water can be used as a mobile phase for siRNA purification. The idea of this experiment arose from the presumption that AF4 separations in water might potentiate the application of siRNA swarms directly to the cells as no buffer exchange or desalting steps are required. However, when water was used as a mobile phase, siRNAs eluted in a void volume, which complicated their separation from small sample components (Figures 6 and S7). This translated into a slightly but statistically significantly increased cellular toxicity of siRNAs swarms purified in plain water (Figure 7). Therefore, to accomplish efficient separation of siRNAs from impurities, a buffered solution containing 10-20 mM salt should be used. The siRNAs purified in 20 mM NaCl, 50 mM Tris-HCl (pH 8.0) were not toxic to the human cells in vitro ( Figure 7) and effectively inhibited HSV-1 infection (Figure 8).
Although AF4 was designed for the separation and characterization of large macromolecules and macromolecular complexes, we demonstrated that it is also applicable for the purification of short ssRNA and dsRNA molecules, including siRNAs. The amounts of RNA loaded into AF4 channels in this study were limited to tens of micrograms. However, large-scale purification would require scaling up of the injected RNA amounts to speed up the siRNA production process. The AF4 semi-preparative channels are currently available on market enabling upscaling from microgram to milligram amounts. Combination of the AF4 separation with MALS, RI, and UV detectors allows monitoring of the quantity of the product as well as a number of quality aspects, such as molar mass, polydispersity, radius of gyration, and length of RNA molecules.
Generation of sRNA Molecules
We generated five types of sRNAs for the study (Table 1). To produce 88-or 108-nt ss-and dsRNA, plasmid pLM659 containing a complementary DNA copy of the genomic S-segment of bacteriophage phi6 (GenBank accession number NC_003714) [75] was used as a template in PCR. The amplified DNA molecules comprise 80 nt or 100 nt from the very 3 end of the S-segment. The primers used for template amplification contained promoter regions for T7 and φ6 polymerases bringing the extra 8 nt in the resulting RNA sequence [59]. DNA templates for the production of siRNA swarms were PCR-amplified from plasmids pET32UL29 [31] or pCR3.1-eGFP [29]. Plasmid pET32UL29 contains 653 bp sequence derived from the UL29 gene of HSV-1 prototype strain 17+ (GenBank JN555585.1, nucleotides 59,302 to 59,954) and plasmid pCR3.1-eGFP harbors complete 717 bp eGFP gene subcloned from pEGFP-C1 (GeneBank U55763). In the manuscript text, these dsRNAs are referred to as 650 bp and 720 bp, respectively. Gel-purified PCR products were used as templates to generate ssRNA by T7 DNA-dependent RNA polymerase. The dsRNA molecules were produced in coupled reactions using T7 and phi6 polymerase [29].
After DNase treatment (RQ1 RNase-Free DNAse, Promega, Madison, WI, USA), the RNAs were directly used for AF4 injections or further purified prior to the AF4 analysis. The pre-purification of ssRNAs included TRIsure-chloroform extraction (TRIsure, Bioline GmbH, Luckenwald, Germany) and precipitation with 4 M lithium chloride (LiCl, Merck, Darmstadt, Germany), while pre-purification of dsRNA molecules included consecutive precipitations with 2 M and 4 M LiCl (Figure 1). The RNAs were stored at −80 • C. A fraction of prepared 88-nt dsRNA molecules was purified with AEX (Section 4.1.3) to serve as a reference for the cell viability analyses.
To generate siRNA swarms, pre-purified UL29 and eGFP dsRNAs were digested into 27-nt siRNAs using the Dicer enzyme from Giardia intestinalis [31]. The enzyme was expressed in Bac-to-Bac baculovirus expression system and purified using HisPur Ni-NTA resin (Thermo Fisher Scientific; Rockford, IL, USA) by the facility of protein services of the University of Tampere. Half of the prepared siRNA swarms were subjected to AF4 purification, and the rest was purified using AEX as previously described [49], with modifications highlighted in 4.1.3. For preliminary AF4 experiments, we used siRNA swarm obtained after digestion of bacteriophage phi6 dsRNA genome (GenBank accession numbers NC_003714, NC 003715, NC 003716) with Dicer. To this end, bacteriophage phi6 [76] was propagated and purified as described previously [77]. Total RNA was extracted from viral particles using TRIzure reagent (Bioline GmbH, Luckenwald, Germany) and chloroform according to the manufacturer's instructions. The dsRNA was fractioned using LiCl precipitation and cleaved with Giardia Dicer into siRNA swarms.
AF4 Setup and Operation
We performed the initial AF4 experiments and data collection with the set-up previously described [78] using the AF2000 MT instrument (Figures 2-4 were collected during the elution step. Retention times (t r ) with focusing and transition time deducted were obtained from the peak maxima. Alternatively, the Eclipse NEON FFF system (Wyatt Technology, Santa Barbara, CA, USA) comprising an Agilent 1260 Infinity II pump and autosampler, analytical long channel, the Eclipse FFF flow instrument, dilution control module (DCM), Agilent 1260 fraction collector, was used for AF4 purification of siRNA molecules ( Figure 5, Figures S3, S4, S6, and S7). The Eclipse NEON FFF system was connected to DAWN, a Multi-Angle Light Scattering detector with 18 angles (MALS, Wyatt Technology, Santa Barbara, CA, USA), one of which is occupied by WyattQELS online dynamic light scattering (DLS) detector, Agilent 1260 Infinity II UV and fluorescence detectors as well as Optilab differential refractometer (dRI) detector (Wyatt Technology, Santa Barbara, CA, USA). A UV detector monitored the nucleic acid signal at 260 nm, whereas a fluorescence detector was utilized to monitor proteins (excitation at 280 nm, emission at 340 nm) The FFF-MALS was controlled via VISION 3.0.0.20 software (Wyatt Technology, Santa Barbara, CA, USA). Data were analyzed with ASTRA 8.0.0 software (Wyatt Technology, Santa Barbara, CA, USA). The optimized flow program for the ECLIPSE NEON is shown in Table S1. The channel flow was 1 mL/min, inject flow 0.2 mL/min, and detector flow 0.5 mL/min.
A Teflon spacer adjusted the FFF separation channel to a nominal width of 350 µm (for Postnova AF2000MT instrument) or 400 µm (for the Eclipse NEON system). The MWCO of the RC membrane was 1 kDa (Postnova) or 10 kDa (Wyatt Technology).
Unless otherwise stated, we performed AF4 experiments at 22 • C using 20 mM NaCl, 50 mM Tris-HCl (pH 8.0) buffer. Alternatively, we used 10 mM Tris-HCl (pH 8.0), or nuclease-free ultrapure water purified with the Milli-Q ® Direct 8 system (Merck, Darmstadt, Germany) as a mobile phase. Each fractionation was repeated at least two times. The instrument was rinsed with the used mobile phase intensively between different RNA samples.
RNA Purification Using AEX
Chromatographic experiments were performed at room temperature using the ÄK-TAPurifier 10UPC system (GE Healthcare, Uppsala, Sweden) controlled by Unicorn 5.2 software (GE Healthcare, Uppsala, Sweden). An anion exchange CIM QA-1 monolithic column (BIA Separations, Ajdovscina, Slovenia) was used for the reference RNA purification. The column has an average pore size of 1.3 µm, outer diameter of 18.6 mm, inner diameter of 6.7 mm, length of 4.2 mm, and a bed volume of 1.0 mL. Its stationary phase is based on poly(glycidylmethacrylate-co-ethylenedimethacrylate) functionalized with quaternary amines. Buffer A containing 20 mM NaCl, 50 mM Tris-HCl (pH 8.0) and buffer B containing 1 M NaCl, 50 mM Tris-HCl (pH 8.0) were used in AEX. Moreover, 88-nt dsRNA was purified as described previously [49]. The generated siRNAs were diluted 3.5-times with nuclease-free ultrapure water and centrifuged at 11,000× g for 10 min before the injection. The injection volume was 5 mL. A linear gradient of 60 column volumes from 20% to 80% buffer B at 1 mL/min was applied ( Figure S8). The column was cleaned with 1 M NaOH solution between samples. Absorbance at 260 nm was monitored and 0.5 mL fractions from the peaks were collected.
Analysis of Recovery, Quality and Quantity
Nucleic acids were recovered from the collected fractions by overnight precipitation at −20 • C in 0.3 M sodium acetate (pH 6.5) and 67% ethanol, followed by centrifugation at 10,000× g for 30 min. The RNA pellets were washed with 70% ethanol, air-dried, and dissolved in 10 µL of sterile nuclease-free water. For functional experiments, the fractions containing maximum amounts of siRNAs were combined and, unless purified in ultrapure water, they were desalted using illustra NAP-25 columns (GE Healthcare, Buckinghamshire, UK) followed by concentrating with the SpeedVac vacuum concentrator (Savant Instruments Inc, Famingdale, NY, USA). Nucleic acid concentrations were measured using NanoDrop2000c Spectrophotometer (Thermo Scientific). Recovery of RNA molecules was calculated with the following formula ([eluted RNA, µg]/[input RNA, µg]) × 100%. Alternatively, an algorithm from Wyatt Technology implemented in ASTRA 8 software enabled the mass recovery calculation of RNA molecules based on the eluted RNA concentration determined by dRI detector, assuming that the refractive index increment (dn/dc) for RNA is 0.17 mL/g.
RNase A Treatment
The fractions collected after AF4 separation of 108-nt ssRNAs were precipitated overnight at −20 • C, as described in Section 4.1.4. Then, 1 µg of RNA from peak 1 or peak 2 was incubated with 0.1 µg of RNase A (Fermentas, Vilnius, Lithuania) either in 1 × SSC buffer (3 M sodium chloride, 0.3 M sodium citrate [pH 7.0]) or in 0.1 × SSC buffer for 15 min at room temperature. The former treatment in a high salt buffer promotes only ssRNA digestions, and the latter one, occurring in low salt buffer, results in the digestion of both ss-and dsRNA species. Enzymatically produced 500-nt dsRNA [59] was used as a control.
Cells and Virus
A human glioma cell line U-373 MG (currently re-classified as U-251) was originally obtained from ATCC (Manassas, VA, USA). Immortalized HCE were from Prof. Arto Urtti (University of Helsinki/University of Eastern Finland, Finland). Both cell lines are susceptible and permissive to HSV-1 and suitable for antiviral siRNA treatment [31,79]. The cells were maintained in high-glucose DMEM (Gibco) with 10% (v/v) fetal bovine serum (FBS, Serana, Silicone Valley, CA, USA) and 2 mM L-glutamine (Sigma, Saint Louis, MO, USA) in 37 • C, 5% CO 2 . African green monkey kidney cells Vero (ATCC), used for plaque assays, were grown in DMEM with 5% FBS (v/v).
The prototype HSV-1 17+ laboratory strain was used for antiviral studies and propagated as previously described [31].
Transfection
The cells were propagated on 96-well plates until their confluency reached about 50% for HCE and 70% for U-373 MG cells. The cells were transfected with 5, 10, or 20 pmol of siRNA preparation per well, 1 pmol per well of 88-nt dsRNA [49], or water (mock transfection) using Lipofectamine RNAiMAX (Invitrogen, Carlsbad, CA, USA) according to the manufacturer's forward transfection protocol. Altogether, 15 independently produced and purified batches of siRNA swarms were used for transfection experiments; the seven batches were purified with AEX to serve as a reference, three batches were purified with AF4 using the standard mobile phase [20 mM NaCl, Tris-HCl (pH 8.0)], and five batches were purified using ultrapure water as a mobile phase. All experiments were repeated two or three times with three technical replicates each.
Cell Viability Assay
Possible cytotoxic effects of siRNA swarms purified with AF4 were assessed 48 hpt with the CellTiter-Glo Luminescent Assay (Promega, Madison, WI, USA) according to the manufacturer's instructions, and luminescence was quantified with VICTOR Nivo Multimode Plate Reader (Perkin Elmer, Waltham, MA, USA). The luminescent signal from untreated cells was set as 100%, and all the other values were normalized to it. Viability similar to mock-transfected cells was considered acceptable. We measured a signal from three technical replicates for each treatment, and the average of these was used in further statistical analyses (Section 4.2.5).
HSV-1 Infection and Plaque Assay
After 4 hpt with 5 pmol/well of either UL29 or eGFP siRNA swarms, the U-373MG or HCE cells were washed with DMEM containing 2% FBS, infected with 1000 PFU/well of HSV-1 17+ in 100 µL DMEM with 2% FBS, and incubated 1.5 h at 35 • C on a shaker. Then, the cells were washed and supplemented with DMEM containing 7% FBS. Live-cell imaging to confirm cytopathic effects caused by HSV-1 was performed 44 h post infection using EVOS Auto FL (Thermo Fisher Scientific). Then, the culture supernatant samples were collected, and their 10-fold dilutions were used to infect monolayers of Vero cells for 1.5 h at 35 • C in 100 µL of DMEM containing 5% FBS. Then, an equal amount of DMEM with 5% FBS and 80 µg/mL of human IgG (Kiovig, Baxalta US Inc., MA, USA) was added. After 3 or 4 days post infection, the cells were fixed with methanol followed by staining with crystal violet and quantification of plaques.
Statistical Analysis
Statistical analysis was conducted with SPSS Statistics version 25 (IBM, Armonk, NY, USA). Statistical significances were calculated with Mann-Whitney's non-parametric U test comparing two individual groups.
Conclusions
We introduce a novel AF4-based method for the purification and analysis of sRNA molecules under native conditions. We found that AF4 enables rapid purification of enzymatically produced antiviral siRNAs from partially digested long dsRNA. In addition, AF4 separates monomeric single-stranded and double-stranded RNA molecules of the same length and promotes the identification of different multimeric forms of ssRNA. Recovery of AF4-purified sRNA was high, about 70% and above. AF4-purified siRNAs were not toxic for mammalian cells and AF4-purified siRNA swarm derived from the UL29 gene sequence of herpes simplex virus 1 was biologically active as it efficiently inhibited herpes virus replication in cell cultures.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/ph15020261/s1, Figure S1: Predicted secondary structures for the 88-nt ssRNA at 22 • C, Figure S2: Predicted secondary structures for the 108-nt ssRNA at 22 • C, Figure S3: Molar mass distribution of the 88-nt and 108-nt RNA sample components; Figure S4: Nature of 108-nt ssRNA species fractionated with AF4; Figure S5: Impact of mobile phase composition on 108-nt ssRNA elution; Figure S6: AF4 separation of siRNA molecules from BSA and NTPs; Figure S7: AF4 separation of siRNA molecules from BSA in water; Figure S8: Anion exchange chromatography of siRNA swarms; Table S1: Method for AF4 separation used with the Postnova AF4 system; Table S2: Method for SF4 separation used with the Eclipse NEON AF4 system. | 10,742.2 | 2022-02-01T00:00:00.000 | [
"Biology",
"Chemistry"
] |
The logos Christology in the fourth gospel (Jn 1:1–5, 14): A soteriological response to an Ewe cosmic prayer
for deliverance from the state of ignorance is the logos incarnate in John 1:14. Contribution: The article contributes to the academic knowledge on the logos Christology in the prologue of the fourth gospel (John 1:1–5, 14), specifically in the use of mother tongue in biblical hermeneutics in Africa.
Introduction
The gospel of John presents us with the interpretation of the Christian message within the context of what appears to be a sophisticated and philosophical recipient community.The evangelist's innovative logos Christology sharply contrasts with the incarnation theologies in Matthew and Luke (Mt 1:18-25, Lk 1:26-38, 2:1-7), and reiterates the case for contextualisation of the gospel message.The encounter between Western Christianity and Ewe indigenous religion in the mid-19th century places the Ewe-Ghanaian Christian in the midst of two religious worldviews -Judeo-Christian and Ewe indigenous religious worldviews, and its attendant cultic practices.An investigation into the Yewe cult, a sect in Ewe religion, reveals a cosmic concept similar to the logos concept in the prologue to the gospel of John.This article seeks to interpret John's logos Christology (Jn 1:1-5, 14) which he developed from Greek religion and philosophy, and concepts prevalent in Palestinian Judaism, within the context of Ewe-Ghanaian cosmology, and assess the implication it brings to bear on Ewe-Ghanaian primal religion (Harris 2002:208-220).The article employs the exegetical and mother tongue hermeneutical approaches as its methodologies (Ekem 2007:77;Fee & Strauss 2003:23-31;Porter & Clarke 2007:3-18;Kuwornu-Adjaottor 2012:11-15).The exegetical and hermeneutical methods were applied in the article as follows: This article interprets the logos Christology in the fourth gospel within Ewe-Ghanaian cosmic setting.The article employs a combination of the exegetical and mother tongue biblical hermeneutics as its methodologies.The article compares the concept of the logos in John 1:1-5, 14 with a similar concept in Ewe cosmology with the aim of finding their points of convergence and divergence.The article also identifies linguistic and theological gaps in the Ewe rendition of John 1:1-5, 14 and proposes a new translation that addresses it.A translation of the logos in Ewe cosmic perspective renders the term nyagbe, the intelligible word which dwells with the female pair of Sogbe-Lisa and manifests itself from the absolute state of existence [logo] to the embodied state [logosu].When nyagbe journeyed from the eternal state of existence to the embodied state of existence, it manifests itself at five different levels of existence until it reaches the final state where humans exist.Although believed to be the epitome of all creation, human beings are at the same time described as being at the shore of ignorance and therefore must continually climb the ladder of knowledge in order to free themselves from it.It is for this reason that humans constantly petition the cosmic forces for their salvation.The article therefore, identifies a soteriological gap created in the Ewe soteriology and concludes that the only soteriological response to the Ewe cosmic cry for deliverance from the state of ignorance is the logos incarnate in John 1:14.
The logos Christology in the fourth gospel (Jn 1:1-5, 14): A soteriological response to an Ewe cosmic prayer
Read online: 1.A lexical inquiry into John 1:1-5, 14, focusing on the preexistence of the logos, its cosmic authority, and incarnation.2. A comparative analysis between John 1:1-5, 14 and its parallel in Ewe cosmic setting with the aim of finding points of convergence and divergence.3.An assessment of the soteriological implication of the logos incarnate for the Ewe-Ghanaian primal religion.
The concept of the logos in the world of the fourth gospel
The logos Christology is viewed hypothetically as a hymn ascribed to Jesus Christ in the fourth gospel (Boring 2012:678;Brown 1997:337;Card 2014:30;Conzelmann 1968:335;Harris 2002:219;Heil 2015:5-15;Kieffer 2001:188;Spivey, Smith & Black 2007:155). 1John employs the concept to interpret the incarnation to what appears to be a highly sophisticated Johannine community said to be influenced by a uniquely high Christology, including belief in Jesus' pre-human existence as cosmic wisdom (Harris 2002(Harris :208, 209, 2012)).The term in the Greek world is rendered, 'word, speech, utterance, revelation', in the sense of things that are 'displayed, clarified, recognized, and understood as against words that are proclaimed by one person and heard by another' (Card 2014:18;ed. Kittel 1973: Lightfoot 2015:81).'It is the content of a thing in terms of its law, meaning, basis, and structure' (Bromiley 1985:506).The concept in Stoicism 'denote the principle of cosmic reason, the intelligent force that orders and sustains the universe, making it accessible and intelligent to the human mind' (Card 2014:31;Cullmann 1959:251;Dodd 1985:263;Harris 2003:465;ed. Kittel 1973:84;Lincoln 2005:95;Morris 1986:225).The logos in Greek philosophy is viewed as the creative power that brought the cosmos out of chaos.Thus, the wisdom behind the creation of the world and everything in it is what is known in Greek philosophy as logos (ed.Kittel 1973:85).Thus, logos is 'common for reason as well as speech'; it either denotes 'the principle which controls the universe, the soul of the world', or 'the generative principle in nature' (Robertson 1960:3; see also Wiles 1966:25).
In Greek mystery religions, for instance, the deities were referred to as 'personifications of the logos or the sons of God' who served as mediators or heralds of 'the divine will and great force of conception…a creative potency, the guide and agent of knowledge', and were described in speculative philosophy as 'the son of God' (demiurge) (Bromiley 1985:507).In the area of prayer, the logos was believed to be the only means by which one could commune with the divine.Its parallel in Jewish Wisdom literature is the personification of Wisdom as a young lady who described herself as the first born of God's creation, the craftsman of creation and a witness of the creation of heaven and earth (Harris 2003:465;Lincoln 2005:96;Robertson 1960:3;Pr 8:30-31; Wisdom of Solomon 6:12-9:18; Ecclus 24).The term, when viewed from the perspective of Old Testament revelation, is the word God used to create the cosmos (Bromiley 1985:507-509).
As Lightfoot (2015:82) summarises it, 'it is a Hebrew expression comprising of any revelation of God's will and a man's heart or understanding whether by direct speech or otherwise'.In his attempt to harmonise Judaism and Greek philosophy, Philo of Alexandria describes the logos as: [A] mediating figure which comes from God, forms a link between the transcendent God and the world, and represents humanity as a high priest and advocate before God…the sum and locus of God's creative power, and as such it orders and governs the visible world.(Bromiley 1985:507) One may conclude by stating that the logos is a complex concept that has evolved from its ordinary meaning in both Jewish and Greek context, to a more technical use in magical practice for incantation, through to becoming a 'metaphysical reality', and a 'cosmological hypostasis' (Bromiley 1985:506).
Christology of the logos in the prologue of the fourth gospel
The logos, according to John, does not only exists with God, but God himself.John theologised that God created the cosmos through the logos and is the source of life and light which dispels darkness (Espin 2007:786;Harris 2003:461;Ngewa 2006Ngewa :1252Ngewa -1253)).He then concludes by interpreting the logos in the light of the Christ event, that is, the logos became incarnated in the person of Jesus and made his dwelling among humans.Although John does not explain how the incarnation of the logos did occur, one can infer that it was through the virgin birth as recorded in Matthean and Lukan's accounts of the incarnation.His omission of the virgin birth in the incarnation process, as Harris posits, may be because of its relatively lower Christology (Harris 2003:458).John's use of en archei (in the beginning) to introduce his logos Christology draws one's attention to the creation narrative in Genesis 1:1 (Card 2014:31-32;Lincoln 2005:18).Thus, the creative principle that brought the cosmos into being existed eternally, were equal, intimate, and in 'perfect fellowship with God' (pros ton theon), a point he re-emphasised in John 1:2.His use of kai theos en ho logos (and the word was God) instead of kai ho theos ēn ho logos (and God was the word), it is argued, is an indication that the logos was God and not vice versa (Robertson 1960:3-5; see also Harris 2015:19-20;Lincoln 2005:97).Harris (2015) paraphrased it as follows: [A]t the very beginning of creation and time, the Word as the perfect expression of God the Father had already always existed, and this Word was in active communion with God, and this Word inherently shared the same nature as God.(p.20) John then continues with the cosmic role the logos played in creation, panta di 'autou egeneto (all things were made through him) kai chōris autou egeneto oide en ho gegonen (and without him was not anything made that was made), 'the intermediate agent in the work of creation', a point which he again re-emphasised in John 1:3 (Harris 2015:22-23;Lincoln 2005:98-99;Robertson 1960:5-6).In other words, the power to create life and sustain it is contained in the logos.In the words of John, zoe (life) is the creative force that emanates from the logos which lighten (phōs) the path of humans (en autō zoe ēn kai he zoe ēn to phōs tōn anthrōpōn) (in him was life and the life was the light of men).His use of zoe to render life makes it unlimited, that is, it includes all life except biographical life; it is '…the very principle or essence of life' which has the power to dispel darkness (Harris 2015
The idea of the logos in Ewe cosmology
John's logos Christology John 1:1-5, 14 depicts how the logos journeyed from its absolute or eternal state of existence to the embodied or incarnated state.Cosmic terminologies prevalent in Ewe cosmology reveal a unique concept that parallels that of the logos.Sakitey & Van Eck (2022:4) have argued that, the SupremeBeing, in Ewe cosmology, exists in three persons, namely, the dualistic pair of Sogbe-Lisa, the male and female pair surrounded by Chi, the awakener and consciousness behind Sogbe-Lisa, constituting the Supreme trinity (Sakitey & Van Eck 2022:4).They also argued that the whole universe in both physical and metaphysical forms is intrinsically linked to the Supreme Being (Sakitey & Van Eck 2022:4) and destroy life 5 respectively (Awoonor 2006:380-381;Sakitey & Van Eck 2022:4).In the second level which is the fifth in the cosmic order, Sogbe-Lisa is known as Xebieso, the god of light and sound (thunder), and is believed to control about 256 deities known in Ewe as Trͻwo who are the embodiment of cosmic law and order.The deities in the third level within logosu and sixth in the cosmic hierarchy are Voduwo, while in the fourth and final level of existence, Sogbe-Lisa manifests the self in visible form -sun, moon and galaxy of stars, plants, animals, the earth, rocks, mountains, rivers and the ocean, with humans as the epitome of all creation.
Apparently, there exists a cosmic hierarchy of a sort in Ewe cosmology with the Supreme Being occupying the absolute position in the hierarchy in a trinitarian relationship and delegating his cosmic powers to create, preserve, and destroy life to the three cosmic creatures of Sovi-Da, Sovi-Agbade, and Sovi-Agbeku in their respective order.Thus, the Supreme trinity and cosmic trinity are the two trinitarian concepts found in Ewe cosmology.The cosmic trinity are creatures of the Supreme trinity who perform cosmic functions on their behalf.The material world is the lowest of the cosmic hierarchy and final state of the manifestation of Sogbe Lisa.Human being (Amegbetͻ in Ewe) is the epitome of creation because it is believed that they are the embodiment of Sogbe-Lisa, the embodiment of all the laws of creation, the embodiment of all mysteries, and the only creature who is endowed with the power of speech (Sakitey & Van Eck 2022:4).
Ewe translation and theologies of the divine logos
Translators who worked on all four existing Ewe translations of the Bible (Agbenya La 2006; Biblia 1931, 2010; Nubabla Yeye La 1990) opted for nya to render logos in John 1:1-5, 14.The choice of nya, according to one of the consultants who worked on the Ewe Bible translation project, is its consistency with John's theology of the logos as God's word used to create the universe (G.Ansre [University of Ghana] interview, pers.comm., 13 February 2023).The term is rendered, word, or know, depending on the context in which it is used.In Ewe cultic language, logos is rendered nyagbe.The elision of gbe from nyagbe in translating logos or word in the Ewe Bible and spoken Ewe may be because of the convenient use of the language.The same reason may be assigned to words such as kadigbe (light) which is also elided to kadi.But amagbe, which is the Ewe name for colour is without elision.Nyagbe is therefore conceptualised in Ewe as the intelligible word which has the power to create, preserve, and 'destroy life'.This concept resonates with both the logos concept in Stoicism and Greek philosophy than it does with Old Testament revelation.
The pre-existence of the logos (Jn 1:1-2) The prologue of John's gospel opens with en archē ēn ho logos, kai ho logos ēn pros ton theon, kai theos ēn ho logos (in the beginning was the word, and the word was with God, and 5.The destruction of life must be understood as taking life back to its cosmic origin. the word was God).Ignatius strongly emphasised the divinity of the logos and clarifies that although the logos became flesh through Jesus Christ, it did not lose either the divine nature nor its essence.
Although Origen asserts that 'the logos was generated from the Father and thus begotten by God', he was certain, unlike Justin Martyr, about its subordination to the Father (McGrath 2007:283;Ryan 2007:785;Wiles 1966:28, 36).Since it has been well established that the intelligible word (nyagbe) dwells with the female pair of Sogbe-Lisa, its eternal existence and hypostatic relationship with the Supreme trinity is therefore unambiguous.
Cosmic authority of the logos (Jn 1:3-5)
John continues by establishing the cosmic authority of the logos in verse 3 of his prologue with the phrase, panta di 'autou egeneto, kai chōris autou egeneto oude en.ho gegonen (all things were made through him, and without him was not anything made that was made) rendered in Ewe as, wowͻ nuwo katã to eyama dzi, eye eya manͻme la womewoa naneke, si wowͻna o.His description of the logos as agent of God's creation, brings us back to the fourth state of existence in the cosmic hierarchy where the three cosmic beings -Sovi-Da, Sovi-Agbade, and Sovi-Agbeku, are believed to exist.As already indicated, these cosmic trio constitute the cosmic trinity and are responsible for creation, preservation, and destruction of life respectively (see also Awoonor 2006:380-381;Sakitey & Van Eck 2022:4).The cosmic trinity are the controllers of life, creating, preserving, and taking life back in due time.The life created by John's logos is expressed in the phrase, en autō zōe ēn, kai hē zōe ēn to phōs tōn anthrōpōn (in him was life, and the life was the light of men).The phrase reads, eya me agbe le, eye agbe la nye amewo fe kekeli, in Ewe.Situating this life in the context of Genesis 1, one may argue that the light in question is none other than the one that came to drive out darkness from the face of the earth; the light without which life, although created, cannot be lived well.Although it is not clear in John's prologue, a spiritual interpretation of the life inherent in the logos is implied in John's gospel and Johannine corpus as a whole (Jn 8:12, 14:6; 1 Jn 1:1-2).John Calvin observed: [T]herefore, understand life in this text very simply not as that life which is in all creatures, whether rational or irrational, but understand that life which believers are restored to their former life, such that freed from death and hell due to their sins, they have eternal life.(eds.Craig & Farmer 2014:19) This life, when viewed from Ewe cosmic perspective, however, is the one that is manifested in the sixth and final level of existence with human life as its epitome (Sakitey & Van Eck 2022:4).Whereas this life precedes light in John's logos Christology, the opposite is the case in Ewe cosmic order.Everything that would become life under the sun exists in the region of light under the control of Xebieso, the god of light and sound.Mawugã, the Supreme Being, just like Elohim the Creator of the universe in Genesis 1, manifested the self in the form of light before life appeared.A change in this cosmic arrangement may result in cosmic chaos.
Incarnation of the logos (Jn 1:14)
The incarnation of the logos concludes the prologue in John's logos Christology, kai ho logos sarx egeneto kai eskēnōsen en hēmin … (and the word became flesh and dwelt among us…) The first half of the verse 14 is rendered eye nya la trͻ zu ŋutilã, eye wonͻ anyi le mía dome… (Biblia 1931), nya la va zu amegbetͻ henͻ mía dome… (Biblia 2010; Nubabla Yeye La 1990), nya la trͻ zu ame va nͻ mía dome le anyigba sia dzi… (Agbenya La 2006).All four translations touched on the fact of the logos assuming flesh, nya la…zu ŋutilã/amegbetͻ/ame.The word egeneto, the aorist mid.indicative of 3rd person singular of ginomai, which translates zu in Ewe, conveys the idea of uniting oneself with something; dissolving something into another, as in salt dissolving in water to become salt solution.The idea, in relation to the logos, suggests that the logos assumes human nature but without necessarily losing its divine nature (eds.Craig & Farmer 2014:28;ed. Pelikan 1957:102, 110).All the arguments to establish the fact of two natures in a single person are difficult to comprehend.Unfortunately, John himself does not give his readers any clue to that effect but concludes that 'we have seen his glory, the glory of the one and only, who came from the Father, full of grace and truth' (Jn 1:14).The question of how the logos became flesh has also remained a mystery.The Old Testament equivalent of the logos is the word that God used to create the cosmos and nyagbe in Ewe cosmology.The place of the logos, in Ewe cosmic hierarchy, is the 'bosom' of the female pair of the Supreme trinity, that is, nyala li kple gbe (the logos dwells with gbe of Sogbe-Lisa).Thus, the incarnation, from a broader Ewe cosmic perspective, can be viewed in terms of everything that is existing in logo, the metaphysical state of existence manifesting itself in logosu, the physical state.
The logos Christology as soteriological response to Ewe cosmic prayer
In Ewe cosmic prayers, there is recognition that the human being is at the shore of the sea of ignorance and thus must continually climb the ladder of knowledge to free the self from the state of ignorance which creates fear, confusion, superstition, disease, sickness, squalor, culminating in death and destruction (Van Eck & Sakitey 2019a:179, 2019b:4).The prayer that is constantly offered to the cosmic forces to be freed from these predicaments is, 'ahe ne to dzi, evͻ ne to dzi, edͻ ne ƒo mia nu, eku ne ƒo mia ta', that is: [M]ay we not be caught in the state of darkness and ignorance; may we not be caught in the state of fear, confusion and superstition; may we not be caught in sickness, disease and squalor; may death and destruction depart from us.(Van Eck & Sakitey 2019a:179) The prayer highlights the four cardinal predicaments of life, namely, ignorance, fear, disease, and death, being the consequence of human's interaction with the pleasures of this life and subsequent loss of godly consciousness.Thus, the fall of man, from Ewe cosmic perspective, is a from consciousness of God.This fall 'creates a new order in man that is always in conflict with [his] godly nature' and creating a 'God-devil/good-evil' dualism in him (Van Eck & Sakitey 2019a:179-180).It is, therefore, human's sole responsibility to free themselves from their predicament in order to restore their godly consciousness.The encounter between Christianity and Ewe indigenous religion in the mid-19th century (1847), coupled with the translation of the Judeo-Christian Scriptures into Ewe, exposed the Ewe-Ghanaian to a soteriology that is higher than what is prevalent in their indigenous religious worldview.Whereas their salvation depends on their effort, the logos Christology portrays Jesus as the agent of that liberation.In other words, Christ is the appropriate response to the Ewe-Ghanaian cosmic cry for deliverance from the power of darkness, ignorance, fear of the unknown, sickness, death and destruction.Among all his faces that have been portrayed in African Christologies, Jesus' life-giving face (Jn 1:3-5; Col 1:15-20) perfectly fits into his soteriological role and resonates with Ewe quest and aspiration for fullness of life (Adams 2010:190-193;Stinton 2004:56-58; see also ed.Bediako 2000;Schreiter 1991).In the Ewe cosmic hierarchy, Jesus combines the cosmic trio that is clothed with the power to create, preserve and destroys life, and that of Xebieso, the controller of life and all forces within the cosmic hierarchy.Above all, he is the intelligible word (nyagbe) that dwells with Sogbe-Lisa, and shares both eternal and hypostatic relationship with the Supreme trinity.The cosmic supremacy he commands over Ewe indigenous religion, therefore, raises the level of godly consciousness of the Ewe-Ghanaian and settles the cosmic war between the biblical worldview and Ewe metaphysical reality.
Conclusion
Arguably, the logos concept is the most complex and controversial Christological concept in Christian theology.
Although the church has put its stamp of authority on the controversies surrounding the divine-human dualism of the incarnate logos, speculations around it have since continued.This may be because of lack of clarity in John's Christological interpretation.However, the theological significance that John brings to bear on his interpretation -the salvation of humanity -should bring finality to any form of speculation because of its supremacy over them.The translation of the logos in Ewe cosmic perspective which renders it nyagbe, the intelligible word, broadens its meaning to the manifestation of everything that exists in logo, the metaphysical state of existence to logosu, the physical state of existence.The place of nyagbe in the Ewe cosmic hierarchy, is gbe, the female pair of Sogbe-Lisa, which is contrary to the translation in the existing Ewe Bibles which suggests that the logos dwells with Mawu, the generic name for the Supreme Being.Therefore, in translating logos in Ewe, annotation in the revised versions of John 1:1-5, 14 is noteworthy.Since incarnation from Ewe cosmic perspective is understood to mean the manifestation of the Supreme Being from logo, the absolute and eternal state of existence to logosu, the embodied state of existence, the human being (amegbetͻ), though not perfect because of his fall from his godly consciousness, is perceived as the epitome of creation and embodiment of the Supreme Being.And because of his inability to free himself from the predicaments of life, he must continually offer supplications to the cosmic forces for his salvation.The incarnation of the divine logos, coupled with his ability to save humanity from sin and death, positions him as the soteriological response to the Ewe cosmic cry for deliverance from ignorance, fear, disease and death, and the restoration of humanity's godly consciousness.
Robertson 1960:12-13).It appears the logos concept is a very familiar concept in the Johannine community, hence its use in theologising about the Christ event.The introduction of Christ as the incarnation of the logos makes the difference between John's theology and what pertained in the worldview of the Greeks.It has brought finality to all the speculations surrounding the logos in Greek philosophy and gives clarity to its theological significance -the salvation of the entire human race.
. Thus, the Supreme trinity in Ewe cosmic sense is made up of father, mother, and spirit, dwelling in the absolute or eternal state of existence referred to in Ewe cosmology as logo, 3 from lo (mystery) and go (container or source).Thus, logo, which coincidentally sounds like the Greek word logos, is the source of the mystery of all things; the unknowable, unnameable, undifferentiated, unfathomable state of existence (D.Kumordzi, interview, pers.comm., 08 May 2019).The state in which the Sogbe-Lisa, the Supreme Being, manifests himself is known in Ewe cosmology as logosu 4 -the world in which all things can be known, named, differentiated, and fathomed.It is the state where the pantheon of deities called Trͻwo, Voduwo and every living creature including human beings exist.It is believed that Sogbe-Lisa manifests himself in five different states within the logosu.In the first which is also the fourth in the cosmic hierarchy, Sogbe-Lisa manifests himself in the form of three cosmic creatures, namely, Sovi-Da, Sovi-Agbade, and Sovi-Agbeku, who are clothed with the powers to create, preserve, 2.Harris (2015:23), however, argues that the life is physical, spiritual, and eternal.3.Ewe people name their children Logo and Logosu.4.The mustard seed in the gospels is rendered logosu in the Ewe Bible (Mark 4:30-32;Matthew 13:31-32, 17:20; Luke 13:18-19, 17:5-6).
This is translated, Le gͻmedzedzea me la, Nya la li, eye Nya la li kple Mawu, eye Nya lae nye Mawu (Agbenya La 2006; Biblia 1931), and Do ŋgͻ na nuwo kata fe gͻmedzedze la Nya la li.Nya la kple Mawue li eye eyae nye Mawu (Biblia 2010; Nubabla Yeye La 1990) in Ewe.All four translations used nya la to translate logos except Biblia (2010) and Nubabla YeyeLa (1990)which paraphrased it.A translation of the logos in Ewe cosmic context would renders it, nyagbe, the intelligible word which dwells with the female pair of Sogbe-Lisa and manifests itself from the absolute state of existence to the embodied state.This concept resonates with the young wise lady in Proverbs 8 who is described as first in the order of creation and clothed with creative power, as alluded to in the works ofHarris (2003:465),Robertson (1960:3), and Lincoln (2005:96).It also resonates with Justin Martyr's assertion that the logos is 'the very principle of the intelligibility of the created world', incarnated in the person of Jesus(McGrath 2007:283; Ryan 2007:785; Wiles 1966:28; ANF02:177).The relationship between the logos and the Supreme Being is what the Evangelist expressed in the last phrase of verse 1, kai theos ēn ho logos (and the Word was God).As Erasmus opines in his work, '… the Word is a sharer in the divine essence, or…homoousion tó patri (of one substance with the Father)' (eds.Craig & Farmer 2014:15).Similar idea is expressed in the work of Ignatius of Antioch (1994) who argues that:[T]here is one God, the Almighty, who has manifested Himself by Jesus Christ His Son, who is His Word [logos], not spoken, but essential.ForHe is not the voice of an articulate utterance, but a substance begotten by divine power, who has in all things pleased Him that sent Him.(p.62) | 6,151.6 | 2023-12-22T00:00:00.000 | [
"Philosophy"
] |
A Biaxial Strain Sensor Using a Single MoS2 Grating
In this paper, we report a new type of MoS2-based grating sensor for in-plane biaxial strain gauges with a precision limit of ~ 1‰. The MoS2 grating is numerically simulated with different biaxial strains up to 5%. Our first-principles calculations reveal that the strain sensitivity of the MoS2 reflectance spectrum can be considered an additional strain sensor integrated with the grating structure, enabling the mapping of in-plane biaxial strains. Our experimental studies on a prototype MoS2-grating sensor further confirm that a strain component perpendicular to the grating period can cause intensity peak shifts in the grating’s first-order diffraction patterns. This work opens a new path towards the sensing of in-plane biaxial strain within a single-grating device. Our new approach is applicable for other materials that have predictable reflectance response under biaxial strains and the capacity to form a two-dimensional single-crystal layer. Supplementary Information The online version contains supplementary material available at 10.1186/s11671-021-03493-3.
Introduction
Flexible electronics technology has received widespread attention from the academic and industrial communities, but the design and application of microscale and nanoscale flexible devices is challenging due to difficulties in dynamic displacement and deformation monitoring [1][2][3][4][5]. Most conventional strain detection methods based on resistance strain gauges require a miniaturized sensor array [4,6,7], which is hard to produce for flexible electronic applications. Optics-based two-dimensional (2D) strain detection techniques, such as speckle interferometry, are superior to those based on piezoresistivity because of their higher precision [8]. However, their image correlation measurement strategy is challenged by the requirements of complex image processing technology [8][9][10]. A reflection grating can provide a high resolution for the strain measurement but lacks the capability to detect 2D strain within a single device [11].
The transition metal dichalcogenides exhibit outstanding optical and mechanical properties [25][26][27]. For instance, MoS 2 can tolerate as much as 19.5% [26] of biaxial strain accompanied by its reflectance modulation [28], and WSe 2 can show notable Berry curvature dipole as well as nonlinear Hall effect via strain engineering [29]. Incorporating the strain-sensitivity of a material's reflectance spectrum into the function of the reflection grating device can be an efficient way to extend strain measurements to biaxial detection within a single device. However, there are no reports of the reflection gratings combined with strain-induced material reflectance modulation for 2D strain sensing applications.
Here, we propose a new type of in-plane biaxial strain sensing technique involving the strain-sensitivity of MoS 2 reflectance in a reflection grating sensor. First-principles calculations reveal that biaxial strains can shift the peak of the intensity distribution in diffraction patterns of a MoS 2 -based grating device because the reflectance of MoS 2 is sensitive to the strain-induced deformation. This nonlinear peak shift is well demonstrated by adding a second-order term to the uniaxial-strain linear equation, from which the strain component perpendicular to the grating period direction can be extracted with a precision limit of ~ 1‰. Our experimental studies on a prototype MoS 2 -grating device confirm that the strain perpendicular to the grating period can induce an intensity peak shift of the grating's first-order diffraction pattern. Our research shows the possibility of one-shot, in-plane biaxial strain gauges with a single grating sensor.
Theoretical Calculations for MoS 2 Flake
The MoS 2 optical responses to the strain are all studied by first-principle calculations performed with the Vienna Ab-initio Simulation Package (VASP) [30]. All-electron projector augmented wave (PAW) potentials [31] were used for all calculations. Geometric relaxation and static calculations were carried out with the Perdew-Burke-Ernzerhof (PBE) generalized gradient approximation (GGA) method [32]. Spin-orbit coupling (SOC) [33] was also included in the static calculations. To determine the optical properties, quasiparticle energies were first obtained by the GW method derived from Green function theory [34]. The wave function obtained from the static calculations and the quasiparticle energy from the GW calculations were used to perform the Bethe-Salpeter equation (BSE) calculations [35] to predict the dielectric constant. A bulk crystal model was used to represent the MoS 2 flake (that the optical property differences between stacks greater than five layers is negligible [36]). In the optimization calculations, the energy cutoff was set to 400 eV, and a Monkhorst-Pack k-point set of 15*15*4 was used to sample the Brillouin zone. The lattice parameters were first optimized as a reference for later calculations of the strained MoS 2 flake. The optimized lattice parameters were a = b = 3.18 Å and c = 13.87 Å. The geometry was relaxed until the energy converged to 10 -5 eV. Adopting a technique from previous literature [37], an energy cutoff of 300 eV and k-point set of 6 × 6 × 2 were used in the optical calculations. The static energy converged to 10 -6 eV in all calculations. The diffraction patterns were simulated based on the Helmholtz-Kirchhoff theorem [38]. More details are provided in the Additional file 1.
Preparation of MoS 2 Sample for Spectrum Measurement
The MoS 2 thin film was mechanically exfoliated from a commercial MoS 2 single crystal (SPI Supplies) and transferred to a polydimethylsiloxane (PDMS) substrate with tape. After the transfer, another layer of PDMS was fabricated on the flake and substrate to enhance adhesion.
Preparation of MoS 2 Grating on the Flexible Substrate
The MoS 2 thin film was mechanically exfoliated from a commercial MoS 2 single crystal (SPI Supplies) and transferred to a polydimethylsiloxane (PDMS) substrate with tape. To fabricate the grating device, the MoS 2 flake was firstly patterned into a grating structure by electron beam lithography (EBL). Then the patterned sample was etched by oxygen plasma with a power of 20 W. Finally, we obtained the MoS 2 -based grating device by washing away the PMMA.
MoS 2 Device Measurements
A supercontinuum white light source (NKT Photonics SuperK Compact) is used as the excitation laser, which passes through one aperture and hits the flake sample or the grating sample at a certain angle relative to the sample plane, as shown in Fig. 1. In the reflectance measurement, the reflected laser is collected through an optic fiber connected to a spectrometer. The reflectance spectra under different strains are calculated from the data measured by the spectrometer. To test the MoS 2 grating, the reflected laser is projected onto a white board and
Results and Discussion
In a conventional reflection grating sensor, a periodic structure of parallel grating strips can diffract the light, and the diffraction is utilized to measure a strain that is along the grating period direction by monitoring a location shift of the diffraction patterns [11]. Due to the periodic structure orientation, the strain sensing function of the reflection grating is limited to the in-plane uniaxial strain gauge (parallel to the periodic direction). To extend the reflection grating function for use in in-plane biaxial strain gauges, we propose that the intrinsic optical properties of the grating material, such as the strain-sensitivity of the material's reflectance, can be considered as an additional strain sensor to detect in-plane strain components perpendicular to the periodic direction. MoS 2 has a layered structure: a layer of Mo atoms sandwiched between two layers of S atoms. The interaction between the layers is a weak van der Waals force. Here, we design a MoS 2 flake-based reflection grating sensor (Fig. 1) and investigate the device diffraction patterns under different in-plane biaxial strains by first-principles calculations. The incident beam wavelength range in our calculation is from 400 to 850 nm. The diffraction grating can be described by: where d is the distance between two adjacent grating strips, θ i is the angle between the incident beam and the normal to the grating, θ m is the angle between the diffraction beam and the normal when the diffraction beam has maxima, n is the diffraction order, and is the beam wavelength [11]. From Eq. (1), we see that incident beams with different must have different θ m . Therefore, a continuous wavelength beam causes a continuous series of diffraction spots corresponding to different θ m , forming an elliptical first-order diffraction pattern. Figure 2a shows the simulated image of the diffraction patterns of the as-designed grating sensor with no strain applied. Figure 2b shows the intensity peak and pattern location evolution of the simulated first-order diffraction pattern of the device under different biaxial strains. The edge of the first-order diffraction pattern corresponding to the incident beam of 850 nm is labeled "LW". When we apply an in-plane uniaxial tensile strain along the grating period direction ( ε x ), this strain can induce an increase in the spacing d between each strip. As a result, θ m decreases because dsinθ m is constant for any given and fixed θ i . Therefore, when we gradually increase the strain ε x from 0 to 4.3%, the location of each point in the firstorder diffraction pattern moves closer to the center of the zero-order diffraction spot in a proportional relationship There is asymmetric behavior between the first-order diffraction spot on both sides of the zero-order beam because the screen in our simulation is set to be parallel to the grating instead of perpendicular to the reflection direction. b Simulated evolution of first-order diffraction spot under different biaxial strains. The intensity partial differential is represented with colors. The horizontal coordinate and vertical coordinate denote the position relative to the center of the zero-order diffraction spot. The peak is marked with a dashed line. From left to right, ε x was set as 0%, 0.9%, 2.6%, and 4.3%, respectively. From top to bottom, the ε y was 0%, 1%, 3%, and 5%, respectively with the corresponding beam wavelength, which is consistent with the function of the conventional reflection grating sensor [11]. An incident beam with a longer wavelength has a larger θ m variation, so the LW edge has the most apparent location shift. However, when an in-plane tensile strain perpendicular to the grating period direction ( ε y ) is simultaneously applied, an intensity peak shift is observed within the first-order diffraction pattern, as marked by a dashed line in Fig. 2b. When the strain ε y increases from 0 to 5%, the intensity peak shifts further away from the center of the zero-order diffraction spot. We attribute this peak shift of the intensity distribution to the strain-induced modulation of the MoS 2 reflectance. Previous literature has reported that the reflectance spectrum of MoS 2 can be tuned by an external strain [28], and the reflectance is equal to the intensity ratio of the diffracted beam to the incident beam of the reflection grating. Therefore, the intensity of the beams with different wavelengths diffracted by the MoS 2 grating can be modulated by the in-plane strains. Meanwhile, no LW edge location shift occurs because the strain ε y exerts no impact on the grating period. Figure 3a shows the linear behavior in the peak shifts of the MoS 2 reflectance spectra when a uniaxial tensile strain along the lattice vector b of MoS 2 is applied. This uniaxial tensile strain causes a peak-position redshift in the MoS 2 reflectance. However, there is a nonlinear modulation in the reflectance peak position shift when we apply an in-plane biaxial tensile strain. The relationship between the peak position in the reflectance spectra and the in-plane biaxial tensile strain can be described by a second-order equation: (2) Peakposition = l(ε a + ε b ) + mε a ε b + n where l, m, and n are three constants, and ε a and ε b are the strain components along the two lattice vectors of the MoS 2 . The first term describes the linear behavior of the peak position shift under uniaxial tensile strains along lattice vector a or b . The second term describes the higher-order behavior in the biaxial tensile strain situation. The third term is the reflectance peak position of the unstrained MoS 2 . Since the MoS 2 lattice vectors a and b are symmetrically equivalent, the tensile strains along the two directions has the same contribution factor. The fitting results show that the highest difference between the fitting curve and the first-principle-calculated peak positions is 1.76 nm, which indicates that a strain-gauge precision limit of ~ 1‰ can be obtained when the reflectance peak position is utilized to calculate the strain with the Eq. (2). Figure 3b shows the mapping image of the reflectance peak position under different in-plane biaxial tensile strains obtained from the fitted Eq. (2) (see detailed plots of reflectance in Additional file 1).
In our simulation, the lattice vector a is perpendicular to the period direction of the simulated grating. Therefore, the strain ε y is equal to ε a , and strain ε x equals √ 3/2 × ε b . Our calculations reveal that in a MoS 2 -based grating sensor, the in-plane strain ε x can be measured by the LW edge location shift of the first-order diffraction pattern. Based on the intensity peak shift in the firstorder diffraction pattern, we can utilize the second-order Eq. (2) to subtract the contribution of the in-plane strain ε x from the peak shift. Then we can quantitatively calculate the in-plane strain ε y .
To further experimentally study the strain-sensitivity of the MoS 2 reflectance, we mechanically exfoliated a MoS 2 flake (thickness of several tens of nanometers; see details in Additional file 1) and attached the flake to a flexible substrate of polydimethylsiloxane (PDMS) by Fig. 4a inset). An inplane uniaxial tensile strain was imposed on this fabricated MoS 2 device by fixing the two sides of the substrate to two translation stages and stretching the substrate. We estimated the in-plane uniaxial tensile strain by calculating ε= δL/L , where L is the length of the substrate between the two clips and δL is the length change. When the strain is varied from 0 to 4%, there is a redshift of the peak position in the MoS 2 reflectance spectrum, and the magnitude of this shift agrees well with our theoretical calculations, as shown in Fig. 4a. Figure 4b, c show the optical image of a MoS2-based reflection grating sensor with a period of 2 μm on a PDMS flexible substrate fabricated by electron-beam lithography (details in Methods). Upon stretching the PDMS substrate, an in-plane tensile strain perpendicular to the period direction is exerted on the MoS 2 -based grating device (Fig. 4d). By monitoring the intensity distribution in the first-order diffraction pattern, we observed that the intensity peak shifts further away from the center of the zero-order spot compared to the unstrained case when we introduce an in-plane tensile strain of 4% perpendicular to the period direction (Fig. 4e). No diffraction pattern location shift is obtained because the tensile strain is perpendicular to the period direction, and the spacing d between each strip changes little.
Conclusion
In summary, we demonstrate a new technique for gauging in-plane biaxial strain using a MoS 2 -based reflection grating sensor. We test the concept by numerically simulating the grating with different biaxial strains up to 5%. In this new technique, the grating structure for detecting the strain component along the period direction ( ε x ) is combined with the strain-sensitivity of the MoS 2 reflectance to act as an additional sensor to obtain the in-plane strain component perpendicular to the period direction ( ε y ). Component ε y is calculated with a second-order approximation equation and the intensity peak shift within the first-order diffraction patterns. The theoretical results are well supported by our experiments. Our work opens a path for the design of flexible grating sensors and provides a novel approach to realize one-shot in-plane biaxial strain gauges with a single grating sensor. Our approach is also applicable for other materials that have predictable reflectance response under biaxial strains and the capability to form two-dimensional single-crystal layers. | 3,693.2 | 2021-02-10T00:00:00.000 | [
"Physics"
] |
Homologous recombination, cancer and the ‘RAD51 paradox’
Abstract Genetic instability is a hallmark of cancer cells. Homologous recombination (HR) plays key roles in genome stability and variability due to its roles in DNA double-strand break and interstrand crosslink repair, and in the protection and resumption of arrested replication forks. HR deficiency leads to genetic instability, and, as expected, many HR genes are downregulated in cancer cells. The link between HR deficiency and cancer predisposition is exemplified by familial breast and ovarian cancers and by some subgroups of Fanconi anaemia syndromes. Surprisingly, although RAD51 plays a pivotal role in HR, i.e., homology search and in strand exchange with a homologous DNA partner, almost no inactivating mutations of RAD51 have been associated with cancer predisposition; on the contrary, overexpression of RAD51 is associated with a poor prognosis in different types of tumours. Taken together, these data highlight the fact that RAD51 differs from its HR partners with regard to cancer susceptibility and expose what we call the ‘RAD51 paradox’. Here, we catalogue the dysregulations of HR genes in human pathologies, including cancer and Fanconi anaemia or congenital mirror movement syndromes, and we discuss the RAD51 paradox.
INTRODUCTION
Genomes are routinely challenged by exogenous and endogenous stresses, leading to genetic instability that can fuel oncogenesis (1,2). To preserve genome integrity, cells have developed the DNA damage response (DDR) that coordinates cell cycle progression and DNA repair.
Homologous recombination (HR), a process that is highly conserved throughout evolution, plays a prime role in genome stability/diversity. HR is involved in the repair of DNA double-strand breaks (DSBs) and DNA interstrand crosslinks (ICLs) and in the protection and resumption of arrested replication forks (3). In particular, HR suppression alters replication dynamics (4)(5)(6). Notably, the activation of the DDR has been observed during the pre/early stages of cancer as a result of endogenous replication stress (7,8), suggesting the potential role of HR as a replication escort in preventing cancer initiation. Moreover, DSBs and ICLs are also important sources of genetic instability, and thus the role of HR in their repair also favours the maintenance of genome stability. Therefore, HR is widely considered a tumour suppressor pathway.
In support of this notion, several HR genes are mutated in tumours. Paradoxically, despite extensive studies, the inactivation of RAD51, which performs the sequence homology search (i.e. the central step of HR that gives the process its name, see Figure 1A) has not been found to be related to cancer development. Notably, RAD51 inactivating mutations are absent in familial breast and ovarian cancer ( Figure 1B). In contrast, overactivation or overexpression of RAD51 has been described in different types of cancer. Moreover, germline mutations in several HR genes are responsible for subgroups of Fanconi anaemia (FA), a rare autosomal recessive syndrome leading to developmental defects and malignancies (9). However, to date, there has been no case of a patient with FA-R (RAD51 mutation) developing cancer. Hence, compared to all its HR partners, RAD51 is an enigma, especially considering its central role in HR (see Figure 1A). Moreover, the fact that RAD51 is an essential gene in mammals has hindered in vivo analysis of RAD51 functions.
Here, we discuss the relationships between HR genes and human pathologies, including cancers. Then, we discuss the 'RAD51 paradox'.
MUTATION AND DOWNREGULATION OF HR GENES IN HUMAN PATHOLOGIES: THE RAD51 PARADOX
HR is downregulated in different cancer contexts, including germline mutations in inherited breast or ovarian cancer ( Figure 1) or in FA subgroups. Some RAD51 mutations have been detected in different types of sporadic cancer, but whether they cause tumourigenesis remains unclear (10). Finally, RAD51 mutations have been identified in congenital mirror movement syndrome but were not associated with cancer predisposition (11,12).
Familial breast and ovarian cancer
Heterozygous germline mutations in different genes confer predisposition to breast or ovarian cancers ( Figure 1B) (13). Genes involved in DDR appear to be markedly overrepresented; specifically, the main represented pathway is HR ( Figure 1B). Indeed, several genes that directly control HR are mutated (BRCA1, BRCA2, PALB2, RAD51C, RAD51D, BARD1, BRIP1, MRE11, RAD51, NBN), as are other genes that indirectly impact HR, such as the DDRcontrolling genes ATM and CHEK2. The association between mutations of the HR mediators XRCC2 and XRCC3 and cancer predisposition remains controversial. Moreover, TP53 precludes HR independently of its roles in the cell cycle and apoptosis (for review, see (14)). PTEN has also been proposed to compromise HR (15,16), but this remains contentious (17,18). As a consequence, the HR genes BRCA1, BRCA2, PALB2, RAD51C, RAD51D, BARD1, BRIP1, MRE11, RAD51 and NBN are included in several hereditary breast/ovarian cancer screening panels to evaluate the tumour for HR deficiency and predict its response to chemotherapy (19).
However, in spite of the importance of HR alteration in hereditary breast/ovarian cancer, mutations of the pivotal HR player RAD51 are surprisingly absent from the lists of genes predisposing individuals to breast or ovarian cancer ( Figure 1A and B).
While many cancer-related mutations affect HR genes, only two germline mutations in RAD51 have been identified, and their impact on cancer risk remains to be established. The RAD51 E258A mutation is a dominant negative germline variant that was identified in breast carcinoma and maps to the interface region between the Nterminal and RecA homology domains of RAD51 (20). The RAD51 R150Q mutation leads to reduced ssDNA and ds-DNA binding abilities (21). However, the mutation is not clearly associated with cancer incidence (22). Therefore, unlike its direct HR partners, RAD51 has not been classified as causal for cancer. Consequently, RAD51 is not included in genetic diagnostic tests or in screening for tumour chemotherapy response.
Nevertheless, both FA syndrome and familial breast and ovarian cancer reveal the association of HR gene germline mutations with cancer susceptibility, with the exception of the central HR player RAD51.
Sporadic cancers
Mutations of HR genes in sporadic cancers. In addition to being implicated in hereditary cancers, mutations in HR genes were also observed in a large panel of sporadic cancers. Two recent studies in 64 791 and 113 927 women respectively confirmed that mutations in HR genes BRCA1, BRCA2, PALB2, BARD1, RAD51C and RAD51D correlate with increased breast cancer incidence (26,27). More generally, the sequencing of a panel of 52 426 tumours, including melanoma, hepatocellular carcinoma, and endometrial, gastroesophageal, ovarian, colorectal, biliary tract, bladder, breast and pancreatic cancers, showed that 15-34% of tumours exhibited mutations in HR or DDR genes (10). Interestingly, in this panel, only the BRCA1, BRCA2, PALB2 and BRIP1 genes were found to be mutated, and no mutation was observed in BARD1, RAD51C or RAD51. The BRCA1 and BRCA2 genes are mutated in 5-6% of breast cancers and 16% and 6% of ovarian cancers, respectively. Mutations in not only BRCA1/2 but also RAD51C, RAD51D, PALB2, BARD1 and BRIP1 are now also screened in clinics to identify HR deficiency, which has recently become a key criterion for treatment with PARP inhibitors. It is noteworthy that no screens include RAD51, as it does not appear to be a frequently mutated gene in these pathologies. In some isolated cases, mutations of RAD51 have been found in tumours. Most of these mutations were classified as variants of unknown significance, but some were functionally characterized (F86L, D149N, Q268P, Q272L) and determined to affect one of the functions of RAD51 (ATPase activity, DNA binding, strand exchange activity and/or thermal stability) (28). However, the HR process itself was not evaluated in cells (for review see (29)). All of the mutations were found in breast cancers, with the exceptions of Q268P (lung cancer) and Q272L (kidney cancer). An additional variant, RAD51 G151D, was found in a triple-negative breast cancer (30,31), but remarkably, this mutation leads to overactivation of RAD51 activity instead of inactivation (30). Finally, it is not known whether these few RAD51 mutations are actually causal for cancer.
Expression of HR genes in sporadic cancers. Expression of HR factors may often be downregulated by promoter methylation rather than (or in addition to) mutations; for instance, the BRCA1 and BRCA2 promoters are methylated in 20% and 5% of epithelial ovarian carcinomas, respectively (32).
In contrast, overexpression of RAD51 is common in many cancers, including cervical, non-small cell lung, breast, ovarian and pancreatic cancers, melanoma and glioblastoma. This overexpression is associated with poor prognosis as a consequence of increased ability to repair lesions induced by DNA-damaging therapeutic agents (33)(34)(35)(36)(37)(38); cancer cells overexpressing RAD51 could be selected during tumour progression because of this survival advantage. RAD51 overexpression of is generally due to overactivation of the promoter in cancer cells (38). In contrast with the general pattern of HR-inactivating mutations promoting breast cancer, RAD51 overexpression is associated with poor prognosis (https://www.proteinatlas.org). Therefore, RAD51 is a pharmacological target, and RAD51 inhibitors are being developed (29).
Regulators of HR are dysregulated in cancer. AKT1 is an oncogenic kinase that is activated in numerous types of cancers. AKT1 exercises its oncogenic activity through the stimulation of proliferation associated with the inhibition of apoptosis. Remarkably, AKT1, which is negatively regulated by PTEN (one of the genes mutated in hereditary breast and ovarian cancer), also inhibits HR through the cytoplasmic sequestration of BRCA1 and RAD51, resulting in at least a BRCA1 defective-like phenotype (39). The fact that AKT1 is activated in 40-60% of sporadic breast or ovarian cancers establishes a potential metabolic link between familial and sporadic breast and ovarian cancer (40).
BCL-2 is one of the most important antiapoptotic genes, and although it facilitates tumour cell survival and proliferation, overexpression of BCL-2 has been consistently associated with good prognosis (50). BCL-2 downregulates HR through the mislocalization of BRCA1 to the mitochondrial membrane (41,42). It may be proposed that the enhanced survival resulting from apoptosis inhibition is at least in part compensated by a lower capacity to resist treatment.
A translocation between chromosomes 9 and 22 (also known as the Philadelphia chromosome) resulting in expression of the BCR-ABL tyrosine kinase is found in chronic myelogenous leukaemia (CML). BCR-ABL expression leads to overexpression of RAD51 and several RAD51 paralogues via STAT5-dependent transcription and inhibition of caspase-3-dependent cleavage, thus resulting in a hyper-RAD51-like phenotype (43).
TP53 is the most frequently mutated gene in cancers. Since TP53 precludes HR, particularly targeting RAD51, mutation of TP53 also leads to elevated HR levels (for review, see (14)).
Congenital mirror movement syndrome
Congenital mirror movement (CMM) disorder is a rare disorder that impairs the patient's capacity to perform normal daily tasks that require bimanual coordination (11,12). Multiple CMM patients (n = 32) have been reported to have RAD51 mutations, usually RAD51 haploinsufficiency. These patients did not show additional developmental malformations, and no mutation in other HR factors has been found in CMM. In an investigation of the function of RAD51 in the context of CMM, a study using primary mouse cortical neurons suggested that RAD51 can negatively regulate neuronal axon growth (44). Importantly, CMM patients bearing RAD51 mutations (but no mutations of other HR genes) do not exhibit cancer predisposition.
The authors propose a cytoplasmic function for RAD51 in neuronal guidance, but the details of this mechanism remain to be characterized. Nevertheless, these data suggest an additional specific function of RAD51 that might contribute to the RAD51 paradox.
THE RAD51 PARADOX IN CANCER: HYPOTHESES AND SPECULATIONS
Several mechanisms that can operate together or in parallel might account for the RAD51 paradox: (i) RAD51 functions that are independent of RAD51 mediators/loaders. CMM syndrome suggests that RAD51 could have BRCA/PALB2-independent functions, potentially cytoplasmic. The alteration of these functions, in addition to defective genome maintenance, might be so toxic that cancer cell proliferation cannot occur. Moreover, RAD51 exhibits some BRCA-independent DNA processing activities.
Indeed, in addition to its function during DSB repair, RAD51 has important functions during replication, which can be independent of its classical loading factors. Blocked replication forks can be regressed, generating a so-called 'chicken foot' structure, and RAD51 is thought to directly promote such structures in a BRCA2-and PALB2-independent manner (45)(46)(47)(48). The protein complex MMS22L-TONSL is recruited to blocked replication forks, and its inactivation decreases survival after replication fork stall. This complex is involved during replication stress-mediated HR and gene conversion (49)(50)(51). MMS22L-TONSL binds directly to RAD51, and its inactivation decreases RAD51 recruitment to blocked replication forks (50). It was proposed that MMS22L-TONSL could load RAD51 onto the ssDNA in blocked replication forks, replacing RPA through a BRCA2/PALB2-independent mechanism (46,50). Based on this evidence, one could propose that because RAD51 possesses both BRCA1/2/PALB2-dependent and BRCA1/2/PALB2-independent functions during replication, its inactivation would be more toxic than that of one of the loading factors, thus impairing the proliferation of tumour cells. (ii) Stimulation of alternative non-conservative DNA repair pathways. The choice of the appropriate DSB repair pathway is important for genome stability maintenance. Different processes cooperate or compete to achieve DSB repair, and inactivation of HR can have several different outcomes and consequences depending on the process involved. The selection of the DSB repair mechanism occurs in two steps ( Figure 2). First, competition between canonical non-homologous end-joining and resection that generates the 3 ssDNA; second, competition between HR and single-strand annealing (SSA) or alternative end-joining (A-EJ) on the 3 ss-DNA (52,53). Importantly, HR is mainly conservative, while SSA and A-EJ are non-conservative and cause genome instability because they ineluctably involve the loss of the sequences surrounding the DSB. RAD51 performs its role(s) in genome stability maintenance at the second step through both enzymatic and non-enzymatic processes ( Figure 2). The enzymatic strand exchange activity of the ssDNA/RAD51 filament triggers the search for homology and strand exchange with a homologous DNA partner. This ultimately leads to HR-mediated repair and resumption of arrested replication forks, which are conservative pathways that preserve genome stability. In parallel, RAD51 protects arrested replication forks and prevents nonconservative DSB repair processes through DNA occupancy (Figure 2), independent of strand exchange activity, in a non-enzymatic manner (54,55). Indeed, the loading of a RAD51 molecule that cannot perform strand exchange still protects against extensive degradation of arrested replication forks (54) as well as the annealing of complementary ssDNA (55), the central step of non-conservative SSA and A-EJ (see Figure 2). Consequently, the absence of RAD51 protein on damaged DNA may result in the stimulation of alternative non-conservative DNA repair pathways that increase genetic instability. Therefore, ablation of the mediators/loaders of RAD51, i.e., the partners of RAD51, results not only in the suppression of conservative HR but also in the concomitant stimulation of non-conservative SSA and A-EJ (55)(56)(57)(58). In addition, the stimulation of alternative repair pathways allows partial rescue from the cell viability defects caused by the failure of HR, thus increasing the viability of cells with increased genetic instability. Therefore, it can be suggested that HR ablation alone would not be sufficient for tumour development and/or viability and that the association with the stimulation of the nonconservative pathways would be required for tumourigenesis. PolQ is overexpressed in cancers and is predictive of a poor prognosis (59). The fact that PolQ removes RAD51 from the filament, thus suppressing HR (60), supports the idea that HR suppression is tumourigenic. However, by removing RAD51 from the DNA, PolQ also stimulates A-EJ (60), overexpression of PolQ might be associated with both the inhibition of conservative HR and non-conservative pathway stimulation (A-EJ).
In this context, the consequences for cancer outcome might depend on the way HR is inhibited and the presence versus the absence of RAD51 protein on damaged DNA. According to this hypothesis, only mutations that suppress RAD51 expression or its loading and/or stabilization on damaged DNA (such as mutations in BRCA1, BRCA2 or PALB2 or other HR factors) should confer cancer predisposition. Note that another essential gene, KNL1 (Kinetochore scaffold 1), which is involved in mitotic spindle assembly and the associated checkpoint, is close to the RAD51 gene on chromosome 15 (both genes are in cytogenetic band 15q15.1). Homozygous codeletion of these two genes is likely too deleterious for cell viability and especially unlikely to be able to support high levels of proliferation as in cancer cells. This would then decrease the frequency of possible RAD51 deletions. Moreover, the hypothesis implies that RAD51 missense mutations that abolish HR but retain the DNA-binding capacity of RAD51 should not promote cancer. These factors would decrease the probability and frequency of carcinogenic RAD51 mutations, especially compared to those of its HR partners, which are not subject to such restrictions.
Finally, we cannot exclude the possibility that other reasons (e.g., different stability of mutant RAD51 proteins or mRNA, different sensitivity of tumours to detection by the immune system, and many other possibilities. . . ) that are not yet identified or characterized might also participate in the RAD51 paradox.
(iii) Overexpression of RAD51. Although RAD51 is not (or is rarely) found to be inactivated in cancer, its overexpression has been described in a wide variety of cancers, leading to poor prognosis (33)(34)(35)(36)(37)(38). One hypothesis is that RAD51 overexpression confers resistance to chemotherapeutic agents that target DNA, accounting for the selection of such cells and for the poor prognosis (33)(34)(35)(36)(37)(38). This hypothesis implies that overexpression of RAD51 initiates strand exchange events that result in increased HR activity, and thus resistance to DNA damaging agents ( Figure 2D). One can propose that, in addition, unresolved HR intermediates accumulate, which can favour genetic instability (61). Finally, overexpression of RAD51 might disequilibrate other RAD51 functions, including cytoplasmic roles. This could alter cell homeostasis, fostering tumourigenesis and/or tumour progression ( Figure 2D).
CONCLUSIONS AND PERSPECTIVES
HR plays a central role in genome stability maintenance. Its suppression generates genome instability and should thus confer cancer predisposition. However, analysis of human pathologies reveals that the causal connection is in fact complex. Indeed, although suppression of most HR factors does trigger carcinogenesis, one important gene, RAD51, escapes this pattern. This is particularly remarkable because RAD51 promotes the pivotal steps of HR, i.e., homology searching and strand exchange with a homologous DNA partner. These steps quite literally define the HR process. In contrast, RAD51 has been found to be overexpressed or overactivated in tumours, which might result from the selection of tumour cells that are resistant to treatment with genotoxic agents. This suggests that the amount of RAD51 protein is a limiting factor for HR. Mouse models should represent useful tools to experimentally address these questions in vivo. Unfortunately, most of the genes involved in the central step of HR, including RAD51, are essential, and their homozygous deletion leads to embryonic lethality in mice (62,63). To overcome these problems, elaborate strategies for partial HR or tissue-specific deletion have been designed (62). These models have confirmed that defects in genes that participate in different steps of HR trigger tumourigenesis. Surprisingly, despite its critical role in HR, there are no such alternative mouse models available for RAD51. Addressing the function of RAD51 in vivo will likely be key to solving the RAD51 paradox.
Different factors can be combined to account for the RAD51 paradox (see above). Among these factors, the presence versus absence of RAD51 on damaged DNA could influence commitment of a cell to a cancerous pathway. Indeed, the HR genes affected in cancer promote the formation and/or stabilization of the ssDNA/RAD51 filament. Depletion of these HR genes results in inefficient formation or stabilization of the ssDNA/RAD51 filament, making DNA accessible to alternative non-conservative repair processes that lead to increased genetic instability associated with partial compensation of the decreased viability. This raises the provocative question of whether HR suppression actually promotes cancer per se or whether oncogenesis in fact results from the stimulation of non-conservative pathways (or a combination of both). Addressing this question will be important to resolving the RAD51 paradox. Studying tumourigenesis when HR is inhibited in the absence of stimulation of alternative non-conservative pathways, when the non-conservative pathways are stimulated without altering HR, or when both pathways are inhibited should provide important clues to understanding why RAD51 disruption is not commonly found in cancers. These concepts will also be important to the design of strategies targeting RAD51 activity in cancer therapy. While such strategies should efficiently sensitize RAD51-overexpressing tumours to radiotherapy and chemotherapy, their use in other types of tumours might be more problematic because of the potential stimulation of non-conservative DNA repair pathways. An ideal strategy would be to repress HR and, in parallel, inhibit or at least ensure not to stimulate the nonconservative DNA repair pathway, thus avoiding the rescue of cancer cell viability and increased genetic instability.
BRCA2-independent functions of RAD51 may also play important roles in the RAD51 paradox. The full identification and characterization of RAD51 functions and the consequences of their inactivation on cell viability and carcinogenesis should also be informative for solving the RAD51 paradox. | 4,792.2 | 2021-04-09T00:00:00.000 | [
"Biology",
"Medicine"
] |
Should co-payments for financially deprived patients be lowered? Primary care physicians’ perspectives using a mixed-methods approach in a survey study in Tokyo
Background Japan’s stagnant economy has produced increasing income disparities, and the effect of poverty on health and health-care utilization is a significant issue. Co-payments could be a trade-off for patients when seeking medical care and limit health-care utilization. We sought primary care physicians’ experiences in dealing with financially deprived patients and their perspectives about reducing co-payments by low-income patients. Methods We used a convergent mixed-methods approach in a mail survey; it was distributed to 1989 primary care physicians practicing in areas with the highest proportions of socially disadvantaged individuals in Tokyo. The survey items included an open-ended question, seeking the participants’ perspectives about reducing co-payments by low-income patients from the current 30%, and closed questions, asking their experience of patient behavior related to financial burdens during the previous 6 months. Results We analyzed the responses of 365 physicians. Sixty-two percent of the primary care physicians agreed with lowering co-payments for financially deprived patients; however, the remainder disagreed or were uncertain. Those who disagreed were less likely to have experienced patient behavior related to financial burdens. The participants suggested challenges and potential measures for reducing co-payments by low-income patients in light of tight governmental financial resources and rapidly increasing health-care expenditures in Japan. The physicians were also concerned about the moral hazard in health-care utilization among patients receiving social welfare who obtain care at no cost. Conclusions From their experience in having dealt with low-income patients, the majority of physicians were positive about lowering co-payments by such patients; the remainder were negative or uncertain. It may be necessary to raise awareness of patients’ socioeconomic status among primary care physicians as a possible deterrent for seeking care. To maintain health-care equity, policy makers should consider balancing co-payments among individuals with differing financial levels and health-care needs.
Background
Japan has been internationally renowned for its universal health coverage, which has long enabled equitable access to health care [1,2]. However, owing to the country's longstanding stagnant economy [3], income disparities are increasing in Japan; accordingly, the impact of poverty on health and health-care utilization has become a highlighted issue [4]. Japan's relative poverty rate has continued to rise since 1985, and its highest-ever figure of 16.1% was recorded in 2012 [5]. In that year, one in six Japanese lived in relative poverty, with incomes of under half of the median equivalent disposable income, i.e., 1.22 million yen (about US$11,700) [6]. Since its launch in 1961, universal health coverage has undergone multiple reforms to meet rapidly changing social needs. Under Japan's current policy, all patients aged under 70 years have to pay 30% of the total health-care costs-regardless of their income-as a co-payment every time they see a primary care doctor [1]. The same copayment applies to outpatient and inpatient care, prescriptions, diagnostic tests, and surgery. Since copayment is not a fixed amount, patients with complex, multiple long-term conditions face a greater financial burden when receiving medical care. A reimbursement system (High-Cost Medical Expense Benefit) exists for medical payments that exceed a set monthly threshold; that system was revised in 2015 to be in line with a patient's household income [7]. For example, patients do not have to pay more than 57,600 yen (about $550) a month out of their own pockets if their household income is below 3.7 million yen (about $35,500) [7]. The exception is for individuals receiving public social welfare: they can obtain medical care at no charge, all such care being covered by the government. In Japan, people receiving social welfare comprised 1.7% of the total population in 2014 [8]. Most individuals in relative or absolute poverty did not receive social welfare: some failed to meet the criteria for receipt; however, many others did not choose to receive it owing to the social stigma or sense of shame.
A current issue is the non-payment of health insurance premiums through financial burdens on deprived households [2]. Specifically, among individuals eligible for the government-run health insurance program (covering almost 40% of the total population; most of those eligible are self-employed or unemployed), 16.7% do not pay the premium regularly [9]; that could result in losing eligibility for insurance. In addition to people at risk of being uninsured, insured individuals with financial insecurity may refrain from receiving medical care owing to the required co-payment at each health-care visit.
Co-payments may operate as a trade-off for patients when seeking medical care, and they affect the amount of health-care utilization [10]. In countries with universal health coverage, such as France, patient cost sharing has been reported to reduce the frequency of physician visits; that was especially true for patients in low social classes [11]. As a result of rapid increases in patient charges when consulting a general practitioner before 1995, more than half of financially deprived patients in Sweden reported that they had forgone medical care owing to costs at least once during the previous year; those who had forgone care perceived worse personal health conditions [12]. In Japan, reports differ as to how often people forgo medical care; however, it is known that low-income individuals more likely forgo or delay seeking care owing to costs compared with the high-or middle-income population [13][14][15]. The socially deprived population, though, has poorer health [16] and more unmet care needs [17][18][19]. To achieve optimal outcomes, patients with chronic illnesses, such as hypertension and diabetes, particularly need to make regular health visits and adhere to treatment. Cost-related medication non-adherence is strongly associated with worse health outcomes, such as hospitalization and death in patients with cardiovascular disease [20,21].
In Japan, physicians are not legally allowed to make repeat prescriptions without seeing the patient [22]; thus, patients must see their doctor every time they need a prescription. For financially deprived patients, the cost and benefit of seeking care is a trade-off; accordingly, if patients are asymptomatic, non-adherence to treatment can easily occur [23][24][25]. That being the case, primary care physicians may be the first to notice that a patient's financial burdens are hindering them from receiving necessary medical care. Those physicians should be responsible for arranging a more acceptable treatment plan with patients to prevent loss to follow-up and nonadherence [26][27][28]. Thus, primary care physicians' perspectives and attitudes regarding the care of low-income patients play an important role in mitigating the inequality of health-care utilization related to economic status.
To the best of our knowledge, no studies have addressed this topic in Japan. Accordingly, to examine experiences and perspectives in seeing patients with financial issues, we conducted a survey of primary care physicians practicing in especially deprived areas in Tokyo; we obtained both quantitative and qualitative data. We have previously published a paper using part of the results of that survey. We reported that approximately 90% of the surveyed physicians had experience of seeing patients with financial burdens; they had made efforts to meet patient expectations by somehow lessening those burdens with respect to the medical care they provided [29]. In the present paper, we use the remaining survey results to report primary care physicians' perspectives about lowering the copayment amount for low-income patients from the current 30%. Employing a convergent mixed-methods approach, we compared the themes derived from the qualitative responses using quantitative data from the respondents to expand our understanding of physician characteristics and perspectives.
Design, setting, and participants
We conducted a mail survey among primary care physicians practicing in especially deprived areas of Tokyo from July to September 2014. We used a convergent mixed-methods approach, whereby we collected qualitative and quantitative data simultaneously and then integrated them for interpretation [30]. We selected the survey areas as the 12 municipalities with the highest proportions of residents on social welfare among the 69 in Tokyo. The participants were physicians working in clinics in those areas that included the specialty of internal medicine; we requested that one physician of the highest rank respond on behalf of each clinic. We distributed 1989 questionnaires.
Ethical considerations
Before answering the questionnaire, the respondents read the explanation of the survey and understood that participation was voluntary. We considered return of the completed questionnaire anonymously to signify consent to participate. After returning the questionnaire, each respondent received a gift card equivalent to 1000 yen (about $9.5). The study protocol was approved by the institutional review board of Nippon Medical School (No. , Tokyo, Japan.
Survey items
We included an open-ended question, asking the participants' perspectives about lowering the co-payment amount for low-income patients from the current 30% in the context of hypothetical legislature. We also included the following closed-ended questions to assess physicians' experience and background. First, we asked if during the previous 6 months, they had experience of patient behavior related to financial burdens, such as loss to follow-up, medication non-adherence, postponing necessary tests, or refusing specialist referrals. Second, we assessed whether the physicians explained medication costs to patients in terms of "always," "usually," "sometimes," "seldom," or "never." Third, we evaluated the self-perceived decisionmaking style using three categories: paternalism, informed consent, and shared decision making. Paternalism signified the physician deciding without asking the patient's opinion or the patient consenting to the decision made by the physician without being presented with other options. Informed consent referred to the patient consenting to what the physician thought best after being presented with alternative options. Shared decision making signified the presentation of possible treatment options and the physician discussing the matter with the patient before the decision being made. We also collected physician demographics, such as age, sex, and specialty.
Analysis and integration of qualitative and quantitative results
We first compared the demographics of the respondents in the open-ended question with the overall survey respondents. To derive underlying themes, we then thematically analyzed the text responses to the open-ended question (qualitative data), which led to the development of several categories. We held discussions to agree upon the development of themes and categories. Based on the developed categories, we chose two groups of respondents: those who responded "agree" or "necessary" and those who responded "disagree" or "unnecessary" to the qualitative question about lowering the co-payment amount for low-income patients. We compared the characteristics of those two groups according to the quantitative data using chi-square tests, i.e., demographics, experience of seeing patients with financial burdens, explaining medication costs to patients, and decision-making styles. We performed statistical analyses using IBM SPSS Statistics version 23 (IBM Corp. New York, USA). We set statistical significance at P <0.05. Finally, as the integration phase in the mixed-methods approach [31], we expanded our understanding of the patterns in the physician characteristics and their perspectives based on our interpretation of both the qualitative and quantitative data.
Results
Among the 617 questionnaires returned (response rate, 31.0%), 550 were complete and used for the overall analyses. Among those, 365 participants responded to the open-ended question, and we used their data for qualitative analyses. Table 1 shows the characteristics of the two groups; no remarkable differences are evident between them. Over half of the participants had experience of patient behavior related to financial burdens during the previous 6 months, such as loss to follow-up and requests to postpone necessary tests. In terms of decisionmaking styles, the majority of respondents employed paternalism or informed consent; shared decision making was uncommon. More than half of respondents explained medication costs to patients sometimes or most of the time. Table 2 presents the primary care physicians' perspectives about lowering out-of-pocket payments for low-income patients. They are grouped into four categories: agree/ necessary; disagree/unnecessary; possible measures and challenges for implementing the legislation; and need to change the current redemption system of providing free care for those on social welfare.
Qualitative results
First, 62.2% of participants (227/365) responded that they agreed it was necessary to lower out-of-pocket payments for low-income patients not on social welfare. The participants mentioned their experience as physicians of losing patients to follow-up, medication non-adherence, and refusing necessary tests, referrals, or treatments. The respondents recognized the relationship between financial burdens and patient behavior as well as the importance of ensuring equal access to health care. The participants stated that they were trying to mitigate the burden on patients by making the consultation interval longer and choosing cheaper drugs. They stated that they felt frustration when negotiating with patients to accept proper care. Further, the physicians recognized that better health could help patients obtain a better-paid job, thereby alleviating financial burdens, and that receiving medical care and health education were necessary in that regard. Moreover, the participants considered that ensuring necessary medical care for low-income patients would prevent worsening of their conditions, resulting in reduced future health-care costs.
Second, one group of physicians (76/365) did not agree with the necessity of lowering out-of-pocket payments. Those doctors considered that patients should share an equal burden in receiving medical care and that lowering the payments for low-income patients would cause inequality with that burden. They also believed that patients should assume responsibility in taking care of their physical condition and improve their lifestyles, such as with smoking and alcohol consumption, which are associated with low incomes. Those physicians regarded payment for medical care as a kind of brake and that it motivated patients to improve their lifestyles. Some physicians thought there was no need or merit in lowering co-payments: they noted that various measures were already in place to mitigate financial burdens for patients with intractable diseases and that utilizing existing measures was sufficient; introducing new measures was unnecessary.
Third, the remaining participants (62/365) were uncertain about the feasibility of legislation to lower out-of-pocket payments for low-income patients; they believed that its implementation would present a number of challenges. Most of all, these physicians were concerned that any measures to mitigate the burden for low-income patients would be impractical owing to the Japanese government's tight financial resources; the physicians believed that such measures would create a dilemma when treating such financially stressed patients. These respondents considered that when discussing affordable treatment with low-income patients so as to avoid loss to follow-up or medication nonadherence, their comments would express disappointment or a sense of hopelessness with the government's financial constraints. The physicians in this group suggested potential measures, such as limiting the redemption for specific diseases and expensive forms of treatment, rather than applying redemption for all patient care costs regardless of the treatment burden. For example, costly medications for diabetes, including insulin, would be particularly problematic in the case of poor adherence. The potential risk of overutilization of health care was another concern raised: some physicians believed that providing free care should be avoided by all means; they mentioned the current issue of health-care overutilization by patients on social welfare who receive care for free.
Fourth, some physicians (50/365) expressed the opinion that some action should be taken to address the current policy of patients on social welfare having to pay no costs. There is a large discrepancy in the out-of-pocket expenses for visiting a doctor among financially deprived patients: low-income patients have to pay 30% of the total cost, whereas patients receiving social welfare do not have to pay at all. Patients and physicians find this gap in the payment burden to be unfair. On a daily basis, physicians observe the moral hazard in health-care utilization by people receiving social welfare: they stated that such patients take the free care for granted. They also pointed out the existing moral hazard for physicians, noting that some physicians tend to order unnecessary tests or overutilize interventions; that is partly because patients demand them and partly because physicians can simply earn more based on the fee-for-service reimbursement system for outpatient care. Table 3 shows the quantitative results, comparing the two groups who responded "agree" or "necessary" and "disagree" or "unnecessary" to the qualitative question we asked. There were no significant differences between the two groups in terms of age, specialty, or decision-making style. However, physicians who disagreed with lowering out-of-pocket expenses for low-income patients or considered it unnecessary were less likely to have had experience of patients' loss to follow-up during the previous 6 months (42.1% vs. 58.0%, P = 0.016), which is one of the common behaviors of patients related to financial burdens. Moreover, those physicians were less likely to explain the cost of medication to patients always, usually, or sometimes (43.4% vs. 58.1%, P = 0.026).
Discussion
This study examined physicians' experiences and perspectives regarding patients with financial issues. To the best of our knowledge, it is the first investigation to explore Japanese physicians' perspectives about reducing copayments for low-income patients from the current 30%. The majority of respondents agreed with the necessity of such a reduction, though a minority considered it unnecessary. Participants regarded the feasibility of Agree/necessary Equal access to health care "High out-of-pocket payments prevent patients from seeking necessary medical care, which leads to poorer health." "If they are asymptomatic, low-income patients soon stop taking medications. Those patients definitely need health education." Better health necessary for better work status "Receiving appropriate health care will enable patients to get a better job, which will lead to a better life." Reducing future health-care costs "Poorer health will lead to more ER visits and hospitalization, which will eventually result in increased overall health-care costs." Disagree/not necessary Sharing equal burden in receiving care "To avoid inequality and misuse of health-care benefits, the rate charged should be the same for everyone." Responsibility in taking care of one's own health condition and improving lifestyle "If patients pay less for medical care, they will be less motivated to improve their lifestyles." "Patients should spend more on medical care and quit spending on cigarettes and alcohol." No need or merit "The existing redemption systems are sufficient to protect the poor. Utilizing those systems should be enough." "Other measures could be used to mitigate the burden for low-income patients." Challenges and potential measures for implementing the legislation Tight government resources for health-care expenditure "Any additional increase in health-care expenditure will make the government unable to maintain universal health coverage." "Some measures need to be undertaken about the issues of high, rising drug costs." Limiting the redemption for specific diseases and expensive treatment "Expensive treatment, including diabetes medication such as insulin and some anti-cancer drugs, should be affordable to all patients who need it." Potential risk of overutilization "Free or low out-of-pocket payments might cause overutilization of medication. That is seen in patients on social welfare, who can get medicines for free, and should be avoided." Necessity to change the current policy of patients on social welfare paying no charges Wide gaps in the payment burden between those on social welfare and those not "Even if their income is at the same level, people on social welfare can receive care for free, but others are charged the 30% co-payment. This gap is too big. Patients on social welfare should pay something for receiving care." Moral hazard in health-care utilization among patients on social welfare "Free care is problematic. Patients take it for granted." "Patients come to see doctors too often and demand unnecessary tests and medications." implementing this reduction an area of huge concern. The physicians who agreed with the reduction or believed it necessary were more likely to have had experience of patients with financial burdens than those who disagreed with it or believed it unnecessary. Over the half of the physicians had experienced patient behavior related to financial burdens, such as loss to follow-up and medication non-adherence during the previous 6 months. These results indicate that the physicians believed the Japanese health-care system to be inequitable for the poor. Most of the primary care physicians who realized that patients' economic status could deter them from undergoing necessary medical care discussed the situation with the patients so as to keep the out-of-pocket payments as low as possible toward avoiding loss to follow-up or nonadherence. Though the physicians aimed to provide the best possible care, they also felt frustrated and powerless because the patients' economic status could not improve.
It was remarkable that some physicians clearly considered lowering the co-payments for financially deprived patients to be unnecessary. Combining our qualitative and quantitative data, we could infer that such physicians were less likely to explain the cost of medication to patients and may have missed signs of the patients' wish to have the medical care-related costs reduced as much as possible. The idea that it is an individual's responsibility to maintain good health through appropriate behavior and to take care of their own physical condition seemed to be prevalent among the participants; physicians who thought this way believed that all patients should accept the same financial burden for medical care regardless of their physical or financial conditions. This notion is related to "victim blaming" [32] and is incongruent with current understandings about the social determinants of health [33,34]. One possible reason for this result is the diverse backgrounds of primary care physicians in Japan. Nearly 80% of respondents in this survey were aged 50 years or more, and may not have been exposed to Engel's biopsychosocial model [35,36]. Most of the primary care physicians in Japan had previously been trained as subspecialists but became primary care physicians in their later careers, which has been observed and called a "Two Career model of specialization" by Saigal and colleagues [37]. Therefore, it is likely that even the younger physicians have received little formal training that includes the skills to build a longitudinal trusting relationship with patients using a biopsychosocial model for understanding patients, the training officially started in the mid-2000s by the Japanese Academy of Family Medicine; that was later integrated to form a new academic body, the Japan Primary Care Association, in 2010 [38]. Primary care physicians clearly have an important role in treating socially deprived individuals and providing appropriate care from both medical and psychosocial aspects. Promoting the importance of social determinants of health should be emphasized in continuous medical education for primary care physicians [39].
Co-payments affect health-care utilization, especially among low-income individuals [10,12,40]. Our results indicate that the participants considered Japan's flat rate of 30% to be sufficiently high to deter impoverished people from seeking medical care. However, changing this copayment system requires ample consideration [41]. The respondents were concerned about Japan's acutely rising health-care expenditures and believed that the government's tight budgets would not allow any reduction in the co-payment proportion [42]. To meet total health-care costs, the government faces very limited options: either increasing revenue by raising taxes, or increasing the proportion of health-care expenditure paid by insurance premiums or as co-payments [43]. Many physicians considered the zero co-payments for individuals receiving social welfare to be a cause of overutilization and a moral hazard and that the system might require some remediation. However, measures for overall cost containment should be undertaken simultaneously: Japan's fee-forservice reimbursement system for outpatient care allows primary care doctors and patients to overutilize health care [44]. As some respondents mentioned in our survey, another option would be to change the rate of copayment according to the type of disease and treatment from the currently employed flat rate. It appears necessary for the Japanese government to consider a policy that prevents inappropriate use of outpatient care while securing access for low-income patients by mitigating the financial burden to receive care. This study has several limitations. First, the qualitative data were obtained as written texts, not from indepth interviews; that could have limited the depth of the ideas expressed by the respondents. However, 365 physicians answered that question and some gave long responses; we consider that we had sufficiently rich data to understand the participants' perspectives. Second, the relatively low response rate indicates that physicians who were especially aware of their patients' economic issues were more likely to have responded. We did not know the characteristics of the nonresponders to the survey, which may affect our interpretation of the results. However, age and gender distributions of our respondents were comparable to those of Japan's national data of primary care physicians [45]. We purposefully chose survey areas that had higher proportions of socially disadvantaged individuals. The transferability of our results should be interpreted in light of the respondents' backgrounds.
Despite these limitations, the results of this study highlight primary care physicians' perspectives on the issue of co-payments as a possible deterrent for low-income patients seeking health care. Under the current policy, Japan's increasing number of socially disadvantaged people may not have equitable access to health care. Widening health disparities may threaten Japanese longevity, of which the country has long been proud. The results of this study imply that the awareness of the possibly widening inequity in health-care should be raised both among primary care physicians and policy makers.
Conclusions
Through their experience of having dealt with financially deprived patients, the majority of the primary care physicians surveyed agreed with lowering copayments for such patients. However, the remaining participants disagreed with or were uncertain about such a move, and they believed its feasibility to be an area of huge concern. It may be necessary to raise awareness of patients' socioeconomic status among primary care physicians as a possible deterrent to undergoing care. At the same time, to maintain healthcare equity, policy makers should consider balancing copayments among individuals with differing financial levels and health-care needs. | 5,889.8 | 2017-02-22T00:00:00.000 | [
"Economics",
"Medicine"
] |
Collaborative Action of Brca1 and CtIP in Elimination of Covalent Modifications from Double-Strand Breaks to Facilitate Subsequent Break Repair
Topoisomerase inhibitors such as camptothecin and etoposide are used as anti-cancer drugs and induce double-strand breaks (DSBs) in genomic DNA in cycling cells. These DSBs are often covalently bound with polypeptides at the 3′ and 5′ ends. Such modifications must be eliminated before DSB repair can take place, but it remains elusive which nucleases are involved in this process. Previous studies show that CtIP plays a critical role in the generation of 3′ single-strand overhang at “clean” DSBs, thus initiating homologous recombination (HR)–dependent DSB repair. To analyze the function of CtIP in detail, we conditionally disrupted the CtIP gene in the chicken DT40 cell line. We found that CtIP is essential for cellular proliferation as well as for the formation of 3′ single-strand overhang, similar to what is observed in DT40 cells deficient in the Mre11/Rad50/Nbs1 complex. We also generated DT40 cell line harboring CtIP with an alanine substitution at residue Ser332, which is required for interaction with BRCA1. Although the resulting CtIPS332A/−/− cells exhibited accumulation of RPA and Rad51 upon DNA damage, and were proficient in HR, they showed a marked hypersensitivity to camptothecin and etoposide in comparison with CtIP+/−/− cells. Finally, CtIPS332A/−/−BRCA1−/− and CtIP+/−/−BRCA1−/− showed similar sensitivities to these reagents. Taken together, our data indicate that, in addition to its function in HR, CtIP plays a role in cellular tolerance to topoisomerase inhibitors. We propose that the BRCA1-CtIP complex plays a role in the nuclease-mediated elimination of oligonucleotides covalently bound to polypeptides from DSBs, thereby facilitating subsequent DSB repair.
Introduction
CtIP was isolated as a binding partner of CtBP (C-terminal binding protein), and has subsequently been shown to interact with a number of molecules, including BRCA1 (Breast Cancer Susceptibility Gene 1) [1]. CtIP is a functional homolog of yeast Sae2 (Sporulation in the Absence of Spo Eleven), and acts at the initial step of homologous recombination (HR)-dependent doublestrand break (DSB) repair [2,3]. HR is initiated by forming 39 single-strand (ss) overhangs at DSBs. In this resection step, Sae2/ CtIP works together with a complex composed of Mre11/Rad50/ Xrs2 in budding yeast, or with Mre11/Rad50/Nbs1 in mammals [4][5][6][7]. The Rad51 recombinase protein polymerizes on the ss DNA overhang, and the resulting ssDNA-Rad51 complex undergoes homology search. Resection activity is upregulated by phosphorylation of a conserved residue in Sae2 by the cyclindependent kinase (CDK) [8]. This phosphorylation site is conserved in human CtIP (Thr847), and is also phosphorylated by CDK [7]. BRCA1 was originally identified as a tumor suppressor gene associated with familial breast and ovarian cancer [9]. BRCA1 contains an N-terminal RING-finger domain, and is associated with structurally related BARD1 to form an E3-ubiquitin ligase. BRCA1/BARD1 forms three distinct complexes with Abraxas, Bach1 and CtIP, and plays a role in DNA repair [10]. BRCA1 binds to CtIP in a manner that is dependent on the phosphorylation of CtIP at Ser327 [11,12]. Following DNA damage, the ubiquitylation of CtIP by BRCA1 causes the migration of CtIP towards a chromatin fraction [12]. However, the biological significance of the complex formed between BRCA1 and CtIP has not yet been clarified.
Topoisomerases 1 and 2 (Topo1 and Topo2) have been drawing increasing attention as important targets for cancer therapy, since the inhibition of these enzymes causes DSBs during DNA replication [13]. Topo1 and Topo2 induce single-strand breaks (SSBs) and DSBs, respectively. Covalent bonds are transiently formed between Topo1 and the 39 end of the SSB and between Topo2 and the 59 end of the DSB [14]. The anti-cancer agent camptothecin (CPT) inhibits Topo1 by stabilizing the Topo1cleavage complex, which interferes with replication, and thereby induces DSBs in one of the sister chromatids [15]. Topo2 inhibitors such as etoposide (VP16) and ICRF-193 also kill cycling cells and are used in cancer therapy. VP16 stabilizes the Topo2cleavage complex, while ICRF-193 stabilizes the closed clamp which forms after the strand passage [16,17]. Topo1-mediated DNA damage caused by CPT is repaired primarily by homologous recombination (HR), while Topo2-mediated DNA damage caused by VP16 or ICRF-193 is mainly repaired by nonhomologous end joining (NHEJ) [18,19]. It should be noted that the repair of CPTand VP16-induced DSBs requires an additional step: the elimination of covalently bound polypeptides from the DNA ends. Hartsuiker et al. demonstrated that Topo2 is removed from DNA by the collaborative action of the MRX complex and ctp1 (the ortholog of CtIP) in fission yeast [20]. It remains to be seen whether vertebrate CtIP shares the same function as in yeast, which does not have BRCA1 ortholog.
To understand the role of the BRCA1-CtIP interaction, we substituted the Ser332 residue (equivalent to human Ser327) of CtIP with alanine in the chicken DT40 B lymphocyte line [21,22]. In addition, to analyze the function of CtIP, we conditionally depleted CtIP in DT40 cells. We here show that the depletion of CtIP is lethal to cells as is the inactivation of Mre11, Rad50, and Nbs1 [23,24], indicating the critical role played by CtIP in HR. Remarkably, although the CtIP S332A mutation had no significant impact on HR, it made cells hypersensitive to CPT and VP16 but not to ICRF193. These observations unmasked an unexpected function of the BRCA1-CtIP interaction: cellular tolerance to the DSBs that are covalently associated with the polypeptides. Our data therefore support two distinct functions of CtIP: the resection of DSBs in HR and the elimination of polypeptides from the cleavage complex.
Results
CtIP is required for the assembly of Rad51 at DNA damage sites In order to determine the function of CtIP, we conditionally disrupted the CtIP gene in chicken DT40 cells, using a chicken CtIP transgene under the control of a tetracycline-repressible promoter (tetCtIP transgene, Figure S1A). We designed CtIP genedisruption constructs, so that the amino acid sequences from 96 to 335 would be replaced by selection-marker genes. Since the gene is encoded on chromosome 2, which is in trisomy in DT40, we disrupted three CtIP alleles ( Figure S1B and S1C). The resulting CtIP 2/2/2 tetCtIP cells tended to grow more slowly than did wildtype cells, presumably due to overexpression of the tetCtIP transgene ( Figure 1A and 1B). To deplete the CtIP in the CtIP 2/2/2 tetCtIP cells, we added doxycycline (modified tetracycline) to the culture medium. One day after the addition of doxycycline, the amount of CtIP was reduced to around 20% of wild-type cells ( Figure 1B), and the cells started dying as evidenced by an increase in the sub-G 1 fraction ( Figure 1C). This lethality can be attributed to abolished HR, because the cells showed a significant increase in the number of spontaneous chromosomal breaks (Table 1), as do Mre11and Rad51-depleted cells [23,25]. By day 3, the vast majority of the CtIP 2/2/2 tetCtIP cells had stopped growing and died ( Figure 1A and 1C). We therefore conclude that CtIP is essential for maintenance of chromosomal DNA and cellular proliferation.
To assess the HR capability of CtIP 2/2/2 tetCtIP cells, we monitored the recruitment of Rad51 and RPA to DNA damage sites one day after addition of doxycycline. Clear Rad51 foci appeared in wild-type cells one hour after ionizing radiation (IR), whereas Rad51 foci were hardly detectable in the CtIP-depleted cells ( Figure 2A). Likewise, the depletion of CtIP abolished the accumulation of RPA on DNA lesions induced by microlaser treatment ( Figure 2B). This is consistent with a phenotype shown in the previous report [3]. Thus, CtIP plays an essential role in the resection of DSBs during HR in DT40 cells as well as in mammalian cells.
We next investigated whether or not CtIP facilitates the activation of BRCA1 at DSBs. To this end, we measured the formation of conjugated-ubiquitin foci at DSBs, since Brca1 promotes extensive ubiquitylation at IR-induced DSBs [26]. Previous studies showed that BRCA1 2/2 DT40 cells exhibit a prominent defect in the formation of conjugated-ubiquitin foci [27]. In contrast, CtIP depletion did not reduce the ubiquitylation of DNA damage sites ( Figure 2C), suggesting that CtIP is not required for the activation of Brca1.
Proficient HR in CtIP S332A/2/2 DT40 clones To functionally analyze the interaction of CtIP with BRCA1, we generated CtIP S332A/2/2 cells, in which the critical amino acid in the binding interface has been mutated ( Figure S2A, S2B, S2C). The CtIP S332A/2/2 DT40 clones were capable of proliferating at a rate similar to the CtIP +/2/2 cells without a prominent change in the cell-cycle profile ( Figure 3A and Figure S2D). Western blot analysis showed that the S332A CtIP proteins were expressed at the similar level to the wild-type protein, indicating that amino acid substitutions do not affect the stability of the CtIP protein ( Figure S2E). As expected, given the results of a previous study [12], the S332A mutation of CtIP indeed inhibited its interaction with BRCA1 ( Figure S2F).
To evaluate the capability of HR in CtIP S332A/2/2 cells, we integrated an artificial substrate, SCneo, in the Ovalbumin locus [28], and measured the efficiency of I-SceI-induced gene conversion. The CtIP S332A/2/2 clones showed no significant decrease in the
Author Summary
Induction of double-strand breaks (DSBs) in chromosomal DNA effectively activates a program of cellular suicide and is widely used for chemotherapy on malignant cancer cells. Cells resist such therapies by quickly repairing the DSBs. Repair is carried out by two major DSB repair pathways, homologous recombination (HR) and nonhomologous end-joining. However, these pathways cannot join DSBs if their ends are chemically modified, as seen in the DSB ends that would arise after the prolonged treatment of the cells with topoisomerase inhibitors such as camptothecin and etoposide. These anti-cancer drugs can produce the polypeptides covalently attached to the 39 or 59 end of DSBs. It remains elusive which enzymes eliminate these chemical modifications prior to repair. We here show evidence that the BRCA1-CtIP complex plays a role in eliminating this chemical modification, thereby facilitating subsequent DSB repair. Thus, BRCA1 and CtIP have dual functions: their previously documented roles in HR and this newly identified function. This study contributes to our ability to predict the effectiveness of chemotherapeutic agents prior to their selection by evaluating the activity of individual repair factors. Accurate prediction is crucial, because chemotherapeutic agents that cause DNA damage have such strong side effects.
appearance of neomycin-resistant colonies compared to CtIP +/2/2 cells ( Figure 3B). The proficient HR in CtIP S332A/2/2 DT40 clones is in marked contrast to the severe phenotype of the Nbs1 p70 hypomorphic mutant, which exhibited a 10-fold reduction of the gene-targeting frequency and a 10 3 -fold decrease in the efficiency of HR in the SCneo substrate [29]. Next, we measured the frequency of gene targeting at the CENP-H and Ovalbumin loci. In contrast to I-SceI-induced gene conversion, the gene-targeting frequency of the CtIP S332A/2/2 clones decreased moderately in comparison with CtIP +/2/2 cells (Table 2). We speculate that this is because unknown recombination intermediates that require processing by CtIP/BRCA1 may arise during gene targeting event (see Discussion).
Fluorescent immunostaining revealed that the kinetics of Rad51 focus formation after c-irradiation was indistinguishable between CtIP S332A/2/2 cells and the CtIP +/2/2 control cells, while BRCA1 2/2 cells showed the significant reduction in the Rad51 focus formation at 1-6 h after irradiation ( Figure 4A). Furthermore, the CtIP S332A/2/2 mutants displayed laser-induced RPA accumulation as did the CtIP +/2/2 cells ( Figure 4B). Lasergenerated RPA accumulation following BrdU incorporation largely arises from the resection rather than other routes of single strand formation such as the damage caused by laser itself or replication-associated single strand formation, because RPA accumulation is abolished specifically in Ubc13 deficient cells [27]. This suggests that CtIP S332A/2/2 cells are proficient in resection at DSB sites. Taken together, we conclude that the S332A mutation of CtIP does not significantly compromise HR. Table 1. Chromosomal aberrations in CtIP 2/2/2 tetCtIP mutants.
IR 2 Gy
Wild-type 10 24 CtIP 2/2/2 tetCtIPDox(2) 6 25 CtIP S332A/2/2 cells display a marked hypersensitivity to both CPT and VP16, which stabilize the Topo-cleavage complexes To determine the role of CtIP in the cellular response to DNA damage, we measured the sensitivity of the CtIP mutant cells to various genotoxic agents using a colony survival assay. CtIP +/2/2 cells exhibited the slightly elevated sensitivity toward CPT and VP16 ( Figure 5A and 5B), though they expressed the similar level of CtIP protein to the wild-type cells ( Figure S2E). It is possible that the difference in the amount of CtIP protein between CtIP +/2/2 and wild-type cells is too subtle to detect, and that even the suboptimal level of CtIP protein renders the cells sensitive to genotoxic stimuli. A compensatory post-translational regulation may be present because CtIP +/2/2 cells exhibited about 80% reductions in CtIP mRNA level compared to the wild-type level ( Figure S2G). In contrast to CtIP +/2/2 cells, CtIP S332A/2/2 mutants showed a significantly increased sensitivity to VP16 and MMS ( Figure 5B and 5D), but not to c-rays (data not shown). Furthermore, the sensitivity to CPT was dramatically elevated in the CtIP S332A/2/2 mutants, in comparison with the CtIP +/2/2 cells ( Figure 5A).
The contribution of CtIP to the cellular tolerance to VP16 indicated that CtIP might play a role in NHEJ [18]. To test this hypothesis, we evaluated NHEJ by measuring the sensitivity of CtIP mutant cells to ICRF-193, because ICRF193-induced DNA lesions are repaired exclusively by NHEJ, whereas a fraction of the VP16-induced DSBs are repaired by HR [18]. The CtIP S332A/2/2 clones exhibited no increased ICRF193 sensitivity ( Figure 5C). NHEJ can also be evaluated by measuring the IR sensitivity of the cell population at the G 1 phase, where NHEJ plays a dominant role in DSB repair [30]. The CtIP hypomorphic mutants synchronized at the G 1 phase did not show significant IR hypersensitivity ( Figure S3). These observations indicate that Figure 2. Impaired c-ray-induced Rad51 focus formation in CtIP-depleted cells. (A) CtIP 2/2/2 tetCtIP cells, with or without 24 h doxycycline treatment, were exposed to 8 Gy ionizing radiation (IR). One hour after irradiation, cells were stained with anti-Rad51 antibody. The graph shows the quantification of Rad51 foci per cell. More than 100 cells were analyzed. (B) Accumulation of RPA protein in the cells exposed to 405 nm pulse laser. Cells were fixed with 4% PFA 1 h post laser stimulation and stained with anti-cH2AX and anti-RPA p32 antibodies. (C) Conjugated-ubiquitin (FK2) focus formation in wild-type (WT) or CtIP 2/2/2 tetCtIP DT40 cells (with or without doxycycline) at 1 h post IR (8Gy). Quantification of FK2 foci per cell is shown in the graph. More than 100 cells were analyzed. doi:10.1371/journal.pgen.1000828.g002 NHEJ is not impaired in CtIP S332A/2/2 clones. In summary, in comparison with CtIP +/2/2 cells, CtIP S332A/2/2 clones exhibited a significantly higher sensitivity to CPT and VP16, although these clones exhibited no decrease in the efficiency of HR or NHEJ. We conclude that CtIP can therefore contribute to cellular tolerance to CPT and VP16, independently of HR or NHEJ, most likely by eliminating covalently bound polypeptides from the DSBs.
We next examined the sensitivity of double mutant cells to CPT and VP16. To this end, we measured the number of live cells after 48-hour continuous exposure to the DNA-damaging agents [32], during which the double mutant cells are able to divide four to five times. We did not use a conventional colony formation assay for this purpose, because CtIP +/2/2 BRCA1 2/2 and CtIP S332A/2/2 BRCA1 2/2 clones grew very badly from a single cell in semi-solid methylcellulose medium. The number of viable cells cultured in the presence of CPT was significantly decreased for CtIP S332A/2/2 and BRCA1 2/2 cells compared to the wild-type cells, whereas CtIP +/2/2 cells grew to the similar extent to the wild-type cells in the presence of CPT ( Figure 6A). The sensitivity of CtIP +/2/2 BRCA1 2/2 cells to CPT was greater than that of BRCA1 2/2 clones. This observation is in agreement with the idea that BRCA1 and CtIP can independently contribute to HR, where CtIP promotes the resection of DSBs, while BRCA1 subsequently loads Rad51 at resected ssDNA overhang. Importantly, although the CtIP S332A mutation significantly increased cellular sensitivity to CPT in the presence of BRCA1, the CtIP S332A/2/2 BRCA1 2/2 and CtIP +/2/2 BRCA1 2/2 clones exhibited a very similar sensitivity to CPT ( Figure 6A). Likewise, the CtIP S332A/2/2 BRCA1 2/2 and CtIP +/2/2 BRCA1 2/2 clones exhibited indistinguishable cellular sensitivities to VP16 ( Figure 6B). These observations suggest that CtIP and BRCA1 can act in collaboration to repair DSBs that are chemically modified by topoisomerases.
Discussion
We here show that conditional depletion of CtIP protein led to cellular lethality with increased frequency of chromosomal aberrations in DT40 cells. CtIP depletion abolished the accumulation of RPA and Rad51 at DNA damaged sites, suggesting that it is required for the resection of DSBs during HR, and that this function is essential for the proliferation of cells. These results are in agreement with previous reports [3]. In contrast, the DT40 cells harboring S332A mutation in CtIP showed the accumulation of RPA and Rad51 upon DNA damage, and were able to proliferate with normal kinetics. Remarkably, compared to the CtIP +/2/2 cells, the CtIP S332A/2/2 clones exhibited significantly increased sensitivity to CPT and VP16, both of which stabilize the Topo-DNA cleavage complex. These observations support the proposition that, in additon to the resection of DSBs, CtIP has the second function, most likely the removal of covalently-bound polypeptides from DSBs. Hence, CtIP S332A/2/2 clones are the novel separation-of-function mutants where CtIP-dependent resection is proficient, whereas the second function required for the tolerance to topoisomerase inhibitors is deficient.
In this study, we demonstrated that the inactivation of CtIP in DT40 cells results in cellular death. We speculate that the defective DSB repair during S phase is the primary cause of cellular death rather than the misregulation of RB/E2F pathway [33,34]. It has been reported that CtIP promotes G1/S progression by releasing RB-imposed repression and by upregulating the genes required for S phase entry such as cyclin D1. MEF from CtIP-deficient mice and NIH3T3 cells transfected with CtIP siRNA arrest at G 1 phase of cell cycle. In contrast, DT40 cells that are depleted of CtIP showed a marked reduction in S phase and an increase in sub-G 1 population with the spontaneous chromosomal aberrations. We speculate that DT40 cells have a lower threshold to enter the S phase in the presence of DNA damage compared to the other types of cells owing to their character that they lack p53 expression [35] and overexpress c-myc [36]. The phenotype of our CtIP-depleted DT40 cells was remarkably different from that of the CtIP-deficient DT40 cells generated by Hiom's group [37]. Surprisingly, their CtIP-depleted DT40 cells were capable of proliferating. However, we believe that CtIP is essential for cellular proliferation because it has been shown that CtIP works together with Mre11/Rad50/Nbs1 complex in budding and fission yeasts as well as in mammalian cells [3,6,38], and the increased spontaneous chromosomal aberrations and cellular death observed in our CtIP-depleted cells are consistent with our previous reports that deficiency of either one of Mre11, Rad50, or Nbs1 was all lethal to DT40 clones [23,24]. The viability of the CtIP-deficient DT40 cells generated by Hiom's group might be due to the occurrence of suppressor mutations during the disruption of the three allelic CtIP genes. Another possibility is that the disruption of exons 1 and 2 in Hiom's group might still allow the residual expression of an N-terminal-truncated CtIP protein, as is observed for the expression of an N-terminaltruncated Nbs1 protein in patients with Nijmegen syndrome [39].
Another critically different point between our study and Hiom's group is that they conclude that the phosphorylation of CtIP-S332 promotes the resection of DSBs, whereas our data do not support this conclusion. The discrepancy between the two studies may be attributable to the different ways of introducing the S332A mutation into the DT40 cells. They randomly integrated wild-type and CtIP S332A transgenes at different loci in their ''CtIP-null'' cells, while we inserted the S332A mutant into one of the CtIP allelic genes. This knock-in approach is essential for the accurate quantitative evaluation of HR and NHEJ, because the endogenous promoter expresses CtIP transcripts differently in each phase of the cell cycle, and this differential expression accounts for the reduced usage of HR in the G 1 phase in fission yeast [38]. Alternatively, the difference between our results could be because Hiom's group reintroduced human CtIP cDNA (wild type or mutants) instead of that derived from chicken into DT40 cells to create individual clones. The human protein may act differently or incompletely in chicken DT40 cells.
The exact function of BRCA1 in HR is controversial. The discovery of the BRCA1-CtIP interaction has led to a proposal that BRCA1 might facilitate the resection step of HR [11,37,40]. However, RPA foci are not completely abolished in BRCA1 mutant cells in these reports, suggesting that ssDNA does form in the absence of functional BRCA1. We found that RPA accumulated at the sites of laser microirradiation in BRCA1 2/2 and CtIP S332A/2/2 cells, while Rad51 focus formation is impaired in BRCA1 2/2 cells. These results indicate that the BRCA1-CtIP interaction is not involved in the promotion of HR including the resection step, and are in agreement with the idea that BRCA1 facilitates the loading of Rad51 on resected ssDNA as does BRCA2 [1,29,41]. Recently, it was found that BRCA1 forms a complex with BRCA2 [42], further supporting the collaborative and overlapping function of BRCA1 and BRCA2. Although we cannot formally exclude the possibility that the RPA accumulation is delayed in BRCA1 2/2 cells (the extent of RPA accumulation induced by laser irradiation cannot be quantified, and we failed to induce RPA foci by other genotoxic stimuli in DT40 cells), our data, together with the fact that BRCA1 deficiency does not lead to cellular lethality in DT40 cells, indicate that BRCA1 has only a minor role, if any, in the resection step. The discrepancies among researchers may arise from different experimental settings including how BRCA1 is inactivated (by gene targeting, siRNA knockdown, or C-terminal truncation), the cell cycle distribution of each cell type, and the extent of DSB end modifications induced by laser or c-ray irradiation. Further studies will clarify the differences among each group.
Accumulating evidence indicates that there are two parallel pathways to eliminate chemical modifications from single-strand breaks and DSBs (Figure 7). Firstly, tyrosyl-DNA phosphodiester-ase1 (Tdp1) removes polypeptides covalently bound at the 39 end of DSBs [43]. Polynucleotide kinase 39-phosphatase (PNKP) and AP endonuclease I (APE1) are also involved in this process. Likewise, PNKP, DNA polymerase b, and aprataxin remove aberrant chemical modifications from the 59 ends of DSBs [44]. These enzymes may be capable of accurately repairing damaged bases at DSBs. On the other hand, the second pathway involves endonucleases and removes damaged bases along with proximal intact oligonucleotides from the 39 or 59 ends of DSBs. Our study showed that this pathway could contribute to cellular tolerance to alkylating agents such as MMS as well as to topoisomerase inhibitors.
A well-known precedent involving the second pathway is the Mre11/Rad50/Nbs1-complex-dependent elimination of oligonucleotides as well as the covalently associated topoisomerase-like protein (Spo11) from DSBs during meiotic HR in S. cerevisiae [2]. A more recent study of the S. pombe CtIP mutant (ctp1D) showed that the level of Top2 protein covalently bound to DNA in the ctp1D mutant increased during treatment with TOP-53, one of the VP16 derivatives, suggesting that Ctp1 plays a role in the endonucleasedependent removal of covalently-bound polypeptides from the 59 end of DSBs [20]. Our study indicates that this conclusion is also relevant to vertebrate cells although there are significant differences between vertebrate and yeast systems. First, yeast Ctp1 or Sae2 seem to be important only for the removal of the peptide covalently bound to 59 of DSB ends as demonstrated for DNA damage induced by TOP-53 or Spo11 [2,20]. Second, yeast does not have BRCA1 counterpart. BRCA1 is involved in degradation of trapped Topo1 cleavage complexes along with proteasome [45]. We hypothesize that BRCA1 may facilitate the removal of Topo1 by degrading them to small polypeptides, which in turn are removed with oligonucleotides by the nuclease activity of CtIP. In summary, we here show compelling evidence that the collaborative action of BRCA1 and CtIP plays a critical role in the endonuclease-dependent removal of damaged nucleotides from DSBs, and acts on the processed DSBs for subsequent HR and NHEJ.
Generation of CtIP conditional mutant DT40 cells
To generate CtIP gene disruption constructs, genomic DNA sequences of DT40 cells were amplified using primers 59-GGATGCGGAGAGGCTTGAAGAGTTTTACAC-39 and 59-TTACAGCACAACGATCACATAATCCCGCTC-39 for the 59 arm, and 59-GGAGCTTCTAGCAATACGCGGAACAACTCA-39 and 59-GCTTCCCCTCCAATTCTTGACTGAGAATCA-39 for the 39 arm. The amplified PCR products were cloned into the pCR2.1-TOPO vector (Invitrogen, CA, USA). The BamHI site in the plasmid that contains the 59 arm was disrupted by blunt-self ligation. The 1.6-kb HindIII fragment was ligated into the partiallydigested HindIII site of the 3.0-kb 39 arm containing the plasmid. A drug-resistance gene (hisD or bsr) was inserted into the BamHI site of the pCR2.1 vector containing both the 59 and 39 arms. To generate CtIP +/2/2 cells, linearized CtIP gene-disruption constructs were transfected sequentially by electroporation (BioRad). The genomic DNA of the transfectants was digested with SacI and the targeted clones were confirmed by Southern blot analysis. The 0.5-kb fragment was amplified using primers 59-GATTGTATGCTTCA-GAGGCTCCTGC-39 and 59-GAAATTCCCAACTTTAGCT- CtIP +/2/2 cells were introduced with the tetracycline-controlled trans-activator (tTA) gene through retrovirus infection. Infected cells were sub-cloned, and tTA expression was confirmed by western blot analysis. The resulting tTA-expressing CtIP +/2/2 cells were transfected with the pTRE2 puro R /CtIP, and puromycin-resistant clones were selected to isolate the CtIP +/2/2 tetCtIP cells. The puromycin-resistance gene was then deleted by transiently expressing the Cre recombinase (Amaxa solution T, program B-23). Puromycin-sensitive CtIP +/2/2 tetCtIP cells were transfected with the CtIP gene-disruption construct carrying the puromycin-resistant cassette to generate CtIP 2/2/2 tetCtIP cells.
Generation of CtIP S332A/2/2 mutant cells
The targeting vectors for the CtIP S332A mutants were generated by site-directed mutagenesis. To generate the S332A knock-in vector, genomic DNA was amplified by PCR, using primers 59-ATTATGCCCCTGAAAGAAGGGAAAC-39 and 59-TTTCCTGGGTTTGCTCTTGATTTT-39, and cloned into the pCR2.1-TOPO vector (Invitrogen, CA, USA). Site-directed mutagenesis was performed using primers 59-GATTCTCAGG-TAGTTGCTCCTGTTTTCGGA-39 and 59-TCCGAAAA-CAGGAGCAACTACCTGAGAATC-39. The puromycin-resistance gene was inserted into the HpaI site of the resulting plasmid. After transfection of the S332A knock-in vector into the CtIP +/+/2 cells, the targeted clones were selected against puromycin and then identified by Southern blot analysis of genomic DNA digested with HindIII. To make probe DNAs, the 0.6-kb fragments were amplified using primers 59-GACTAACAAAGATCAAC-CTGTC-39 and 59-GTGCATGAGATTTTGGTCGTTG-39. After the deletion of the puromycin-resistance gene by transiently expressing Cre recombinase by nucleofection (Amaxa, Germany), the third allele of the CtIP gene was disrupted by transfecting the CtIP gene-disruption construct carrying the puromycin-resistance gene. The insertion of the S332A mutation into the endogenous CtIP gene was confirmed by RT-PCR followed by sequencing amplified DNA.
Chromosome aberration analysis
Karyotype analysis was performed as described previously [25].
To measure the number of c-ray-induced chromosome breaks in mitotic cells, we exposed cells to 2 Gy c-rays and immediately added colcemid. At 3 hours after irradiation, mitotic cells were harvested and subjected to chromosome analysis.
Measurement of cellular sensitivity to DNA-damaging agents
Methylcellulose colony formation assays were performed as described previously [30,46]. Since in this assay the plating efficiency of BRCA1-deficient cells was less than 50%, we used a different assay to measure cellular sensitivity to DNA-damaging agents. Cells (1610 3 ) were seeded into 24-well plates containing 1 ml culture medium per well and the DNA-damaging agents, and then incubated at 39.5uC for 48 hours. To assess the number of live cells, we measured the amount of ATP in the cellular lysates. We confirmed that the number of live cells was closely correlated with the amount of ATP. This ATP assay was carried out with 96well plates using a CellTiter-Glo Luminescent Cell Viability Assay Kit (Promega Corporation, WI, USA). Briefly, we transferred 100 ml of cell suspension to the individual wells of the plates, placed the plates at room temperature for approximately 30 minutes, added 100 ml of CellTiter-Glo Reagent, and mixed the contents for 2 minutes on an orbital shaker to induce cell lysis. The plate was then incubated at room temperature for 10 minutes to stabilize the luminescent signal. Luminescence was measured by Fluoroskan Ascent FL (Thermo Fisher Scientific Inc., MA, USA).
I-Sce-I-induced gene conversion and targeted integration frequencies
The measurement of homologous recombination frequencies using a SCneo cassette [28] and CENP-H-EGFP was performed as described previously [47]. After the I-Sce-I vector was transfected into the cells, the frequency of neomycin-resistant colony formation was measured.
Synchronization of cells
To enrich DT40 cells in the G 1 phase, cells were synchronized by centrifugal counterflow elutriation (Hitachi Industrial, Japan). The cell suspension (,5610 7 ) was loaded at a flow rate of 11 ml/ min into an elutriation chamber running at 2,000 rpm. The first 50 ml was discarded, and the following 100 ml was used as a G 1phase cell fraction.
Microscopy imaging and generation of DNA damage
Fluorescence microscopy was carried out and images were obtained and processed using the IX81 (Olympus, Japan). Cells were cultured in medium containing BrdU (10 mM) for 24-48 h to sensitize them to DSB generation by means of a 405 nm laser from a confocal microscope (FV-1000, Olympus, Japan). During laser treatment, cells were incubated in phenol red-free Opti medium (GIBCO, NY, USA) to prevent the absorption of the laser's wavelength. c-irradiation was performed using 137 C (Gammacell 40, Nordion, Kanata, Ontario, Canada). Antibodies against Rad51 (Ab-1, Calbiochem, CA, USA), FK2 (Nippon Biotest Laboratories, Japan), RPA p32 (GeneTex, TX, USA), rabbit Ig (Alexa 488-conjugated antibody, Molecular Probe, OR, USA), and mouse Ig (Alexa 594-conjugated antibody, Molecular probe, OR, USA) were used for visualization. Figure S3B were introduced into the CtIP +/+/2 cells. The insertion of the S332A mutation was verified by Southern blot analysis of HindIII-digested genomic DNA. Cre recombinase were transiently expressed in the resulting CtIP S332A/+/2 clones to delete the drug-resistant marker. The remaining intact CtIP allele was targeted by the CtIP disruption construct to obtain CtIP S332A/2/2 clones. (B) The knock-in constructs for the generation of CtIP S332A/2/2 mutant clones. Black and white boxes represent exons and the drug-marker cassettes, respectively. (C) Nucleotide sequence analysis of CtIP cDNAs derived from the CtIP +/2/2 and CtIP S332A/2/2 mutant. The total RNA was subjected to reverse transcription. The regions spanning the mutations were amplified by PCR and the sequence was determined. (D) Cell cycle profiles of CtIP S332A/2/2 and CtIP F871A/2/2 mutant cells. Cells were pulse-labeled with BrdU for 10 min and subsequently stained with FITC-conjugated anti-BrdU antibody (Y axis, log scale) and propidium iodide (PI) (X axis, linear scale). (E) Western blot analysis of wild-type, CtIP +/2/2 and CtIP S332A/2/2 DT40 clones. b-actin was used as a loading control. (F) FLAG-BRCA1 association with CtIP is dependent on Ser332. 293T cells were transfected with plasmids encoding FLAG-tagged chicken BRCA1 together with either wild-type CtIP or S332A CtIP. Cell lysates were subjected to immunoprecipitation with anti-FLAG antibody, and the precipitated proteins were detected with anti-FLAG or anti-CtIP antibody. (G) Quantitative real time PCR of CtIP mRNA in wild-type, CtIP +/2/2 and CtIP S332A/2/2 DT40 clones. PCR amplification was performed in triplicate. The expression level of CtIP was normalized against b-actin using the comparative CT method. Found at: doi:10.1371/journal.pgen.1000828.s002 (19.89 MB TIF) Figure S3 Sensitivity of CtIP S332A/2/2 mutant to IR at G 1 phase. Cells at G 1 phase were separated by centrifugal elutriation and were c-irradiated for colony survival assay. The dose of c-ray irradiation is displayed on the X axis on a linear scale, while the percent fraction of surviving colonies is displayed on the Y axis on a logarithmic scale. | 7,427 | 2010-01-01T00:00:00.000 | [
"Biology",
"Medicine"
] |
Photocatalytic Reforming of Lignocelluloses, Glycerol, and Chlorella to Hydrogen Photocatalytic Reforming of Lignocelluloses, Glycerol, and Chlorella to Hydrogen
Bioethanol, biodiesel, and biogas have gained much attention as sustainable energy alternatives to petroleum‐based fuels. Bioethanol production is the most typical method to provide liquid fuel. Recently cellulosic materials have been recognized as one of the promising sources for bioethanol, since they are not directly in competition with food sources. However, ethanol concentration is usually too low to separate by distillation at a low‐energy cost. Gaseous H 2 is spontaneously isolated without operation to separate. Therefore, H 2 production is an economical approach to biofuels. Photocatalytic H 2 production over a Pt‐loaded TiO 2 is initiated by the charge separation. Electrons reduce water to generate H 2, while holes oxidize hydroxide to hydroxyl radicals. Generally, the use of sacrificial agents remarkably accelerates the H 2 production since the hydroxyl radical is consumed by them. This chapter deals with the photocatalytic H 2 production (PR) using sacrificial water‐soluble materials derived from lignocelluloses, lipids, and Chlorella. Lignocellulosic Italian ryegrass (2.00 g) was turned into H 2 (78.7 mg) through alkali treatment, hydrolysis, and PR processes. The PR process of glycerol (10.4 g) and methanol (11.3 g), which were by‐products in biodiesel synthesis, formed H 2 (3.10 g). Dried Chlorella (10 g) was turned into H 2 (578 mg) by protease hydrolysis and PR.
Introduction
Plants collect sunlight energy through photosynthesis and store it as a variety of polymeric saccharides. Polymeric saccharides are converted into monomeric saccharides, which are then converted into energy in all living organisms. Thus, saccharides are energy-storage substances which are produced from CO 2 and easily converted to energy along with CO 2 emission. Therefore, saccharides have highly potential resources to produce renewable energy. Bioethanol production from starch of maize, sugarcane, and sugar sorghum is the most typical method to provide renewable liquid fuel [1,2]. Recently in order to avoid the direct competition with food sources, cellulosic materials have been widely recognized as one of the promising sustainable resources to produce second-generation bioethanol [3]. However, the ethanol concentrations (<5.0 %) were still too low to separate by distillation at a low-energy cost [4]. On the other hand, gaseous H 2 is spontaneously isolated from reaction mixtures without operations to separate. Therefore, H 2 production from saccharides and biomass-derived materials is one of the economical approaches to biofuels [5].
In this chapter, I will show the photocatalytic reforming over titanium dioxide (TiO 2 ) using saccharides, glycerol, and amino acids, which are derived by hydrolysis of lignocelluloses, lipids, and Chlorella, respectively. This will lead to construct the sustainable energy system alternatives to petroleum-based fuels.
Outline of photocatalytic biomass reforming
A general procedure of biomass reforming is started by the production of water-soluble materials from biomass through biological treatment as well as chemical reaction. The resulting water-soluble materials are converted to biofuels such as ethanol, methane, and hydrogen through various catalytic reactions involving methane fermentation and steam reforming. It was demonstrated that the photocatalytic H 2 production from biomass-derived materials had an advantage compared with other thermal catalytic reforming by Shimura and Yoshida in their review in 2011 [6].
Our biomass reforming was performed in aqueous solution through enzymatic and chemical hydrolysis of biomass (lignocelluloses, lipids, and Chlorella) followed by photocatalytic reaction of water-soluble materials (saccharides, glycerol, and amino acids) over TiO 2 under UV-irradiation (Figure 1). Saccharides were produced by enzymatic hydrolysis of lignocelluloses using cellulase and xylanase. Glycerol was obtained by transesterification of lipid with methanol. Amino acids were obtained from hydrolysis of Chlorella by protease. These watersoluble materials were served as sacrificial agents for the photocatalytic H 2 production in aqueous solution. Details of each process were described in the following sections.
Cellulose and hemicellulose (holocellulose), which were composed of glucan and xylan, were hydrolyzed to glucose and xylose by the enzymatic saccharification (SA, Eq. 1). The powdered and pre-treated lignocellulose (4.0 g) was dispersed in an acetate buffer solution (80 mL, pH 5.0, 0.1 M) which was prepared by mixing 0.808 g acetic acid and 3.05 g sodium acetate in 500 mL of water. Cellulase (200 mg) and xylanase (200 mg) were added to the suspension of lignocellulose. The SA was performed by stirring the solution vigorously with a magnetic stirrer at 45°C for 120 h. After centrifugation of reaction mixture, the supernatant solution involving glucose and xylose was analyzed by HPLC and used as sacrificial agents in the following photocatalytic reaction.
Cellulase Xylanase Xylan Glucan
Xylose Glucose water Also, lignocellulose could be turned into ethanol and xylose through simultaneous saccharification and fermentation (SSF, Eq. 2) using cellulase and xylanase as well as S. cerevisiae as follows [8]. An acetate buffer solution (10 mL, pH 5.0, 0.1 M) was added to pre-treated lignocelluloses (3.0 g) in the reaction vessel. The reaction vessel was autoclaved at 120°C for 20 min. After cooling, cellulase (180 mg) and xylanase (120 mg) in an acetate buffer solution (8.0 mL) and the cell suspension of S. cerevisiae (0.36 mL) were introduced into the reaction vessel. After air was purged with N 2 stream for 15 min, the SSF was performed at 34°C under stirring vigorously with a magnetic stirrer. The evolved CO 2 was collected by a measuring cylinder to monitor the volume of CO 2 gas. The SSF reaction was continued for about 96 h until CO 2 evolution ceased. After unreacted biomass was removed from the reaction mixture by centrifugation, the supernatant solution was analyzed by gas chromatography (GC) and highperformance liquid chromatography (HPLC) to determine the concentrations of ethanol and saccharides, respectively. Ethanol was collected from the SSF solution by evaporation under reduced pressure while the residual xylose was subjected to the photocatalytic reaction. Cellulase, Xylanase X OH ylan Glucan Xylose C .
S cerevisiae water
Another process to convert lignocellulose to ethanol is simultaneous saccharification and cofermentation (SSCF). A recombinant Escherichia coli KO11 which can ferment xylose was used. Glucan and xylan in lignocellulose are turned to ethanol by SSCF using cellulase, xylanase, yeast, and E. coli KO11. An example is an SSCF process of the low-moisture anhydrous ammonia (LMAA)-treated Italian ryegrass (Section 6.1), which produced ethanol in 84.6% yield [9]. In this case, it was not necessary to undergo the photocatalytic process.
, . , Cellulase Xylanase Xylan Glucan CO C H OH KO water
Titanium dioxide (TiO 2 ) as the photocatalyst
TiO 2 is a white powder material which is thermally stable, non-flammable, and no health hazards. Therefore, TiO 2 has been used for many years in industrial and consumer goods, including paints, coated fabrics and textiles, cosmetics, and so on. The photocatalytic H 2 production was performed by use of an anatase-type TiO 2 . It has a semi-conductor structure whose band gap is known to be 3.20 eV, which corresponds to 385 nm. Therefore, TiO 2 can be excited by 366 nm-emission from a high-pressure mercury lamp. Irradiation induces charge separation into electrons and holes on the TiO 2 [10]. Electrons (e -) reduce water to generate H 2, while holes (h + ) oxidize hydroxide anions to hydroxyl radicals ( Figure 2) [11]. In most cases, noble metals (Pt, Pd, and Au) were loaded on TiO 2 to accelerate the reduction of water by electrons. We used a Pt-loaded TiO 2 (Pt/TiO 2 ) throughout the present investigation. Moreover, it was well known that the use of sacrificial agents remarkably accelerates H 2 production because the hydroxyl radical is consumed by them. Especially, we have elucidated that sacrificial agents with all of the carbon attached heteroatoms (O and N) are superior sacrificial agents because they continued to serve as electron sources until their sacrificial ability was exhausted [12,13]. Glucose, xylose, glycerol, and glycine meet this requirement.
The photocatalytic H 2 production using sacrificial agents is called "sacrificial H 2 production."
Preparation of Pt-loaded TiO 2 photocatalyst
For photocatalytic reaction, almost researches have continued to use a P25 (Degussa Co. Ltd, Germany) and a ST01 (Ishihara Sangyo Co. Ltd., Japan). The P25 is prepared through hydrolysis of TiCl 4 and composed of 75% of anatase and 25% of rutile, while the ST01 was prepared through hydrolysis of TiOSO 4 and composed of 100% of anatase.
The Pt-loaded TiO 2 (Pt/TiO 2 ) was prepared by the method reported by Kennedy and Datye [14] as follows. An aqueous solution (400 mL) containing TiO 2 (4.0 g, ST01, particle size 7 nm and surface area 300 m 2 g -1 ), K 2 PtCl 6 (200 mg), and 2-propanol (3.06 mL) was introduced into a reaction vessel which is illustrated in Section 4.3. After O 2 was purged by N 2 gas, the solution was irradiated by a high-pressure mercury lamp with stirring for 24 h when the gas evolution reached over 100 mL. After the irradiation, water was entirely evaporated. The resulting gray precipitate was moved on a filter and washed with water and then dried and ground to produce Pt/TiO 2 powder. The Pt-content on TiO 2 was optimized to be 2.0 wt% from the photocatalytic H 2 evolution by various Pt-content TiO 2 using glucose as a sacrificial reagent. Identification of Pt/TiO 2 was usually performed by an XRD pattern and TEM image [15]. Figure 3 shows a TEM image and an X-ray diffraction pattern of a P/TiO 2 (2.0 wt% of Pt content).
Experimental method
The photocatalytic H 2 production was performed using a photo-irradiation apparatus (Figure 4). The catalyst (100 mg) and the given amounts of aqueous solution of sacrificial agent were introduced into a reaction vessel. The volume of the reaction solution was adjusted to 150 mL with water. The reaction vessel was connected with a measuring cylinder through a gas-impermeable fluororubber tube to collect the evolved gas. A high-pressure mercury lamp (100 W, UVL-100HA, Riko, Japan) was inserted into the reaction vessel, which was set in a water bath to keep it at a constant temperature (usually 20°C). After O 2 was purged from the reaction vessel by N 2 gas for 15 min, the reaction mixture was irradiated with vigorous stirring using a magnetic stirrer until the gas evolution ceased. The evolved gas was collected by a measuring cylinder to measure the total volume of the evolved gas. The evolved gas (0.5 mL) was obtained using a syringe and was subjected to the quantitative analysis of H 2 , N 2 , and CO 2 , which were performed on a Shimadzu GC-8A equipped with a TCD detector at a temperature raised from 40 to 180°C using a stainless column (3 mmΦ, 6 m) packed with a SHINCARBON ST (Shimadzu). In the absence of sacrificial agents, the H 2 evolution from water was small (<2 mL).
Analysis of photocatalytic reaction
Theoretically, the photocatalytic reaction can convert glucose and xylose to 12 and 10 equivalents of H 2 (Eq. 4). Indeed, the photocatalytic reaction using glucose and xylose produced 11.8 and 10.0 mol of H 2 from 1 mol of glucose [15] and xylose [16], respectively. I show a method to determine the amounts of H 2 evolved from 1 mol of sacrificial agent. A typical example is the photocatalytic H 2 production using saccharides obtained from enzymatic saccharification of Napier grass. Although the saccharides contained not only xylose but also glucose, the evolved H 2 and CO 2 were plotted against the moles of xylose in a mixture of xylose and glucose, as shown in Figure 5A. Gas volumes of H 2 and CO 2 increased with the increase of xylose. However, the molar ratios of H 2 to xylose (H 2 /xylose) were not constant to the amount of xylose used. It was speculated that the colored material in the solution and the carboxylic acids formed during the photocatalytic reaction may lower the activity of photocatalyst. Therefore, the H 2 /xylose ratio was plotted against the molar ratio of xylose to catalyst (xylose/catalyst), as shown in Figure 5B. As the xylose/catalyst ratios decreased, the H 2 /xylose ratios increased. The intercept of the plots was equaled to H 2 max , which is the limiting mole amount of H 2 produced from one mole of xylose (sacrificial agent) at an infinite amount of catalyst [17]. Thus, the total molar amount of H 2 was calculated by the equation: H 2 max × (moles of sacrificial agent).
Energy recovery efficiency (E ff )
Total energy recovery efficiency (E ff ) from biomass to biofuels was calculated using combustion energy: E ff = 100H F /H 0 where H 0 and H F were the combustion energies of biomass and biofuels, respectively. The combustion energies of sacrificial agents such as glucose, xylose, and glycerol are 2803 [18], 2342 [19], and 1654 kJ/mol [18], respectively. The combustion energies of biofuels such as ethanol and H 2 are 285 and 1367 kJ/mol [18], respectively. In the case of lignocellulose, the H 0 value was combustion energy of glucose and xylose at the complete hydrolysis of glucan and xylan which were determined by the National Renewable Energy Laboratory (NREL) [20].
Lignocelluloses
Lignocellulosic biomass was composed of cellulose, hemicellulose, lignin, and other components. The components of glucan, xylane, lignin, ash, and others in non-treated lignocelluloses are summarized in Table 1. Since the contents of cellulosic components in lignocelluloses were in the range of 41.0-66.5 wt%, only a half of lignocelluloses were utilized for production of H 2 . The method to determine the content of each component was shown as follows. Lignocelluloses were cut by a cutter and dried at 70°C for 72 h. The dried matter was powdered by a blender until the powder passed through a sieve with 150 μm of mesh. The powdered lignocellulose (30 g) was treated with a 1% aqueous solution of NaOH (400 mL) at 95°C for 1 h. The reaction mixture was centrifuged and filtered to isolate the holocellulose (a mixture of cellulose and hemicellulose) as a pale yellow precipitate. The supernatant solution was made acidic (pH 5.0) with a dilute HCl solution to isolate dark brown precipitate which was identified as lignin. The precipitate was collected via centrifugation at 10,000 rpm for 10 min.
The contents of saccharides in holocellulose were analyzed according to the methods published by NREL [20]. Sulfuric acid (72 wt%, 3.0 mL) was added slowly to holocellulose (300 mg) in a reaction vessel and kept at 30°C for 1 h. Water (84 mL) was added to the reaction vessel so that the concentration of sulfuric acid became 4.0 wt%. Acid hydrolysis was performed by autoclaving at 121°C for 1 h in an autoclave. The treated solution was neutralized with CaCO 3 and was centrifuged. The supernatant solution (ca. 87 mL) was concentrated to 30 mL by evaporation. The solution was analyzed by HPLC to determine the amounts of glucose and xylose. The amounts of glucan and xylan were determined from the amounts of glucose and xylose. The ash component in lignocellulose was obtained by the burning of the lignocellulose (2.0 g) in an electric furnace (KBF784N1, Koyo, Nara, Japan) for 2 h at 850°C.
The pre-treatments to promote an enzymatic digestibility of the cellulosic components and to remove the lignin component were usually performed. Alkali (AL) treatment is a popular method to remove lignin from lignocelluloses [21]. A powdered lignocellulose (30 g) was added to a 1% aqueous solution of NaOH (400 mL). The mixture was heated under stirring at 95°C for 1 h. The reaction mixture was subjected to centrifugation at 10,000 rpm for 10 min. The lignin remained in the supernatant solution. The holocellulose, which is a mixture of cellulose and hemicellulose, is isolated as a pale yellow precipitate, which was washed by dispersion in water to remove the contaminated lignin. After the pH adjustment to 7.0, the washed precipitate was collected by centrifugation and dried. Thus, lignin-removed holocellulose was obtained. The AL treatment is effective for saccharification of the lignocellulose with higher lignin contents. However, in the case of lignocelluloses with low lignin content such as Napier grass, the AL treatment retarded the yeast-fermentation rate because AL treatment removed not only lignin but also nutrients to help yeast fermentation [22].
Another useful pretreatment of lignocelluloses is LMAA (low-moisture anhydrous ammonia pretreatment), described as follows [23]. Dry powdered lignocelluloses (100 g, volume 320 mL) were mixed homogeneously with water (100 g) in a flask (1 L). The flask containing wet lignocellulose was evacuated with a pump and then gaseous NH 3 was introduced into the flask repeatedly until the atmosphere inside the flask was entirely replaced with NH 3 gas. The moist powdered lignocellulose was kept under an NH 3 gas atmosphere at room temperature for 28 days. After NH 3 was removed with an evaporator, the treated lignocellulose was washed with water to liberate the brownish aqueous alkali solution of the lignin. This washing operation was continued until the pH became below 7.7. The treated lignocellulose was dried at 60°C. Here, NH 3 served for transformation of the cellulose crystal phase to a highly reactive structure toward enzymatic degradation rather than the removal of lignin [24]. As a special pretreatment method, TiO 2 -photocatalytic pretreatment was developed by our group [25].
The photocatalytic reforming was applied to lignocelluloses such as Italian ryegrass [26], Napier grass [26], bamboo [27], rice straw [27], and silver grass [27]. The results are summarized in Table 2. The SA→PR method is a process through the enzymatic saccharification (SA) of the pretreated lignocelluloses into glucose and xylose which were then used as sacrificial agents for the photocatalytic H 2 production over Pt/TiO 2 (PR). For example, the dried Italian ryegrass (2.00 g) was subjected to the AL treatment to give the AL-treated Italian ryegrass (1.00 g) which was turned into 554 mg of glucose and 193 mg of xylose by SA. The SA of xylan was more inefficient than that of glucan. Glucose and xylose were turned into H 2 (78.7 mg) by PR. As a result, the total energy recovery efficiency (E ff ) from AL-treated Italian ryegrass to H 2 was calculated to be 71.9% (Figure 6). In the case of Napier grass, dried Napier grass (2.075 g) was subjected to the AL treatment to give the AL-treated Napier grass (1.00g) which was turned into 487 mg of glucose and 197 mg of xylose by SA. The PR of glucose and xylose gave 84.0 mg of H 2, which corresponded to 77.0% of E ff . In the case of the SSF→PR method, the LMAA treatment of the dried Italian ryegrass (1.458 g) gave the LMAA-treated Italian ryegrass (1.0 g) which was turned into ethanol (250 mg), xylose (121 mg), and glucose (19 mg) by SSF process. Ethanol was removed from SSF solution, whereas the residual xylose and glucose were converted to H 2 (17.3 mg) by PR. The E ff value of H 2 combined with ethanol was 82.7% from the LMAA-treated Italian ryegrass. We have reported the ethanol production through an SSCF process of Italian ryegrass [9]. The E ff value was 82.7%. These E ff values showed similar values. In the cases of Napier grass, the LMAA treatment of the dried Napier grass (1.637 g) gave the LMAA-treated Napier grass (1.0 g) which was turned into ethanol (177 mg), xylose (167 mg), and glucose (13 mg) by SSF process. After ethanol was removed from SSF solution, the residual xylose and glucose were converted to H 2 (21.0 mg) by PR. The E ff value of H 2 combined with ethanol was 77.2% from the LMAAtreated Napier grass. In the cases of bamboo, rice straw, and silver grass, the AL treatment of bamboo (1.656 g), rice straw (2.092), and silver grass (2.439 g) produced the AL-treated lignocelluloses (1.00 g). They were turned into ethanol and H 2 by the SSF→PR process with E ff of over 73.4%.
Glycerol
Biodiesel (BDF) is one of new sustainable energy alternatives to petroleum-based fuels. BDF market has significantly increased in Europe to adhere energy and climate policies [28]. BDF (methyl alkanoate) is produced by transesterification of vegetable oil or animal fats with methanol under basic conditions [29]. However, glycerol as co-production and unreacted methanol was not utilized and went to waste. Glycerol has a potential to produce H 2 in maximum theoretical yield of seven equivalents (Eq. 5). Also methanol can produce three equivalents of H 2 . Hydrogen transformation of glycerol and unreacted methanol isolated from the BDF synthesis was performed by sacrificial H 2 production over a Pt/TiO 2 [30].
As starting material, we used vegetable oil which was mainly composed of oleic acid (C 17 H 33 CO 2 H) triglyceride. The average molecular weight of vegetable oil was thought to be 884 g/mol. Vegetable oil (150 mL, 136.5 g, 0.154 mol) was set in a reaction vessel. Methanol (30 mL, 23.8 g, 0.743 mol) was mixed with NaOH (0.485g, 0.012 mol). About half of the mixture of methanol and NaOH was poured into a reaction vessel and then kept at 61°C for 1 h. Moreover, the remaining mixture of methanol and NaOH was added into the reaction vessel and the reaction mixture was kept at 61°C for another 1 h. After cooling, the reaction mixtures were separated into a lower layer and an upper layer. The procedure of the follow-up process is shown in Figure 7. The lower layer (GL layer) contained glycerol (GL, 0.113 mol) and methanol (0.214 mol). The upper layer (BDF layer) was washed with water (300 mL) to give BDF (114.5 g, 0.387 mol) and the aqueous washing solution which contained 0.137 mol of methanol. The total recovery yield of unreacted methanol was 47.5%. The yields of GL and BDF were 73.3 and 83.7%, respectively. The photocatalytic reaction was performed by irradiation of aqueous solution containing Pt/TiO 2 powder (100 mg, 1.25 mmol) and GL layer, which was added to the reaction vessel so that the amounts of GL became 0.25, 0.50, 0.75, 1.00, and 1.25 mmol. The limiting mole amount of H 2 (H 2 max ) per 1 mol of GL was obtained from the plots of the H 2 /GL against the GL/catalyst. Similarly the photocatalytic reaction was performed for the washing solution, which contained methanol. Using H 2 max values, it was calculated that 2.82 and 0.28 g of H 2 was obtained from the GL layer and washing solution, respectively. The E ff value of H 2 was determined to be 100.8% using H F of H 2 (444 kJ) and the sum of combustion energy of glycerol (H 0 = 187 kJ) and unreacted methanol (H 0 = 255 kJ).
We examined the photocatalytic H 2 production from Chlorella [32]. The frozen Chlorella was thawed and dried in a drying machine and then ground. Gas evolution did not occur from the non-enzymatic-treated solution, which was prepared by magnetic stirring of the Chlorella powder (10 g) in a phosphate buffer (60 mL) for 48 h at 50°C. Therefore, the enzymatic hydrolysis of Chlorella powder (10 g) was performed using protease (1.0 g) in a phosphate buffer (0.1 M, pH 7.6, 60 mL) under stirring at 50°C for 48 h to give the enzymatic hydrolyzed solution. The solution was subjected to centrifugation to remove the precipitate. The supernatant solution (EH solution) was collected. The EH solution was subjected to freezing-drying in order to weigh the water-soluble components in the EH solution. It was determined to be 117 g/L. Since the weight of the solid was 167 g/L before hydrolysis, more than 70% of the solid was hydrolyzed into water-soluble components. The EH solution was composed of 98.0 g/L of amino acids and 18.3 g/L of glucose which were determined by colorimetric analysis using ninhydrin and by HPLC analysis, respectively.
The photocatalytic H 2 production was performed using the EH solution (0.10 -0.50 mL) over a Pt/TiO 2 (100 mg) in 150 mL of water. The limiting volume of H 2 per 1 mL of the EH solution (H 2 max ) was determined to be 119 mL/mL from the plots of the H 2 /(EH solution) against the (EH solution)/catalyst. We successfully produced 579 mg of H 2 from 10.0 g of dry Chlorella (Figure 8). This yield is higher than 394 mg for the H 2 production through AL treatment, saccharification, and photocatalytic H 2 production from non-treated Italian ryegrass (10.0 g) [25,27]. Thus, the photocatalytic reforming is applicable to not only saccharides but also amino acids. Chlorella includes colored materials such as chlorophyll which may disturb the light absorption by the catalyst. Therefore, dried Chlorella (20 g) was subjected to refluxing in ethanol (100 mL) for 6 h to remove the colored materials. Almost all amount of colored materials remained in the ethanol solution. However, the decolorization did not affect the amount of H 2 but could shorten the irradiation time.
In this chapter, biohydrogen production was discussed from the viewpoints of feedstock and methodology to transform biomass to fuels. This will help life recycle assessment (LCA) to evaluate CO 2 emission during cultivation, transportation, and manufacturing, as performed for bioethanol from cellulose [33]. | 5,856.2 | 2017-01-25T00:00:00.000 | [
"Engineering"
] |
Direct growth inhibition assay of total airborne fungi with application of biocide-treated malt extract agar
Graphical abstract
G R A P H I C A L A B S T R A C T A B S T R A C T
Indoor air pollution by airborne fungi has risen to become a common issue all over the world and it is hazardous to indoor occupants' health as it is associated with a series of respiratory-related and skin-related diseases. Selected bioactive compounds from the food industry have been suggested to be effective against individual fungus isolated from indoor environment. However, the techniques used to evaluate these compounds were lengthy and unsuitable against total airborne fungi. Therefore, this paper describes an assay to assess the effectiveness of a bioactive compound to inhibit growth of total airborne fungi.
A combination and modification of previous methods and the NIOSH Manual Analytical Standard Method (NMAM 0800) is proposed. This method concurrently samples the total airborne fungi and evaluates the ability of bioactive compounds (potassium sorbate in this paper), as a biocide, to treat these indoor airborne fungi.
Study background
Indoor airborne fungi contamination has became a serious issue in indoor air quality (IAQ) management as it is correlated with various diseases such as damage of the respiratory tract that involves nose and lung, skin infection, mucous membrane irritation and a series of symptoms classified under the sick building syndrome [1]. The effect of conventional fungicides used in disinfecting the environment is not long-lasting [2] and might be toxic to humans. Besides, microorganisms have developed resistance against existing fungicides [3]. As the conventional fungicides are not suitable for indoor usage, more environmental friendly compounds that are nontoxigenic to humans are required [4]. A few biocides used in the food industry have been evaluated and shown to be effective against isolated indoor waterborne fungi [4][5][6][7][8][9] and single isolated indoor airborne fungus [9]. Nevertheless, the techniques used in the previous method focused on evaluation of bioactive compounds against single isolated fungus by taking into consideration that the fungus growth is measured by diameter of fungus colony, which is very hard to measure, and required 30 days to accomplish [4]. While, the previous NMAM 0800 method is meant for bioaerosol sampling only [10]. Thus, these previous methods were not suitable for evaluation of growth inhibition of total airborne fungi. Hence, a method to evaluate the performance of bioactive compounds (potassium sorbate in this study), as a biocide [11,12] in growth inhibition of total airborne fungi was reported here.
The biocide's antifungal activity was assessed by a direct growth inhibition assay of total airborne fungi that comprises the air samplings of total airborne fungi, incubation and enumeration of fungal colonies formed. Biocide-incorporated and untreated control MEA were used in these procedures. The assay takes into account that the total number of the viable fungi can be indicated by colony forming unit (CFU) analysis [13].
Preparation of potassium sorbate-incorporated malt extract agar
Firstly, 0.03% (w/v) of potassium sorbate was incorporated into malt extract agar (MEA). The mixtures were sterilized in an autoclave at 121 8C for 15 min. Pre-sterilized Petri dishes measuring 90 mm  15 mm were filled with 20 mL of the biocide-treated MEA under aseptic conditions. The solidified biocide-treated MEA plates were sealed with Parafilm. The control MEA plates without potassium sorbate were prepared under the same conditions. The whole process was carried out in a laminar flow hood. fungi onto the MEA plates was carried out for 5 min for each sample. Before the first air sampling and between every two consecutive measurements, the air sampler was cleaned with 70% ethanol to avoid contamination. Air sampling was carried out in triplicate onto the biocide-incorporated MEA and untreated control MEA, respectively. After the field sampling for 5 min, the plates were removed from the sampler, sealed with Parafilm and immediately placed in a cooler box with an ice pack at 4 8C to inhibit microbial growth. The air samples were then transferred to laboratory aseptic conditions within 2 h. All air samplings at a particular testing area were performed on the same day. The airborne fungi samples were cultured at 37 8C for 5 days.
Viable counts of total airborne fungi
The enumeration of the samples is indicated by colony forming unit (CFU) analysis [13]. The counting process was done by mounting the agar plate on digital colony counter and the colonies were counted manually (Fig. 1). The total number of the fungi colonies formed on the agar plate was then divided with the total volume of air drew by the sampler. The calculation is as follows [14]: The fungal colonies formed were observed and counted daily until the fifth day to identify the colonies, to ensure no growth of bacterial colonies and to solve the difficulties in recognizing and counting the colonies. Since replicate samples were collected, the data was averaged.
Biocide's antifungal activity toward total airborne fungi
The biocide inhibitive performance was determined by calculating the percentage of the reduction of the average total counts of the viable airborne fungi found on both types of agar plate, as shown in the equation below: where, X is the average total counts of airborne fungi found on the control MEA, and Y is the average total counts of airborne fungi found on the biocide-treated MEA. It was shown that, with this method, the total airborne fungi can grow on both types of MEA but with different total fungi counts. Moreover, consistent total counts of fungi can be found on the triplicates of sampling using the same type of MEA. Therefore, with this tabulation, the ability or effectiveness of a biocide (potassium sorbate in this study) against general microenvironment of the total airborne fungi at the testing site can be determined. This determination is important as it [ ( F i g . _ 1 ) T D $ F I G ] provides a new eco-friendly alternative to circumvent indoor air pollution by airborne fungi and therefore to provide a safe and comfortable indoor environment with good indoor air quality. In this study, potassium sorbate was shown to effectively reduce the total counts of airborne fungi with around 84% of biocide inhibitive performance (Table 1).
Additional information and recommendations
The incubation temperature of 37 8C was used in this study to selectively sample fungi that are pathogenic to humans [15]. However, a more common and lower incubation temperature, such as 25 8C could be used for general purposes. A 5 days incubation period was used according to the standard method, NMAM 0800 [10] and previous indoor airborne fungal sampling studies [16][17][18]. It is a standard incubation period for fungi samples in indoor air quality studies.
Potassium sorbate was used as an example of the subject of the assay in this study because of its previous performance in controlling the growth of individual indoor fungus [4][5][6][7][8][9]. However, this method could be used to assess the biocide inhibitive performance of other new alternatives/bioactive compounds against the total indoor airborne fungi. Table 1 The biocide inhibitive performance of the biocide, potassium sorbate against total indoor airborne fungi. | 1,661.8 | 2015-07-20T00:00:00.000 | [
"Biology",
"Environmental Science"
] |
Two-Dimensional Audio Compression Method Using Video Coding Schemes
: As video compression is one of the core technologies that enables seamless media streaming within the available network bandwidth, it is crucial to employ media codecs to support powerful coding performance and higher visual quality. Versatile Video Coding (VVC) is the latest video coding standard developed by the Joint Video Experts Team (JVET) that can compress original data hundreds of times in the image or video; the latest audio coding standard, Unified Speech and Audio Coding (USAC), achieves a compression rate of about 20 times for audio or speech data. In this paper, we propose a pre-processing method to generate a two-dimensional (2D) audio signal as an input of a VVC encoder, and investigate the applicability to 2D audio compression using the video coding scheme. To evaluate the coding performance, we measure both signal-to-noise ratio (SNR) and bits per sample (bps). The experimental result shows the possibility of researching 2D audio encoding using video coding schemes.
Introduction
As consumer demands for realistic and rich media services on low-end devices are rapidly increasing in the field of multimedia delivery and storage applications, the need for audio or video codecs with powerful coding performance is emphasized, which can achieve minimum bitrates and maintain higher perceptual quality compared with the original data. As the state-of-the-art video coding standard, Versatile Video Coding (VVC) [1] was developed by the Joint Video Experts Team (JVET) of the ITU-T Video Coding Experts Group (VCEG) and ISO/IEC Moving Picture Experts Group (MPEG). VVC can achieve a bitrate reduction of 50% with similar visual quality compared with the previous method, named High Efficiency Video Coding (HEVC) [2]. In the field of audio coding technology, Unified Speech and Audio Coding (USAC) [3] was developed by the MPEG audio group as the latest audio coding standard, which integrates speech and audio coding schemes. Whereas VVC can accomplish significant coding performance in the original image or video data, USAC can provide about 20 times compression performance of the original audio or speech data.
In this study, we conducted various experiments to verify the possibility of compressing two-dimensional (2D) audio signals using video coding schemes. In detail, we converted 1D audio signal into 2D audio signal with the proposed 2D audio conversion process as an input of a VVC encoder. We used both signal-to-noise ratio (SNR) and bits per sample (bps) for performance evaluations.
The remainder of this paper is organized as follows: In Section 2, we overview the video coding tools, which are newly adopted in VVC. In Section 3, we describe the proposed
Overview of VVC
VVC can provide powerful coding performance compared with HEVC. One of the main differences between HEVC and VVC is the block structure. Both HEVC and VVC commonly specify coding tree unit (CTU) as the largest coding unit, which has a changeable size depending on the encoder configuration. In addition, a CTU can be split into four coding units (CUs) by a quad tree (QT) structure to adapt to a variety of block properties. In HEVC, a CU can be further partitioned into one, two, or four prediction units (PUs) according to the PU splitting type. After obtaining the residual block derived from the PU-level intra-or inter-prediction, a CU can be partitioned into multiple transform units (TUs) according to a residual quad-tree (RQT) structure similar to that of CU split.
VVC substitutes the concepts of multiple partition unit types (CU, PU, and TU) with a QT-based multi-type tree (QTMTT) block structure, where the MTT is classified into binary tree (BT) split and ternary tree (TT) split to support more flexibility for CU partition shapes. This means that a QT can be further split by the MTT structure after a CTU is first partitioned by a QT structure. As depicted in Figure 1, VVC specifies four MTT split types: vertical binary split (SPLIT_BT_VER), horizontal binary split (SPLIT_BT_HOR), vertical ternary split (SPLIT_TT_VER), and horizontal ternary split (SPLIT_TT_HOR), as well as QT split. In VVC, a QT or MTT node is considered a CU for prediction and transform processes without any further partitioning schemes. Note that CU, PU, and TU have the same block size in the VVC block structure. In other words, a CU in VVC can have either a square or rectangular shape, whereas a CU in HEVC always has a square shape. The remainder of this paper is organized as follows: In Section 2, we overview the video coding tools, which are newly adopted in VVC. In Section 3, we describe the proposed methods for converting a 1D audio signal into a 2D audio signal. Finally, the experimental results and conclusions are provided in Sections 4 and 5, respectively.
Overview of VVC
VVC can provide powerful coding performance compared with HEVC. One of the main differences between HEVC and VVC is the block structure. Both HEVC and VVC commonly specify coding tree unit (CTU) as the largest coding unit, which has a changeable size depending on the encoder configuration. In addition, a CTU can be split into four coding units (CUs) by a quad tree (QT) structure to adapt to a variety of block properties. In HEVC, a CU can be further partitioned into one, two, or four prediction units (PUs) according to the PU splitting type. After obtaining the residual block derived from the PU-level intra-or inter-prediction, a CU can be partitioned into multiple transform units (TUs) according to a residual quad-tree (RQT) structure similar to that of CU split.
VVC substitutes the concepts of multiple partition unit types (CU, PU, and TU) with a QT-based multi-type tree (QTMTT) block structure, where the MTT is classified into binary tree (BT) split and ternary tree (TT) split to support more flexibility for CU partition shapes. This means that a QT can be further split by the MTT structure after a CTU is first partitioned by a QT structure. As depicted in Figure 1, VVC specifies four MTT split types: vertical binary split (SPLIT_BT_VER), horizontal binary split (SPLIT_BT_HOR), vertical ternary split (SPLIT_TT_VER), and horizontal ternary split (SPLIT_TT_HOR), as well as QT split. In VVC, a QT or MTT node is considered a CU for prediction and transform processes without any further partitioning schemes. Note that CU, PU, and TU have the same block size in the VVC block structure. In other words, a CU in VVC can have either a square or rectangular shape, whereas a CU in HEVC always has a square shape. Table 1 shows newly adopted coding tools between HEVC and VVC. In general, intra-prediction generates a predicted block from the reconstructed neighboring pixels of the current block. As shown in Figure 2, VVC can provide 67 intra-prediction modes, where modes 0 and 1 are planar and DC mode, respectively, and the others are in angular prediction mode to represent edge direction. According to [4], the VVC test model (VTM) [5] achieves 25% higher compression performance than the HEVC test model (HM) [6] under the all-intra (AI) configuration recommended by JVET Common Test Conditions (CTC) [7]. This improvement was mainly realized by the newly adopted coding tools, such as position dependent intra-prediction combination (PDPC) [8], cross component linear model intra-prediction (CCLM) [9], wide angle intra-prediction (WAIP) [10], and matrix- Table 1 shows newly adopted coding tools between HEVC and VVC. In general, intra-prediction generates a predicted block from the reconstructed neighboring pixels of the current block. As shown in Figure 2, VVC can provide 67 intra-prediction modes, where modes 0 and 1 are planar and DC mode, respectively, and the others are in angular prediction mode to represent edge direction. According to [4], the VVC test model (VTM) [5] achieves 25% higher compression performance than the HEVC test model (HM) [6] under the all-intra (AI) configuration recommended by JVET Common Test Conditions (CTC) [7]. This improvement was mainly realized by the newly adopted coding tools, such as position dependent intra-prediction combination (PDPC) [8], cross component linear model intraprediction (CCLM) [9], wide angle intra-prediction (WAIP) [10], and matrix-based intraprediction (MIP) [11]; the computational complexity of encoder substantially increased approximately 26 times [4]. Inter-prediction fetches a predicted block from previously decoded reference frames using motion estimation (ME) and motion compensation (MC) processes. The motion parameter can be transmitted to the decoder in either an explicit or implicit manner. When a CU is coded as SKIP or MERGE mode, the encoder does not transmit any coding parameters, such as significant residual coefficients, motion vector differences, or a reference picture index. In cases of SKIP and MERGE modes, the motion information fo the current CU is derived from the neighboring CUs including spatial and tempora candidates. VVC adopted several new coding tools for SKIP/MERGE and inter-prediction Inter-prediction fetches a predicted block from previously decoded reference frames using motion estimation (ME) and motion compensation (MC) processes. The motion parameter can be transmitted to the decoder in either an explicit or implicit manner. When a CU is coded as SKIP or MERGE mode, the encoder does not transmit any coding parameters, such as significant residual coefficients, motion vector differences, or a reference picture index. In cases of SKIP and MERGE modes, the motion information for the current CU is derived from the neighboring CUs including spatial and temporal candidates. VVC adopted several new coding tools for SKIP/MERGE and inter-prediction, such as affine inter prediction [12], geometric partitioning mode (GPM) [13], merge with motion vector difference (MMVD) [14], decoder-side motion vector refinement (DMVR) [15], combined inter-and intra-prediction (CIIP) [16], and bi-prediction with CU-level weight (BCW) [17]. VVC also adopted adaptive motion vector resolution (AMVR) [18], symmetric motion vector difference (SMVD) [19], and advanced motion vector prediction (AMVP) to save the bitrate of the motion parameters. Although these tools can significantly improve the coding performance, the computational complexity of a VVC encoder is up to nine times higher under the random-access (RA) configuration compared with that of HEVC [4].
Per the JVET CTC [7], VVC supports three encoding configurations [20]: all-intra (AI), low-delay (LD), and random-access (RA). In the AI configuration, each picture is encoded as an intra picture and does not use temporal reference pictures. Conversely, in case of the LD configuration, only the first frame in a video sequence is encoded as an intra frame; subsequent frames are encoded using the inter-prediction tools to exploit temporal reference pictures as well as intra-prediction tools. In the RA coding mode, an intra frame is encoded in an approximately one second interval in accordance with the intra period, and other frames are encoded as hierarchical B frames along the group of pictures (GOP) size. This means that a delay may occur in RA configuration because the display order is different from the encoding order.
Although the aforementioned VVC coding tools provide significant compression performance compared with HEVC, their computational complexity [21] of the encoder is high. Therefore, many researchers have studied various fast encoding algorithms for block structure, intra-prediction, and inter-prediction of VVC while maintaining the visual quality. In terms of block structure, Fan et al. proposed a fast QTMT partition algorithm based on variance and a Sobel operator to choose only one partition from five QTMT partitions in the process of intra-prediction [22]. Jin et al. proposed a fast QTBT partition method through a convolutional neural network (CNN) [23]. To reduce the computational complexity of intra-prediction, Yang et al. proposed a fast intra-mode decision method with gradient descent search [24] and Dong et al. proposed an adaptive mode pruning method by removing the non-promising intra-modes [25]. In terms of inter-prediction, Park and Kang proposed a fast affine motion estimation method using statistical characteristics of parent and current block modes [26].
Overall Frameworks of 2D Audio Coding
In general, the original 1D audio signal is generated from pulse code modulation (PCM), which involves sampling, quantization, and binarization processes. In this study, we used a 48 kHz sampled audio signal, and each sample was already mapped to signed 16 bits binary number by PCM coding. As the inputs of VVC specify a video sequence, which is an unsigned 2D signal with a max 12 bits integer number [7], we converted the original 1D audio signal into an unsigned 2D audio signal as the pre-processing before VVC encoding.
As shown in Figure 3, the overall architectures for 2D audio compression can be divided into the encoder and decoder side, where the encoder side consists of the 2D audio conversion process and the VVC encoder. In the 2D audio conversion process, signed 1D audio signals with a bit-depth of 16 bits are mapped to unsigned 1D audio signals with the bit-depth of 8 or 10 bits and then unsigned 2D audio signals are generated by 1D-to-2D packing as an input of the VVC encoder. After receiving a compressed bitstream, the decoder side conducts VVC decoding and inverse 2D audio conversion process to reconstruct the 1D audio signal with a bit-depth of 16 bits. Note that it may strongly influence the coding performance of the VVC depending on how the 1D signal is converted to a 2D signal.
Proposed 2D Audio Conversion Process
To generate unsigned 1D audio signals, we exploited three approaches as follows: 1.
Adaptive linear mapping (ALM) to generate unsigned 10 bits per sample.
Proposed 2D Audio Conversion Process
to generate unsigned 1D audio signals, we exploited three approaches as follows: 1. Non-linear mapping (NLM) to generate unsigned 8 bits per sample; 2. Linear mapping (LM) to generate unsigned 10 bits per sample; 3. Adaptive linear mapping (ALM) to generate unsigned 10 bits per sample.
NLM by Mu-law [27] is primarily used in 8 bits PCM digital telecommunication systems as one of two versions of ITU-T G.711. This reduces the dynamic range of an original signed 1D audio signal, as expressed in Equation (1): where and denote the sign function and the bit-depth of output samples, respectively.
Similarly, linear mapping (LM) maps the dynamic range of an original signed 1D audio signal into a specified sample range depending on the maximum value (Max) of samples. As shown in Figure 4, the sample's dynamic range of LM is from -Max to Max. This range can be divided into uniform intervals and each interval is represented with the type of unsigned N bits by Equation (2).
where ⌊ ⌋ and | | indicate the rounding operation and the maximum value of all samples, respectively. Note that the encoder side transmits the value of | | with the bit-depth of 16 bits for signed 1D conversion. Since this method has to search for the maximum value in the overall audio samples, this process is slow compared with the length of the audio signal. In addition, an unused sample range, called the dead zone, may exist in the range of the unsigned 1D signal, as shown in Figure 4. To resolve the aforementioned problems, we considered adaptive linear mapping (ALM). As shown in Figure 5, ALM firstly searches for both the and values among the signed 1D audio samples and then performs the LM conversion to represent the unsigned N bits 1D audio signal, as expressed in Equation (3). Because a video encoder generally NLM by Mu-law [27] is primarily used in 8 bits PCM digital telecommunication systems as one of two versions of ITU-T G.711. This reduces the dynamic range of an original signed 1D audio signal, as expressed in Equation (1): where sgn(I) and N denote the sign function and the bit-depth of output samples, respectively. Similarly, linear mapping (LM) maps the dynamic range of an original signed 1D audio signal into a specified sample range depending on the maximum value (Max) of samples. As shown in Figure 4, the sample's dynamic range of LM is from -Max to Max. This range can be divided into uniform intervals and each interval is represented with the type of unsigned N bits by Equation (2).
where x and Max(|I|) indicate the rounding operation and the maximum value of all samples, respectively. Note that the encoder side transmits the value of Max(|I|) with the bit-depth of 16 bits for signed 1D conversion. Since this method has to search for the maximum value in the overall audio samples, this process is slow compared with the length of the audio signal. In addition, an unused sample range, called the dead zone, may exist in the range of the unsigned 1D signal, as shown in Figure 4. To resolve the aforementioned problems, we considered adaptive linear mapping (ALM). As shown in Figure 5, ALM firstly searches for both the Min(I) and Max(I) values among the signed 1D audio samples and then performs the LM conversion to represent the unsigned N bits 1D audio signal, as expressed in Equation (3). Because a video encoder generally compresses input data in the unit of a 2D frame, ALM can reduce the encoding delay in the unit of a frame and avoid the dead zone problem compared with LM. Therefore, the encoder side
ALM(I k ) =
Electronics 2021, 10, x FOR PEER REVIEW 6 of 12 compresses input data in the unit of a 2D frame, ALM can reduce the encoding delay in the unit of a frame and avoid the dead zone problem compared with LM. Therefore, the encoder side has to transmit both and values for each frame, where is the index of the frames. After performing unsigned 1D conversion, the 2D audio signal is generated from the unsigned 1D audio signals used as the inputs of the VVC encoder. For example, if the spatial resolution (width height) of a frame is 256 256, the unsigned 1D audio signals, whose total number is 256 256, are arranged a frame according to the raster-scan order, as shown in Figure 6. Figure 7 shows the 2D packing results according to the aforementioned 2D conversion approaches. Although 2D audio signals showed different and complex texture patterns between consecutive frames, we confirmed that the 2D signals had a horizontally strong directionality within a frame. compresses input data in the unit of a 2D frame, ALM can reduce the encoding delay in the unit of a frame and avoid the dead zone problem compared with LM. Therefore, the encoder side has to transmit both and values for each frame, where is the index of the frames. After performing unsigned 1D conversion, the 2D audio signal is generated from the unsigned 1D audio signals used as the inputs of the VVC encoder. For example, if the spatial resolution (width height) of a frame is 256 256, the unsigned 1D audio signals, whose total number is 256 256, are arranged a frame according to the raster-scan order, as shown in Figure 6. Figure 7 shows the 2D packing results according to the aforementioned 2D conversion approaches. Although 2D audio signals showed different and complex texture patterns between consecutive frames, we confirmed that the 2D signals had a horizontally strong directionality within a frame. After performing unsigned 1D conversion, the 2D audio signal is generated from the unsigned 1D audio signals used as the inputs of the VVC encoder. For example, if the spatial resolution (width × height) of a frame is 256 × 256, the unsigned 1D audio signals, whose total number is 256 × 256, are arranged a frame according to the raster-scan order, as shown in Figure 6. Figure 7 shows the 2D packing results according to the aforementioned 2D conversion approaches. Although 2D audio signals showed different and complex texture patterns between consecutive frames, we confirmed that the 2D signals had a horizontally strong directionality within a frame.
Experimental Results
In this study, we used an original audio signal (48 kHz, 16 bits) whose length was approximately 195 s and consisted of five music, five speech, and five mixed items, as described in Table 2. To compress a 2D audio signal using the VVC encoder, we first set the spatial resolution to 256 256. Because the original audio signal had 9,371,648 samples, we generated 143 frames with a 256 256 spatial resolution. As demonstrated in the spectrograms in Figure 8, ALM is superior other conversion methods in terms of
Experimental Results
In this study, we used an original audio signal (48 kHz, 16 bits) whose length was approximately 195 s and consisted of five music, five speech, and five mixed items, as described in Table 2. To compress a 2D audio signal using the VVC encoder, we first set the spatial resolution to 256 256. Because the original audio signal had 9,371,648 samples, we generated 143 frames with a 256 256 spatial resolution. As demonstrated in the spectrograms in Figure 8, ALM is superior other conversion methods in terms of
Experimental Results
In this study, we used an original audio signal (48 kHz, 16 bits) whose length was approximately 195 s and consisted of five music, five speech, and five mixed items, as described in Table 2. To compress a 2D audio signal using the VVC encoder, we first set the spatial resolution to 256 × 256. Because the original audio signal had 9,371,648 samples, we generated 143 frames with a 256 × 256 spatial resolution. As demonstrated in the spectrograms in Figure 8, ALM is superior other conversion methods in terms of SNR. After generating 2D audio sequences with 143 frames, they were compressed by the VVC encoder under the experimental environments presented in Table 3. SNR. After generating 2D audio sequences with 143 frames, they were compressed by the VVC encoder under the experimental environments presented in Table 3.
Category Item Name Description
Music salvation Classical chorus music te15 Classical music Music_1 Rock music Music_3 Pop music phi7 Classical music Speech es01 English speech louis_raquin_15 French speech Wedding_speech Korean speech te1_mg54_speech German speech Arirang_speech Korean speech
Mixed twinkle_ff51
Speech with pop music SpeechOverMusic_1 Speech with chorus music SpeechOverMusic_4 Speech with pop music HarryPotter Speech with background music lion Speech with background music To evaluate the coding performance, we measured the SNR between the original and reconstructed 1D audio signals after decoding the compressed bitstream. Table 4 shows that NLM produced the worst coding performance due to the incorrectness of the signed 1D conversion. In addition, 10 bits LM had a higher SNR than ALM regardless of the encoding mode, as shown in Table 4 and Figure 9. To evaluate the coding performance, we measured the SNR between the original and reconstructed 1D audio signals after decoding the compressed bitstream. Table 4 shows that NLM produced the worst coding performance due to the incorrectness of the signed 1D conversion. In addition, 10 bits LM had a higher SNR than ALM regardless of the encoding mode, as shown in Table 4 and Figure 9. We further investigated if the objective performance was well-reflected subjective quality. In terms of subjective assessment, we conducted a MUSHRA te which is commonly used to evaluate subjective quality in the MPEG audio group. experiment, we used 10 audio items with a length of 10 s, and seven listeners partici To enable comparison with the state-of-the-art audio coding method, the same signal was coded by USAC [29]. Because the encoded frames have different quant parameters (QPs) according to the temporal layer of LD and RA configuration, SN frame can fluctuate between consecutive frames. Therefore, we performed MUSHR under the AI configuration without QP variations. In the AI configuration, we target to 0.5 bps and conducted MUSHRA listening tests for the original, USAC, LM, and NLM. As shown in Figure 10, the subjective quality of 10 bits ALM was than that of 10 bits LM, which opposes the results in Table 4. We further investigated if the objective performance was well-reflected in the subjective quality. In terms of subjective assessment, we conducted a MUSHRA test [28] which is commonly used to evaluate subjective quality in the MPEG audio group. In this experiment, we used 10 audio items with a length of 10 s, and seven listeners participated. To enable comparison with the state-of-the-art audio coding method, the same audio signal was coded by USAC [29]. Because the encoded frames have different quantization parameters (QPs) according to the temporal layer of LD and RA configuration, SNR per frame can fluctuate between consecutive frames. Therefore, we performed MUSHRA tests under the AI configuration without QP variations. In the AI configuration, we set the target to 0.5 bps and conducted MUSHRA listening tests for the original, USAC, ALM, LM, and NLM. As shown in Figure 10, the subjective quality of 10 bits ALM was higher than that of 10 bits LM, which opposes the results in Table 4. Electronics 2021, 10, x FOR PEER REVIEW 10 of 12 Figure 10. Scores of the MUSHRA listening test with a 95% confidence interval for different methods.
Although the average score of ALM or LM was lower than that of USAC, it was necessary to further investigate the possibility of researching 2D audio encoding with respect to 2D packing and video codec optimization. Firstly, the coding performance is strongly dependent on how the 1D signal is converted to a 2D signal. Because we used simple 2D packing in this study, it was necessary to consider a more elaborate 2D packing method suitable for video coding tools. For example, if we generate the 2D signals to efficiently perform the block based inter prediction, the coding performance of LD or RA configurations can be improved. Secondly, the current VVC has to transmit many unnecessary coding parameters due to the adoption of new tools, even though these tools are not used in the 2D audio encoding process. If a video encoder is optimized with core coding tools and light syntax structure, the coding performance will be improved.
Conclusions
In this paper, we proposed a pre-processing method to generate a 2D audio signal as an input of a VVC encoder, and investigated its applicability to 2D audio compression using the video coding scheme. To evaluate the coding performance, we measured both signal-to-noise ratio (SNR) and bits per sample (bps). Additionally, we conducted a MUSHRA test to evaluate subjective quality. The experimental results showed the possibility of researching 2D audio encoding using video coding schemes.
Conflicts of Interest:
The authors declare no conflict of interest. Although the average score of ALM or LM was lower than that of USAC, it was necessary to further investigate the possibility of researching 2D audio encoding with respect to 2D packing and video codec optimization. Firstly, the coding performance is strongly dependent on how the 1D signal is converted to a 2D signal. Because we used simple 2D packing in this study, it was necessary to consider a more elaborate 2D packing method suitable for video coding tools. For example, if we generate the 2D signals to efficiently perform the block based inter prediction, the coding performance of LD or RA configurations can be improved. Secondly, the current VVC has to transmit many unnecessary coding parameters due to the adoption of new tools, even though these tools are not used in the 2D audio encoding process. If a video encoder is optimized with core coding tools and light syntax structure, the coding performance will be improved.
Conclusions
In this paper, we proposed a pre-processing method to generate a 2D audio signal as an input of a VVC encoder, and investigated its applicability to 2D audio compression using the video coding scheme. To evaluate the coding performance, we measured both signal-tonoise ratio (SNR) and bits per sample (bps). Additionally, we conducted a MUSHRA test to evaluate subjective quality. The experimental results showed the possibility of researching 2D audio encoding using video coding schemes. | 6,414.8 | 2021-05-06T00:00:00.000 | [
"Computer Science"
] |
A Physical Layer Multicast Precoding and Grouping Scheme for Bandwidth Minimization
Physical layer multicasting exploits multiple antennas at the transmitter side to deliver a common message to a group of K users. To this end, two formulations have been well addressed in the literature: i) the max-min-fair criterion, which maximizes the signal-to-noise ratio (SNR) of the worst user for a fixed transmit power; and ii) the quality of service (QoS) formulation, which minimizes the transmit power subject to a target SNR. Nevertheless, it is known that the performance and complexity of these approaches is severely degraded as the group size grows. In this paper, we propose a different formulation that aims at minimizing the required bandwidth needed to provide the multicast service. This is achieved by dividing the users into smaller groups and assigning the bandwidth required to provide a target rate to each group. Contrary to the common belief, it is shown that dividing the users into different groups that use orthogonal bandwidth allocations can lead to a smaller aggregated bandwidth than the single-group with single bandwidth allocation counterpart, if an intelligent grouping scheme is used. An iterative algorithm to derive the optimal number of groups is presented with an stopping criterion to reduce the numerical complexity. It is shown through simulation that our proposed approach greatly reduces the required bandwidth compared to existing schemes that rely on single bandwidth allocation. Interestingly, results reveal that our proposed scheme also leads to a greater SNR for a randomly chosen user, and it reduces the variance of the required bandwidth, which eases the implementation in real networks.
I. INTRODUCTION
The increasing request and popularity of on-demand video and broadcast-like applications for smartphones has fueled an intensive area of research and standardization activities. This wave was initiated with Long Term Evolution (LTE) under multimedia broadcast/multicast service (MBMS), multicast/broadcast single frequency network (MBSFN) and single cell-point to multipoint (SC-PTM) technologies; and it is now under discussion for the Fifth Generation (5G) New Radio (NR), which has a work item targeted for Release 17 on June 2022 [1], [2].
MBMS defines MBMS service areas, that are composed by a number of cells that announce a list of available broadcast services. User equipments (UEs) can subscribe to these services to receive broadcast/multicast data [3]. Since both channel state information (CSI) reports and hybrid automatic repeat request (HARQ) protocols are not supported by this technology, achievable rates provided by multicast services are limited. To overcome this limitation, MBSFN (also known as enhanced MBMS) was introduced in Release 9. With this scheme, MBMS service areas are divided into MBSFN service areas, which are composed by a group of cells that are time/frequency synchronized to mitigate intercell interference. While this technology efficiently increases the achievable rates, it lacks of flexibility since the service areas and the resources are allocated statically [4].
SC-PTM solves these issues, since it allows a flexible resource allocation on a per-cell basis. Although this tech-nology does not support CSI reports nor HARQ on current releases of the LTE standard, it has been shown that SC-PTM outperforms MBMS and MBSFN in some scenarios [4]. More importantly, SC-PTM might support CSI reports on 5G, which would open the door to great performance improvements mainly for two reasons. Firstly, broadcast/multicast precoding could be exploited at the transmitter side [2], [5] to determine a sub-optimal beamforming vector for the intended group of users. Secondly, multicast channelaware user grouping and resource allocation can be used to maximize a given metric, i.e., system throughput or fairness among users.
A. RELATED WORK
There has been an extensive area of research on communication and signal theory communities to determine the optimal broadcast/multicast precoding, which is known as physical layer multicasting [6]. Attending to the specific scenario under consideration, existing studies focus either on singlegroup (i.e., broadcast) precoding [7]- [10] or multi-group (i.e., multicast) precoding [11]- [15]. The former involves that a single stream is delivered to a group of K UEs, using N transmit antennas, in the same time/frequency block of resources. The latter considers a type of group spatial division multiple access, where G groups are served in the same block of resources, but a different stream is transmitted to each group. Although multi-group multicast schemes have a greater flexibility, single-group multicast is normally preferred since it does not suffer from inter-group interference and it is simpler than the multi-group counterpart [6].
The determination of the optimal broadcast/multicast precoding admits two formulations [7]. The quality of service (QoS) formulation considers the minimization of the transmit power subject to a target signal-to-noise ratio (SNR) that must be fulfilled by each UE. An alternative formulation is the max-min-fair, which aims at maximizing the SNR of the worst UE, this being the key factor that limits the performance of the whole group, subject to a power constraint per UE.
Achieving the upper bound given by the multicast capacity requires precoding with high rank transmit covariance matrices, which is not feasible in practice [16]- [18]. For this reason, sub-optimal solutions that restrict to unit rank (i.e., transmit beamforming precoding) are widely adopted as single-group multicast schemes [7]- [10], [19]. Despite restricting to unit rank precoding, the aim of these works is to reduce the numerical complexity, since the computation of the multicast precoding requires treating the channel of a high number of users jointly.
A pioneering work is described in [7], where it is shown that max-min-fair and QoS formulations are equivalent NPhard problems, which can be expressed as a non-convex quadratically constrained quadratic program (QCQP). The authors propose a semi-definite relaxation (SDR) programming, which relaxes the non-convex unit rank constraint to have a convex problem that can be solved by semi-definite programming (SDP) followed by relaxation and Gaussian randomization (SDR-G). It is shown that solving the SDP problem leads to the upper bound on the min SNR, whereas the SDR-G yields good sub-optimal results close to the upper bound.
The performance of SDR-G deteriorates as N and K grows, and this motivated a number of research works to propose better approximations to the multicast beamforming problem. One of the best solutions in terms of performance is the successive linear approximation (SLA) algorithm, which is proposed for QoS problem in [20]. This approach involves an iterative algorithm where the non-convex constraints are linearized at each iteration by using first-order Taylor series expansion. The resulting convex problem is solved and the obtained vector is used in the next iteration. As shown in [9] with simulations, SLA outperforms SDR-G, although at the expense of a higher computation time.
A different approximation for the QoS problem is presented in [21]. In this work, a low-complexity algorithm based on QR decomposition and channel orthogonalization is proposed. The proposed algorithm is shown to provide a better performance than SDR-G when K N with a smaller complexity.
On the other hand, the case of max-min-fair problem is addressed in [22], where the non-convex part of the problem is replaced with an equivalent non-convex bilinear trace constraint, that is solved with alternating maximization (AM). It is shown that AM leads to a greater min SNR than SDR-G, but at the cost of a higher computational complexity.
The computational complexity is highly reduced in [19], which presents an appealing algorithm named successive beamforming (SB). This algorithm exhibits a high performance with moderate and small number of users and it reaches the upper bound for the case of K = 2. The algorithm performs orthogonalizations of the subspace spanned by each user's channel vector in an iterative fashion until there are no more spatial degrees of freedom left, which results in a reduced number of iterations (i.e., min(M, K)).
An interesting approach that achieves a good tradeoff between performance and complexity is proposed in [9] where class adaptive algorithms are developed. At each iteration, the beamforming vector is updated in the direction of an inverse SNR weighted linear combination of the SNRgradient vectors of all the users. It is shown that the proposed algorithms feature guaranteed convergence and state-of-theart performance at low complexity.
An algorithm to find the global solution is proposed in [23]. The algorithm is based on branch-and-bound strategy, combined with a new argument-cut technique that is used to design convex relaxations of non-convex constraints. Simulation results show that the proposal greatly outperforms state-of-the art techniques when N and K are high, but the computational complexity makes this approach unfeasible when the channel varies quickly.
All of the above works focus on multicast precoding given a fixed number of groups; nevertheless, user grouping and resource allocation offer additional degrees of freedom that can be used to improve the system performance. The resource allocation and user grouping problems are often posed as maximization problems with an extremely large solution space [24], [25], which makes low-complexity solutions specially appealing. In [26] an opportunistic multicast scheduling (OMS) algorithm is proposed to exploit multiuser diversity and to increase aggregated throughput. Nevertheless, OMS does not guarantee fairness among multicast users. The fairness between multicast and unicast users is addressed in [27]- [29], which is important from the operators perspective, since it maintains a good balance between both services. Another strategy is to form groups that maximize the aggregated throughput [30]- [33]; however, these works restrict to single antenna case and do not consider multicast precoding. User grouping has been also studied for multiantenna systems with hybrid beamforming. In this context, different user grouping strategies based on channel sparsity in the beam domain has been proposed in [34]- [37] for other applications such as physical layer security or wireless information and power transfer (SWIPT). These methods assume a spatial basis expansion model, e.g., beam-domain and angular-domain, which compress the dimension of the channel. This reduces the complexity of related tasks such as channel estimation, suppression of pilot contamination or user grouping. Here, users are grouped based on the similarity of the angle of arrivals [36] or the active beam domain sets to eliminate inter-group interference [35].
B. CONTRIBUTIONS
Despite of their relevance, none of the aforementioned works have addressed two paramount aspects: i) the problem of bandwidth minimization; and ii) the interactions between multicast precoding and user grouping.
The former aspect is closer to the problem that face operators and vendors in practical deployments, which aim to deliver a broadcast/multicast service with a given rate, using as less resources as possible to maximize their profits. Regarding multicast services, this bandwidth minimization problem has been only addressed for core networks [38]- [40]. These works propose different techniques to group service demands of the same content, thus saving bandwidth in the backbone. As for wireless access networks, the bandwidth minimization problem has been investigated only for unicast transmission, e.g., with non-orthogonal multiple access (NOMA), [41], [42] and wireless relaying scenarios [43]. Nevertheless, to the best of the authors' knowledge, this problem has not been investigated yet for physical layer multicast on wireless access networks.
The latter aspect is related to the fact that physical layer multicast has been mainly investigated under two different directions: i) multicast precoding; and ii) user grouping/resource allocations. On the one hand, the works that focus on multicast precoding normally assume fixed groups. On the other hand, the works that address user grouping and resource allocations either ignore the multi-antenna setups or assume simplified models that do not capture the dependence between the user group and the performance of multicast precoding.
These reasons motivated us to investigate the interactions between user grouping/resource allocation and multicast precoding. As a result, we propose an scheme that minimizes the required bandwidth to deliver a given broadcast/multicast service. The contributions of the present work can be summarized as follows: • We propose a new problem formulation for multicast precoding and grouping. The proposed formulation aims at determining a user partitioning into groups and a multicast precoding for each group that minimizes the required bandwidth. • We propose a novel low complexity algorithm for the above problem named adaptive multicast grouping (AMG). The proposed algorithm considers three grouping criteria that differ in terms of the required information and performance. • The implementation aspects of the proposal are addressed and specialized for the context of 5G NR networks. It is discussed the availability of the information required by the AMG algorithm in current 5G networks. To this end, the signaling mechanisms, the available measurements and UL reports that might be used to implement the proposed algorithm are identified. • The performance of the proposed scheme is assessed with extensive simulations. It has been investigated a plethora a performance indicators, which includes the average, the variance and the distribution of the minimum required bandwidth, the distribution of the optimal number of groups, the distribution of the SNR and the average computation time. Results reveal that the proposed approach greatly reduces the required bandwidth compared to existing schemes that rely on single bandwidth allocation. It is also shown that it leads to a greater SNR for a randomly chosen user, and it reduces the variance of the required bandwidth, which eases the implementation in real networks.
The remainder of this manuscript is structured as follows. The system model and problem formulation is depicted in Section II whereas the proposed multicast scheme is described in Section III. Section IV illustrates the benefits of our proposal with extensive simulation results. Finally, some conclusions are drawn in Section V. Notation: The following notation is used throughout the text. Matrices and vectors are represented with boldface uppercase and lowercase letters, respectively. If Σ is a matrix, [Σ] n,m is used to identify its (n, m)-th element. A matrix with N 1 rows and N 2 columns where all its elements are equal to 0 is written as 0 N1×N2 . (•) * denotes conjugate of a complex number whereas Re{•} and Im{•} denote the real and imaginary parts, respectively; j = √ −1 stands for the imaginary unit. (•) T
RF Chain
Transmit RF beam matrix, then X 0 indicates that such a matrix is positive semi-definite. N, R and C stand for the sets of natural, real and complex numbers respectively. If A is a set, then |A| is the cardinality of that set; however, if a is a complex number, |a| represents its modulus. CN (0 K×1 , Σ) denotes the circular symmetric complex Gaussian distribution with zero mean and covariance Σ.
II. SYSTEM MODEL
We consider a hybrid beamforming scheme, as considered in 5G for millimeter-wave (mmW) bands, where the precoding is composed of two stages: i) a digital base-band beamforming; and ii) an analog/radio frequency (RF) beamforming [44]. On the one hand, base-band digital precoding is implemented by multiplying the complex IQ constellation symbols by a complex beamforming vector, and thus different precoding vectors can be applied to different blocks of time/frequency resources. By contrast, RF beamforming is implemented with phase shifters after digital-to-analog converters (DACs) and frequency up-conversion in the transmitter side, and thus RF beams can be multiplexed in time domain only, and not in the frequency domain [44], [45]. It is considered a hierarchical and modular RF and digital beam management as in 5G networks [46], [47]. This implies that RF beam management and digital precoding computation are performed independently. Fig. 1 illustrates a simplified block diagram of the hybrid beamforming scheme considered in this work. On the base station (BS) side, it is observed that the base-band digital beamforming block receives a single stream of IQ constellation symbols and outputs N streams which are delivered to N RF chains for digital-to-analog and frequency up-conversion. In the context of 5G, the output of each RF chain is named logical antenna port [48], and it is connected to a different antenna array. The number of physical antennas that form the different antenna arrays determine the beamwidth and gain of the transmit RF beams that can be synthesized [44], [49]. In this work it is assumed that the UE has a single RF chain with a single physical antenna to reduce the cost of mobile handsets.
The transmit RF beams are modeled using the widely adopted sectored-pattern model [50]- [53], where it is considered that the main lobe has a constant gain of G m with a beamwidth of θ radians, and it is centered at the steering angle ϕ. Angles that do not fall within the main lobe have a constant back lobe gain G b . It is assumed that the RF gain of the main lobe is related to the beamwidth as G s ≈ 2π/θ [52], [54]. It is considered that UE receiver has a single RF chain connected to a single physical antenna for the sake of cost saving. For this reason it is assumed that the gain of the receive beam is just 1. Thus, with this system model, the relevant parameters are the beamwidth of the transmit RF beam, which also determines its beam gain, and the number of RF chains, N , which determines the gains that can be achieved by digital multicast beamforming.
Finally, it is considered that RF beams are time-domain multiplexed so there is no inter-beam interference. In addition, it is considered that mmW systems are noise-limited rather than interference-limited and thus inter-cell interference can be neglected, e.g., due to the high path-loss exhibited at mmW bands, the directivity of the RF beam patterns and the use of frequency planning strategies [55], [56].
A. SPATIAL MODELING
The probe BS is assumed to be placed at the origin, giving service to a cell of radius d c meters and it is equipped with N RF chains, whereas the UEs have a single RF chain. We focus the analysis on a probe RF beam, and thus we consider the set of K UEs that are served by that RF beam. The K UEs can be divided into G ∈ [1, K] ⊂ N groups, and a different digital beamforming vector, w g ∈ C N ×1 , can be used for Each UE is identified by a given index, k ∈ K = [1, K] ⊂ N, where K represents the set of UEs. The set of UEs that belong to the group g ∈ [1, G] ⊂ N is expressed as K g ⊆ K.
Each UE is assigned to a single group, i.e., G g=1 K g = K and G g=1 K g = ∅. Besides, we define the function K g (u), as the u-th ordered element of the set K g . For instance, if we have the following UE set, K = {1, 2, 3, 4, 5}, with K 1 = {1, 3} and K 2 = {2, 4, 5} for G = 2; then K 1 (1) = 1, K 1 (2) = 3, The UEs are associated with the RF beam that provides the highest received power [47], [57], [58] and thus the UE locations fall within the region R ∈ R 2 , which is defined by its main lobe as , being x the Euclidean norm of x, and ∠x ∈ (−π, π] its angle. In addition, the UE locations are drawn randomly according to a point process (PP), In this work we have considered two spatial distributions: i) the uniform Binomial point process (BPP), that models zero interaction between node locations; and ii) a clustered point process (CPP) that models spatial correlation between nodes [59]. The BPP places randomly K points within the region R with uniform distribution. The CPP is expressed where C is the number of clusters and x i represents the center of the i-th cluster. The cluster centers are drawn randomly within the region R with uniform distribution. The PP Φ i represents the distribution of daughter points and it is modeled as a BPP that places C i points within a disk centered at the origin with radius d d . The sum of all daughter points placed by each of the clusters is K, i.e., Fig. 2 shows a given spatial realization for the two PPs considered in this work.
B. CHANNEL AND SIGNAL MODELING
For a given allocation block of time/frequency resources, the channel is assumed to be flat in the time and frequency domains. A base-band equivalent channel model is considered, and thus, the channel between the BS and the k-th UE, h k ∈ C N ×1 , is modeled as a zero-mean complex random vector that accounts for the RF beam patterns gains, the pathloss and fast fading. Thus, the received signal by the k-th UE, which belongs to user group g, is given by where w g ∈ C N ×1 is the multicast beamforming vector and s g represents the IQ constellation symbols intended for group g with zero mean and unit power, i.e., E[|s g | 2 ] = 1 and E[|s g |] = 0. p t represents the transmit power per constellation symbol which spans a bandwidth of ∆f Hz. Wide-sense stationary additive Gaussian noise is represented as z k , and its power is p n on ∆f Hz. It is assumed a constant power spectral density (PSD) in the frequency domain, and thus t = p t /∆f W/Hz and N 0 = p n /∆f W/Hz represent the transmit and noise power spectral densities, respectively. The channel vectorḧ k ∈ C N ×1 is expressed as where x k is the distance between the k-th UE and the probe BS, andh k = [h 1,k , ..,h N,k ] T ∈ C N ×1 models the fast fading. The path-loss model is based on the classical power law, x k −α/2 , where α > 0 is an environmental dependent path-loss exponent. The complex gain due to fast fading between the i-th transmit logical antenna and the k-th user has a marginal distribution according to a zero mean unit power complex Gaussian distributionh k,i ∼ CN (0, 1).
Besides the uncorrelated case, it is considered the case where the fading of different users can be correlated based on their distance. This latter distance dependent correlation model is based on the observation made in some works, (e.g., [60] and references therein), where it is shown that there exists a correlation between nearby locations that tend to decrease as the distance increases. Therefore, it is proposed to model the correlation between the k-th and q-th UEs as , being β ≥ 0 a factor that models how strong the correlation is and x k −x q the distance between the two UEs. It has been chosen a decreasing exponential to capture the fact that the correlation is stronger between nearby locations and decreases as the distance increases. Besides of this, the proposed model considers the case of maximal correlation, i.e., ρ k,q = 1, and the independent case as particular cases that are modeled with β = 0 and β → ∞ respectively. Finally, the vector of complex gains for the i-th transmit logical antenna w.r.t. the K UEs is generated according to a multivariate Gaussian distribution ash
C. BANDWIDTH ALLOCATION AND RATE ADAPTATION
It is assumed that a broadcast/multicast service is intended to be delivered to K UEs associated with the probe RF beam. The target binary rate of such a service is R T bps. The K UEs are divided into G groups that use a different beamforming vector, w g , on a different bandwidth allocation, B g , of |B g | Hz. The bandwidth allocations to different groups are orthogonal to avoid inter-group interference and thus The overall bandwidth allocated to the broadcast/multicast service is The SNR of UE k that belongs to group g, γ k (w g ), is given by where h k ∈ C N ×1 is the scaled channel gain, which is expressed as the average SNR when N = 1. The transmission rate of each group is adapted to the min SNR of the group, and thus the spectral efficiency (SE) (i.e., rate) of group g is expressed as log 2 (1 + min k∈Kg γ k (w g )) bps/Hz. VOLUME 4, 2016
III. PROPOSED USER GROUPING, BANDWIDTH ALLOCATION AND PRECODING SCHEME A. PROBLEM FORMULATION AND PROPOSED ALGORITHM
We consider a precoding and grouping scheme that aims at minimizing the bandwidth |B| needed to provide a broadcast/multicast service that requires R T bps. Therefore, the problem can be posed as obtaining the user division into G groups, K g , with g ∈ [1, G], and multicast beamforming vector for each group, w g ∈ C N ×1 , that minimizes the overall required bandwidth. More formally, this problem can be formulated as where (7) comes after (3) and the fact that the achievable binary rate of every group g ∈ [1, G] ⊂ N must be equal to R T . To seek for a sub-optimal solution, we split this problem in two parts: user grouping and multicast precoding. The user grouping assumes a number of groups and outputs the partition of users into G groups, K g ∀g ∈ [1, G] ⊂ N.
1) User grouping
The proposed user grouping algorithm relies on the observation that highly correlated channels increases the multicast beamforming gain [9], [18]. Thus, the proposed algorithm aims at assigning the same group to users whose channel is similar. To this end, a K-means++ (KM) algorithm has been selected due to its reduced complexity and quality of final solution [61]. This clustering algorithm partitions a data set of K points into G groups using an iterative algorithm to minimize the sum of data sample-to-centroid distances, summed over all G clusters. The centroid of each group is the mean of the points that belong to the group and it is also an output of the algorithm. We define three different data set types, which are used as input to KM, to devise three grouping algorithms that differ in performance and complexity. i) Scaled channel matrix: This option considers that the data set is the matrix H ∈ C N ×K , which is built by stacking the scaled channel gain of each user, i.e., h k ∈ C N ×1 ∀k ∈ K as column vectors. Since KM algorithm restricts to real data samples, the input data set for KM algorithm is formed by stacking the real and imaginary parts of H as [Re{H} T , Im{H} T ] T ∈ R 2N ×K . Hence, each data sample is a real point in R 2N . This grouping algorithm is labeled as KM-CSI.
ii) Location information: Here the input to the KM algorithm is the position x k ∈ R 2 of every user k ∈ K. The data set is the PP of all UEs locations, Φ, which can be arranged as a matrix Φ ∈ R 2×K . Compared to KM-CSI, this type of data set lacks of information about the instantaneous channel gain of every user; nevertheless, the size of the data set is smaller as well as its complexity. This grouping algorithm is labeled as KM-loc.
iii) Reference signal received power (RSRP): The RSRP represents the average received power by a given user k and it is expressed as: µ k = x k −α G s t . The data set is then a vector of K real elements µ = [µ 1 , .., µ K ] ∈ R 1×K . Contrarily to KM-loc and KM-CSI, this method lacks of information about distances between the different users and it only has information about distances towards the BS. However, it has a smaller data size and complexity than the other two options. It is labeled as KM-RSRP.
The proposed algorithm, AMG, requires any of these three types of information. The availability of such information in 5G networks is discussed in Section III-B whereas the impact of the type of information used on the performance is assessed in Section IV-B.
2) Multicast Beamforming
The multicast precoding considers a partition of the users into G groups and then it computes a beamforming vector for each group that maximizes the min SNR. The following stateof-the art multicast precoding vectors have been considered in this paper: i) SDR-G: This algorithm was proposed in [7]. The algorithm approximates the max-min-fair problem into the following convex problem that can be solved via SDP. Hence, with SDR-G, the multicast beamforming vector for group g can be computed as follows: 1) Solve the relaxed SDP problem to obtain the positive semi-definite matrix, X ∈ C N ×N : s.t. trace(XQ k ) ≥ t, ∀k ∈ K g (10) trace(X) = 1, X 0 where Q k ∈ C N ×N and Q k = h k h H k 0.
2) Perform Gaussian randomization. This involves that
M triplets of candidate beamforming vectors are randomly generated and the best one after M realizations is selected. Each triplet consists on the following vectors, w where U ∈ C N ×1 is obtained after eigendecomposition of X = UΛU H and [e where e (b) g ∈ C N ×1 are generated using the same procedure as for e (a) g ∈ C N ×1 , and finally where e (c) g ∈ C N ×1 is a vector of zero-mean, unitvariance complex circularly symmetric uncorrelated Gaussian random variables. ii) Adaptive update (AU): This is an iterative algorithm which is proposed in [9]. With this algorithm, each update takes a step in the direction of an inverse SNR weighted linear combination of the SNR-gradient vectors of all |K g | users. At iteration m, the beamforming vector of group g is updated as follows where ξ is the fixed positive step-size for every iteration and is a positive constant that is introduced for numerical stability.
iii) SB: this algorithm involves a maximum of min(N, K) iterations. With iteration k the algorithm equalizes the SNRs of the k UEs with smallest SNR, and it stops if the SNRs of the rest of users is greater than that value or if there are no more degrees of freedom, i.e., k + 1 > min(N, K). The details are described in [19]. iv) Random beamforming (RBF): this scheme considers that the beamforming vector is randomly generated according to a complex Gaussian distribution where [w g ] i ∼ CN (0, 1/N ).
Both stages, user grouping and multicast precoding, assume a number of groups, G. Hence, the proposed algorithm goes through an increasing number of groups in an iterative fashion, starting with a single group. The proposed algorithm for user grouping, bandwidth allocation and beamforming, referred to as AMG, is summarized in Algorithm 1, and it is explained as follows. Firstly, the algorithm sets the initial required bandwidth with the highest value, and the variable stop, which is used as stopping criterion (lines 1 to 3). Then, a loop is executed to search for the best number of groups, , within the range [1, K − 1] (lines 4 -11). At each iteration, the algorithm partitions the users in groups according to either KM-CSI, KM-loc or KM-RSRP criteria (line 6). Afterwards, it computes the multicast precoding for each group, w g ∈ C N ×1 , and the required bandwidth, |B ( ) |, (lines 7 -10). To save computation time, we can limit the maximum number of iterations (i.e., groups) to a given value, G max . The algorithm iterates until the maximum number of groups to explore, G max , is reached, or the stopping criterion is fulfilled. We propose as stopping criterion whether the required bandwidth of current iteration, , is greater than the bandwidth of the previous iteration (line 11). The effect of such an stopping criterion will be assessed in Section IV with simulation results. The obtained solution, which can Algorithm 1 AMG Input: K, h k ∀k ∈ K Output: G , γ min,g , |B g |, w g ∈ C N ×1 , K g ∀g ∈ [1, G ] Data: Data for user grouping: explore a number of groups up to K − 1, is finally compared with the results of unicast transmission, where the number of groups is K (lines 12 -16). This unicast transmission uses MRT beamforming which achieves the capacity of the MISO channel. If the bandwidth of multicast transmission with groups is smaller than the one required by unicast transmission, then G = , whereas unicast transmission is selected otherwise.
B. IMPLEMENTATION ASPECTS
In this subsection the implementation aspects of the proposed AMG scheme in real systems and its application to 5G NR are discussed. The block diagram of the proposed scheme is illustrated in Fig. 3. The time is divided into transmission time intervals (TTIs), and it is considered that the channel is time invariant within the TTI although it varies between different TTIs. It is observed that the AMG block computes the number of groups, G , the min SNR, γ min,g , and the required bandwidth, |B g |, per group. All these metrics are forwarded to the Rate Adaptation block, which adapts the symbol rate intended to each group to the link conditions of the worst UE, given by γ min,g . In real systems, e.g., 5G NR, this is achieved by selecting an appropriate modulation and coding scheme (MCS).
The constellation symbols for each group, s g , are delivered to the precoding stage, where they are multiplied by the beamforming vector per group, w g ∈ C N ×1 , which is obtained by the AMG block. The precoded symbols for each group are delivered to the Bandwidth Allocation block which stacks blocks of symbols for each group and maps them into the portion of bandwidth allocated to each group. As seen in the figure, the bandwidth that is not used by the broadcast/multicast service is available to other services. Since the required overall bandwidth, |B g |, depends on the channel of every UE, it varies on a TTI basis. Finally, the digitally beamformed stream to each group is transmitted using the same RF beam, which is received by the K UEs.
The 5G standard supports several reporting quantities that can be used to develop the proposed scheme. Firstly, the scaled channel matrix per UE, h k ∈ C N ×1 , is used by the multicast beamforming stage as well as user grouping in case of KM-CSI. This metric can be obtained with the precoding matrix indicator (PMI) and channel quality indicator (CQI) report quantities [47], [57]. The former report quantity leads to a precoding vector that belongs to a given codebook, whose type is specified by higher layer configuration [62]. This precoding vector can be understood as a quantized version of the Hermitian channel vector, i.e., h H k ∈ C 1×N , since this is the optimal precoder of the single user MISO channel [63]. The CQI, on the other hand, indicates the appropriate MCS to achieve a block error rate (BLER) below a given target value, which can be configured to either 10 −1 or 10 −5 in 5G networks [64]. This CQI is computed by the UE based on the estimated SNR [65], and thus, it can be used in real systems in combination with the PMI to get the scaled channel matrix given by (5).
Another required metric is the RSRP, which is needed at the user grouping stage if KM-RSRP algorithm is used. This metric actually corresponds to an existing report quantity which is named L1-RSRP in the 5G standard [66].
Lastly, the location information is used by the user grouping stage in case of KM-loc. This can be achieved thanks to 5G NR positioning protocol of current releases 15 and 16. Some methods that are part of the standard and can be used to this end are uplink time difference of arrival (UTDOA), enhanced cell ID (E-CID), multicell round trip time (Multi-RTT) and uplink angle of arrival (UL-AoA) [67].
As it was justified in the introduction, a paramount challenge of multicast precoding algorithms is the computational time due to the complexity of the underlying optimization problems that needs to be solved. The solution of the user grouping and multicast precoding determined by the AMG algorithm is valid for a time period where the joint channel of all the users can be considered as roughly constant. Thus, this period of time, which is named coherence time, acts as a system requisite for the chosen multicast precoding and user grouping algorithm. As shown in [68], the coherence time highly depends on different factors such as the velocity of the users, the frequency band and the beamwidth. From Fig. 4 it is observed that the coherence time ranges from thousands of ms down to a few ms or even fractions of ms. It is seen that reducing the beamwidth increases the coherence time, whereas it can be reduced by increasing the frequency band and/or the user velocity. More specifically, for a beamwidth of 45 degrees a pedestrian that moves at 3 km/h and receives a multicast transmission at 7 GHz has a coherence time around 50 ms. If the set of users has this coherence time, the AMG algorithm should take less than 50 ms to compute the user partition into groups and the related multicast beamforming vectors. Nevertheless, as explained in Section III, the AMG algorithm can use different precoding (e.g., SDR-G, AU, RBF) and grouping algorithms (i.e., KM-CSI, KM-loc, KM-RSRP) that lead to different computational time as well as different performance. So an appropriate choice can be selected based on the expected coherence time to have an appropriate balance between performance and computational time.
IV. SIMULATION RESULTS AND DISCUSSIONS
The performance of the proposed scheme is assessed with simulations. It is considered a path-loss exponent of α = 4, and a beamwidth of θ = 2π/8 radians with a main lobe gain of G s = 8. The thermal noise is assumed to be N th = −174 dBm/Hz, with a noise figure, N F = 7 dB, and thus the noise PSD is N 0 = N th + N F = −167 dBm/Hz. The results have been obtained through Monte Carlo simulations with 10 3 realizations. The default parameters considered in this section are summarized in Table 1. These parameters have been used to obtain each figure unless otherwise stated in the caption.
The performance of the proposed algorithm, AMG, is compared to two extreme alternatives: i) unicast transmission, which considers MRT beamforming and orthogonal bandwidth allocation to each UE; and ii) broadcast, which involves a single group with single bandwidth allocation and a single multicast beamforming vector.
Three multicast beamforming algorithms are considered as described in Section III. The AU algorithm uses a step size ξ = 0.1, with a factor of = 10 −3 to avoid numerical instability and 100 iterations. The SDR-G algorithm uses M = 30N K randomizations as recommended in [7], which involves 7200 random trials for the default parameters.
The proposed algorithm described in Algorithm 1 considers an stopping criterion as described in line 11. To assess the impact of such a stopping criterion it is also considered the case where the algorithm executes G max iterations (line 11 is removed). This allows us to evaluate the increment in the required bandwidth due to the early stopping but also the savings in computation time.
Next subsections illustrate the performance of the proposed algorithm under diverse scenarios to get insights about the performance trends and interplay between different parameters.
A. PERFORMANCE WITH INDEPENDENT FADING AND BPP SPATIAL DISTRIBUTION
Firstly, the performance of AMG is assessed for a BPP spatial distribution. This type of distribution models independent locations of the UEs, which are drawn randomly with uniform distribution. The fading is also assumed to be independent. Figures 5 and 6 show the mean and CDF of the required bandwidth versus the transmit PSD for the proposed AMG algorithm, with and without an stopping criterion. Results for the case of unicast and broadcast transmissions are also presented.
It is observed that our proposal greatly outperforms broadcast and unicast transmissions for a wide range of transmit PSD values. This demonstrates that it is beneficial to divide the users into different groups, since the required bandwidth can be potentially reduced. On the one hand, the performance of multicast precoding algorithms is deteriorated as the group size increases [6]. This involves that dividing the users into smaller groups will lead to a higher min SNR, min k∈Kg γ k (w g ), of the groups than the single group counterpart. If the increase in the min SNR is high enough, then it leads to an smaller required bandwidth than using a single group. On the other hand, grouping users with similar channel realizations increases the gain of the multicast precoding. This statement is based on the observation that correlated scaled channels (including fast fading and average SNR) increase the performance gain of multicast precoding [9], [18]. Hence, our proposed AMG scheme relies on these two ideas to find a sub-optimal number of groups, G , user partition, K g ∀g ∈ [1, G ], and multicast beamforming that minimizes the required bandwidth.
The performance loss due to early stopping is negligible as it can be observed from the average required bandwidth in Fig. 5 as well as from its distribution in Fig. 6. Interestingly, the average required bandwidth is smaller for unicast than for broadcast transmission in the low transmit power regime. This means that, on average, the impact of the increase in min SNR, which yields an increase on the SE, leads to an smaller aggregated bandwidth than the single group counterpart. However, as the transmit power is increased, the required bandwidth of broadcast transmission tends to be much smaller than the unicast alternative. Fig. 7 illustrates the PMF of the sub-optimal number of groups, G , with (red) and without (blue) an stopping criterion, for different transmit PSD values. It can be observed that the distribution of G takes a wider set of values when the transmit power is small. Nevertheless, as the transmit power is increased, the range of values is narrower and more concentrated around small values. This might be expected in view of Fig. 5 and 6 since the performance of broadcast gets closer to the performance of AMG as the transmit power is increased.
If we compare the PMF of the number of groups with and without stopping criterion, we observe that the effect of the stopping criterion is to concentrate the distribution around smaller values of G . This is related to the fact that the early stop prevents from searching solutions related to greater number of groups.
As it is seen in the pseudo-code of Algorithm 1, the algorithm always searches a solution in the unicast case, since computing the unicast beamforming vector is trivial in terms of computational complexity. When an stopping criterion is used and t = −100 dBm/Hz, there is a probability of 0.067 to find the unicast case as sub-optimal solution (i.e., G = K).
Nevertheless, for the same transmit PSD of −100 dBm/Hz, the unicast case is not selected if no stopping criterion is considered, which is labeled in the legend as G max = K. This involves that the optimal solution is never the unicast solution, i.e., G = K, if an exhaustive search is considered. Yet, the unicast solution is better than the solution found by the loop of lines 4-11 in Algorithm 1, if an stopping criterion is used, with a probability of 0.067.
B. EFFECT OF GROUPING ALGORITHMS
The effect of the three proposed grouping algorithms is illustrated in Fig. 8. As it is mentioned in Section III, among the 3 grouping algorithms, KM-CSI is the one that requires a higher amount of data, i.e., 2N K real numbers. This algorithm is followed by the KM-loc, which requires 2K real numbers and KM-RSRP, which requires K real numbers. As it can be seen in Fig. 8, the performance of the different grouping algorithms follows the amount of data used. Thus, the smallest bandwidth in statistical terms is obtained by KM-CSI, whereas KM-RSRP leads to the highest bandwidth and KM-loc obtains intermediate results.
C. SNR DISTRIBUTION AND VARIANCE OF REQUIRED BANDWIDTH
The CCDF of the SNR of a randomly chosen UE is shown in Fig. 9 for AMG, unicast and broadcast schemes with different transmit PSD values. It can be observed that unicast transmission, which is based on MRT, achieves the highest SNR in statistical terms. This is expected since multicast precoding uses the same beamforming vector for a group of users, and thus its performance deteriorates as the group size increases. Nevertheless, broadcast transmission, which uses single group multicast precoding, achieves a much smaller required bandwidth than unicast transmission for a broader set of transmission power values, as it was discussed in Fig. 5 and 6. Interestingly, AMG achieves a smaller required bandwidth than broadcast and unicast approaches, but at the same time, it leads to a higher SNR than broadcast transmission. Fig. 10, on the other hand, shows the variance of the required bandwidth for AMG, unicast and broadcast schemes versus the transmit PSD, t . The variance of the required bandwidth is an important metric for frequency planning. As it was mentioned in Section III-B, the bandwidth that is not used by the multicast/broadcast service can be used by other services. Therefore, having an small variance of the required bandwidth is highly appealing, since it eases the frequency planning of the other services that can be accommodated in the available bandwidth. It can be observed that AMG achieves an smaller variance than broadcast transmission in the considered range of transmit power and it also achieves an smaller variance than unicast transmission for t > −85 dBm/Hz, while for t < −85 dBm/Hz the difference in terms of performance between AMG and unicast is negligible.
D. PERFORMANCE VERSUS COMPLEXITY
The trade-off between performance and complexity for different multicast precoding algorithms is shown in Fig. 11 and 12. Results reveal that unicast transmission and broadcast with RBF lead to the smallest average computation time. Nevertheless, the required bandwidth for broadcast with RBF is the highest, and its performance is clearly inferior to other alternatives. These results highlight the importance of multicast precoding, which leads to great performance improvements compared to the RBF case, thanks to the use of channel information of all the users and different RF chains. It is observed that AMG greatly improves the performance compared to broadcast and unicast alternatives. Even with RBF, AMG leads to an smaller required bandwidth than the unicast case and its broadcast counterpart. This is specially relevant at high transmit powers, where the required bandwidth of AMG with RBF tends to be also greatly smaller than the unicast case.
As seen from Fig. 11, the smallest required bandwidth is obtained with SDR-G precoding. However, as it is observed in Fig. 12, this precoding leads to the highest computation time with both (the broadcast and AMG) alternatives. Hence, SDR-G precoding does not seem appropriate for real implementations due to its high numerical complexity. After SDR-G, AU is the precoding technique that leads to the smallest required bandwidth for broadcast transmission as it is illustrated in Fig. 11. In case of broadcast, SB requires a greater bandwidth than AU as it is expected, since AU leads to a higher minimum SNR for a high number of users [9]. Nevertheless, it is observed that AMG/SB leads to a smaller bandwidth than AMG/AU, which suggests that SB performs better with AMG. The reason behind this is that SB exhibits a high performance for a small number of users, and it reaches the upper bound for the case of K = 2 as shown in [19]. Since AMG partitions the users into smaller groups, this can explain why SB leads to an smaller bandwidth when it is used by the AMG algorithm.
The greater differences observed in terms of computation time are due to the chosen multicast precoding option. For instance, with broadcast transmission under the simulated parameters listed in Table 1 and = −100 dBm/Hz, SDR-G precoding requires around 176 times more computation time than AU; AU consumes roughly 42 times more time than SB; whereas SB requires around 8 times more time than RBF.
The increment in terms of computation time due to the use of AMG is clearly smaller than the differences observed between different multicast precoding algorithms. More specifically, the increment of computation time with AMG/AU with respect to broadcast/AU is around 4.10 times; whereas the increment of AMG/SDR-G with respect to broadcast/SDR-G is roughly 4.28 times.
As it was discussed in Section III-B, the coherence time of the channel imposes a system requisite for the chosen multicast precoding and user grouping algorithm. It is observed from Fig. 12 that AMG/AU leads to around 50 ms of average computational time, whereas AMG/SB leads to around 8 ms and AMG/RBF leads to 4 ms. AMG/SDR-G requires the highest computation time, which is around 5 s due to its high complexity. As it is observed, SB requires a small computational time compared with SDR-G and AU. The reason behind this is that the number of iterations performed by SB is limited to min(N, K), which is greatly smaller than the number of iterations and randomizations required by AU and SDR-R (i.e., 100 iterations and 7200 random trials respectively). These simulation results have been obtained with an Intel i7 processor and MATLAB R2020b. Nevertheless, such computation times can be greatly reduced implementing and optimizing the code, e.g., for C++ and using more powerful processor as used in commercial 5G base stations.
Finally, the saving in computation time due to the use of an early stopping is also assessed for the case of AMG/AU in Fig. 12. It is observed that such a saving is around 26 times, which is considerable bearing in mind that the early stopping leads to minor reduction in required bandwidth.
E. EFFECT OF SPATIAL AND FADING CORRELATIONS
The effect of the spatial and fading correlations is assessed in terms of the CDF of the required bandwidth, in Fig. 13, and CCDF of the SNR of the typical UE, in Fig. 14 It is confirmed that both the spatial and fading correlations are beneficial in terms of required bandwidth and SNR of the UEs since they increase the SNR and reduce the bandwidth in statistical terms.
F. EFFECT OF THE NUMBER OF UEs AND RF CHAINS
To conclude this section, the effect of the number of served users and available RF chains is shown in Fig. 15 and 16 respectively. Fig. 15 shows the average required bandwidth ranging from 2 up to 32 RF chains. It is shown the high impact of the number of RF chains. For instance, with AMG, the decrease of average required bandwidth between N = 2 and N = 32 is around 17.5 times. In all the simulated range, the improvement of AMG with respect to the other alternatives is notorious.
The average bandwidth versus the number of users, K, is illustrated in Fig. 16. As it is seen, the required bandwidth of all the considered techniques increases as the number of users increases. Nevertheless, the performance of broadcast transmission greatly worsens as the number of users increases, compared to the AMG and unicast approaches. More specifically, the bandwidth reduction of AMG with respect to broadcast transmission is 42% for K = 35, 60% for K = 100, and 82% for K = 200 users. This confirms the observation made in other papers (e.g., [6], [7], [9]) that the performance of multicast precoding deteriorates as the group size grows. This exacerbation of the increase of required bandwidth that happens with broadcast transmission does not happen with AMG, as observed in Fig. 16. This is due to the fact that AMG smartly divides the UE set into smaller groups and thus it benefits from higher multicast beamforming gains, even when the number of users increases. Since AMG searches for a sub-optimal number of groups, the growth in the required bandwidth as K increases is less notorious than in the broadcast case, and thus AMG is even more appealing as K increases.
V. CONCLUSIONS
In this paper a novel grouping and precoding scheme, named AMG, has been proposed. This algorithm relies on a new formulation of the multicast problem that aims at minimizing the required bandwidth, since it is a key metric to increase the operator profits. The implementation aspects of the proposal have been addressed. To this end, its suitability to be integrated in the context 5G NR, using the signaling mechanisms and the available measurements has been discussed. Extensive simulation results have been provided to demonstrate the benefits of the proposal. Hence, the proposal has been assessed in terms of the mean and distribution of the required bandwidth, the complementary distribution of the SNR, the PMF of the optimal number of groups, and the average computation time. Different grouping and multicast precoding algorithms have been compared under different fading and spatial correlation models. Results reveal that the proposed approach reduces the required bandwidth up to 82% for 200 users compared to existing schemes. It is also shown that AMG leads to a greater SNR for a randomly chosen user, and it reduces the variance of the required bandwidth, which eases the implementation in real networks. he was an associate researcher at the University of Málaga, where he participated in contracts with several industry partners related to cellular and satellite communications. He was with Keysight Technologies as a SW R&D engineer from 2017 to 2021, developing 5G cutting-edge technology for real-time communication systems. | 12,520.2 | 2021-01-01T00:00:00.000 | [
"Engineering",
"Computer Science"
] |
Synergistic Effect of Precursor and Interface Engineering Enables High Efficiencies in FAPbI3 Perovskite Solar Cells
Formamidinium lead iodide (FAPbI3)-based perovskite solar cells have gained immense popularity over the last few years within the perovskite research community due to their incredible opto-electronic properties and the record power conversion efficiencies (PCEs) achieved by the solar cells. However, FAPbI3 is vulnerable to phase transitions even at room temperature, which cause structural instability and eventual device failure during operation. We performed post-treatment of the FAPbI3 surface with octyl ammonium iodide (OAI) in order to stabilize the active phase and preserve the crystal structure of FAPbI3. The formation of a 2D perovskite at the interface depends on the stoichiometry of the precursor. By optimizing the precursor stoichiometry and the concentration of OAI, we observe a synergistic effect, which results in improved power conversion efficiencies, reaching the best values of 22% on a glass substrate. Using physical and detailed optical analysis, we verify the presence of the 2D layer on the top of the 3D surface of the perovskite film.
Introduction
Hybrid organic-inorganic perovskite solar cells (PSCs) represented by the chemical formula ABX 3 have been an interesting topic of research in the field of photovoltaics (PV) ever since they were first reported in 2009 with a PCE of 3.8% [1]. In the formula, A represents organic or inorganic cations such as methylammonium (MA + ), cesium (Cs + ), and formamidinium (FA + ), B represents metals such as Pb, Sn, or Ge, and X represents a halide ion (I − , Cl − , or Br − ). With the ease of fabrication, which involves simple processing steps such as spin coating and low-temperature annealing, coupled with the excellent opto-electronic properties of perovskite PV absorbers, rapid advances were made in the upcoming years, and at the moment, the maximum PCE for PSCs has reached an astonishing 26.1% (https://www.nrel.gov/pv/cell-efficiency.html, accessed on 25 July 2023) within a time span of approximately 13 years. For a major part of this time, most of the best PCEs reported in laboratories involved the use of MA + -based structures. Around the year 2020, the PCEs reported for MA + -based solar cells started to plateau, with PCEs not exceeding 22%, and at the same time, the frailties of MA + -based solar cells, such as poor thermal stability, were exposed [2]. Around this time, FA + -based PSCs of the FAPbI 3 type
Fabrication of Solar Cells 2.2.1. Substrate Preparation
FTO substrates were consecutively cleaned using a 2% Hellmanex aqueous solution, milli-Q water, and isopropanol by sonicating for 5 min for each solvent. After drying with a nitrogen gun, the samples were kept in a box.
TiO 2
A precursor solution of tetra ethyl ortho titanate (0.57 g) was dissolved in a mixture of ethanol and hydrochloric acid (5 + 0.1 mL), respectively. The solution was then spincoated on FTO substrates at 2000 rpm for 15 s. Prior to spin coating, the substrates were cleaned with O 2 plasma for 15 min. After spin coating, the wet substrates were sequentially annealed at 200 • C for 10 min and 500 • C for 30 min, respectively.
Mesoporous TiO 2
A precursor solution of meso-TiO 2 was prepared by dissolving 30 NR-D paste in ethanol in a mass ratio of paste: EtOH = 1:6. The solution was then spin-coated on top of the compact layer at 3500 rpm for 15 s. The wet substrates were sequentially annealed at 200 • C for 10 min and 500 • C for 30 min, respectively.
Preparation of the Precursor
For perovskite FAPbI 3 , two kinds of precursor solutions were used: Stoichiometric and non-stoichiometric. A stoichiometric precursor solution (1 FAI:1 PbI 2 ) was prepared by dissolving a mixture of lead iodide (903.2 mg, 1.96 mmol), formamidinium iodide (336.2 mg, 1.96 mmol), and methylammonium chloride (39.04 mg) in a 1 mL mixed solution of DMF and DMSO (DMF (v):DMSO (v) = 4:1) in the glove box and was stirred using a magnetic stirrer without heating. A perovskite obtained from a stoichiometric solution is hereinafter referred to as a PVK (S).
Non-stoichiometric precursor solutions were prepared in the same manner but with a higher PbI 2 content in relation to FAI. We introduced non-stoichiometry in the precursor in such a way that the total mass of the constituents (FAI and PbI 2 ) participating in the reaction remained constant. The perovskite obtained with excess Pb and further denoted by PVK (NS) is referred to as a lead-rich perovskite, which will be justified on the basis of SEM and XPS tests.
Perovskite Layer Fabrication
The perovskite active layer was deposited using an anti-solvent method, with ethyl acetate as the antisolvent. The perovskite precursor solution was deposited on the freshly prepared FTO/c-TiO 2 /mpTiO 2 substrate, and a two-step spin-coating method was applied. The first step proceeded at 1000 rpm with an acceleration rate of 200 rpm/s for 5 s. The second step was followed by 6500 rpm with an acceleration rate of 2000 rpm/s for 25 s. Furthermore, 1000 µL of EA was applied at the 10th second after the spin-coating had ensued. After spin-coating, the substrate was annealed at 150 • C for 10 min to enable the formation of the black-phase FAPbI 3 .
2D Perovskite Fabrication
The surface 2D perovskite was fabricated by the post-treatment of the 3D perovskite with a solution of the 2D reagents, octyl ammonium iodide (OAI) in isopropanol with a concentration of 0.01 M-0.04 M.
The solutions were spin-coated on the surface of the perovskite film for 15 s at 3000 rpm. The full perovskite film was then annealed at 100 • C for 5 min. The procedure was carried out in a glovebox.
Hole-Transporting Layer
Spiro-OMeTAD was used as the hole-transporting layer (HTL) material. First, 73.6 mg of Spiro-OMeTAD powder was dissolved in 1 mL of chlorobenzene. The solution was doped with 17 µL of LiTFSI (prepared by dissolving 520 mg of LiTFSI in 1 mL of acetonitrile) and 30 µL of 4-tertbutylpyridine, respectively. The mixed Spiro-OMeTAD solution was then spin-coated on the surface of the perovskite at 2000 rpm for 30 s.
Top Surface Contacts
The gold electrode was thermally evaporated on the surface of the HTL with the shadow mask, with an area of 0.25 cm 2 . The thickness of the gold electrode was 80 nm, and the evaporation rate was adjusted to 0.01 nm/s for the first 10 nm and 0.08 nm/s for the rest of the procedure.
Current Voltage (I-V) Measurements
Photovoltaic performance measurements were carried out under AM1.5G standard conditions by I-V curve tracing using a Keithley 2401 source meter under simulated AM1.5G irradiation (100 mW cm −2 ). A Photo Emission Tech AAA class solar simulator calibrated against certified reference Si solar cells with a KG-3 filter (Institute Fraunhofer ISE, Breisgau, Germany) was used.
Solar cells were masked to 0.25 cm 2 . I-V measurements were performed in two scan directions, from −0.2 V to 1.5 V, which we call the forward scan, and from 1.5 V to −0.2 V, which we call the reverse scan. The scan rate was set at 100 mV/s.
Scanning Electron Microscopy (SEM)
The topography investigations were performed using scanning electron microscopy (tabletop TM3030, Hitachi High-Tech, Tokyo, Japan) with an accelerating voltage of 10 V in the backscattered electron mode.
UV-Vis-NIR Spectroscopy
The transmittance and reflectance of the perovskite films were measured using an optical spectrophotometer (Lambda 950S, Perkin Elmer, Waltham, MA, USA). The band gap of the perovskite films was calculated by Tauc's plot, which uses the values of absorption coefficient α of the film calculated from the transmittance (T) and reflectance (R) data according to the formula: α = −1/d × ln[(1−R)T]. We used the reflectance data in order to increase accuracy.
Transient Absorption Measurements (TAS)
Ultrafast dynamics were determined using a broad-band transient absorption (TA) setup (Helios spectrometer, Ultrafast Systems, and Spectra Physics laser system), described previously [25]. The IRF (pump-probe cross-correlation function) was approximately 200 fs (full width at half maximum), and transient absorption measurements were performed in the time range of up to 3 ns. One excitation wavelength was used, 490 nm, and the spectra were probed in the range of 500-830 nm. Transient absorption spectra were analyzed with Surface Xplorer 2.4.3.153 Software (Ultrafast System).
X-ray Photoelectron Spectroscopy (XPS)
The X-ray photoelectron spectra (XPS) were recorded using the hemispherical analyzer EA 15 (PREVAC, Rogow, Poland) equipped with the dual anode X-ray source RS 40B1 (PREVAC). The measurements were performed using Al Kα (1486.6 eV) radiation and an analyzer pass energy of 100 eV. The spectra were recorded in normal emission geometry with an energy resolution of 0.9 eV. The spectrometer was calibrated with Ag, Au, and Cu foil according to ISO 15472:2010 standards [26]. An ultra-high vacuum (UHV) of 8·10 −9 mbar was maintained during the measurements. The analyzed area was approximately 3 mm 2 , and the penetration depth was approximately 10 nm.
The perovskite samples were mounted and positioned at the dedicated holder and pumped out to a high vacuum and then transferred into the UHV chamber. The survey and high-resolution spectra were acquired for every sample. The spectra were analyzed using the analysis software CasaXPS 2.3.24PR. The electron binding energy (BE) scale was calibrated for the Fermi edge at 0.0 eV. The Shirley-type spectrum background was used. The high-resolution spectra were deconvoluted with the Voigt function (Gaussian to Lorentzian profile ratio of 70:30). The spectra were compared respectively to the background level.
X-ray Diffraction (XRD)
The XRD measurements of the perovskite films were carried out using a Bruker D8 Discover diffractometer equipped with a Cu Kα X-ray source. Bragg-Brentano (Θ-2Θ) X-ray diffraction continuous scans were performed over the range of 2Θ = 10-100 • at 2 s per step with a step size of 0.02 • . The phase composition of the perovskite films was analyzed using Diffrac.EVA v.3.0 software with the ICDD PDF-4+ crystallographic database.
Atomic Force Microscopy (AFM)
Kelvin Probe Force Microscopy (KPFM) experiments for the standard perovskite solar half-cells at room temperature (22 • C) and humidity below 44% were performed using the Dimension3100 setup with the NanoscopeV controller. The KPFM mode was used in order to determine the electrical response on the top of 3D and 2D perovskite layers at the microscale. To obtain information related to the surface potential, a tip (Type: Arrow™ EFM) with PtIr thin layer was used. The measurement was performed in twopass, lift mode (that is, the tip of the AFM was approximately 120 nm above the surface of perovskite solar half-cells) in air. The obtained data were processed using Gwyddion software 2.60 [27]. KPFM measurements in the dark and under irradiation were conducted. For the illumination of working perovskite solar half-cells, the following light source was used: A white cold LED COB with electrical power of 10 W, a viewing angle of 140 • , a color temperature of 6500 K, and a luminosity of 850 lm. In Figure S1, a schematic representation of a KPFM measurement setup is shown (see Supporting Information).
In order to understand the significant effect on electrical properties at the microscale of a 2D perovskite (PVK) layer spread on top of a 3D PVK, atomic force microscopy studies have been carried out in the dark and under light conditions. Based on Kelvin probe force microscopy (KPFM) experiments, changes in the surface potential (SP) in different light conditions and for stoichiometric and lead-rich perovskite 3D PVK with and without 2D passivation layers have been estimated.
Obtaining the Right Precursor Stoichiometry for Obtaining High-Efficiency Solar Cells
As a starting point, we fabricated 3D FAPbI 3 solar cells by combining 1 molar mass (1M) of the constituents. The reaction leading to the formation of the 3D perovskite can be written as follows: 1M FAI + 1M PbI 2 → 1M FAPbI 3 (1) The 3D perovskite films were processed following the optimized procedure explained in the previous section. The solar cells were fabricated in the n-i-p configuration. The following sequence of layers was used: Glass-fluorine doped tin oxide (FTO)/compact titanium oxide (c-TiO 2 )/mesoporous (m-TiO 2 )/3D-FAPbI 3 perovskite/2D-perovskite/spiro-OMeTAD/Au. The current-voltage characteristics of the corresponding PSCs (the best cells) are shown in Figure 1a. The solar cells prepared with stoichiometric precursors show a PCE of 9.6%, which is very different from many published works. In order to prevent the formation of the δ phase, we deposited a thin film of 2D perovskite made with OAI (10 mM dissolved in 1 mL IPA) on the surface of the 3D perovskite film. Even after modifying the interface, the PCE of the solar cell did not improve but rather degraded to 7.5%. The photovoltaic parameters of the solar cells are shown in Table 1.
Obtaining the Right Precursor Stoichiometry for Obtaining High-Efficiency Solar Cells
As a starting point, we fabricated 3D FAPbI3 solar cells by combining 1 molar mass (1M) of the constituents. The reaction leading to the formation of the 3D perovskite can be written as follows: The 3D perovskite films were processed following the optimized procedure explained in the previous section. The solar cells were fabricated in the n-i-p configuration. The following sequence of layers was used: Glass-fluorine doped tin oxide (FTO)/compact titanium oxide (c-TiO2)/mesoporous (m-TiO2)/3D-FAPbI3 perovskite/2D-perovskite/spiro-OMeTAD/Au. The current-voltage characteristics of the corresponding PSCs (the best cells) are shown in Figure 1a. The solar cells prepared with stoichiometric precursors show a PCE of 9.6%, which is very different from many published works. In order to prevent the formation of the δ phase, we deposited a thin film of 2D perovskite made with OAI (10 mM dissolved in 1 mL IPA) on the surface of the 3D perovskite film. Even after modifying the interface, the PCE of the solar cell did not improve but rather degraded to 7.5%. The photovoltaic parameters of the solar cells are shown in Table 1. When we compare the photovoltaic performance of the two solar cells, the PSC made with the 3D/OAI film appears more resistive, resulting in a decreased J sc and FF. We suspect that this resistive behavior comes from the addition of OAI to the surface of the film. The insulating nature of the long alkyl chain could have contributed additional series resistance to the solar cell, thereby decreasing its efficiency. This suggests that a 2D perovskite film may not have formed on the surface of FAPbI 3 as per expectations, and instead, we merely deposited a thin layer of OAI and partially insulated the film's surface.
In Figure 1b, we show the morphology of the bare perovskite film made with the stoichiometric precursor. The perovskite film has a dense microstructure with large, pinholefree grains on the surface. The presence of the cubic α phase is confirmed by the XRD pattern of the perovskite film shown in Figure 1c. The characteristic peaks of the black phase have been indexed using the PDF card 00-069-0999 from the ICDD PDF 4 database. Apart from FAPbI 3 , there is a very small amount of PbI 2 in the film, as indicated by the peaks at 14.72 • and 45.2 • (indexed by the PDF card no: 00-007-0235) and reflections from the back contact, FTO (PDF card no: 04-003-0649). Both the bare perovskite film and the film with OAI exhibit the same diffraction pattern, but the intensity of the FAPbI 3 peaks in the 3D/OAI film is diminished considerably. We attribute this effect to the presence of a thin amorphous layer at the surface formed by OAI. We calculated the band gap of the perovskite films using Tauc's plot, which shows a sharp absorption onset around 800 nm, corresponding to a band gap of approximately 1.54 eV for both films (Figure 1d).
In order to gain more understanding of the 3D perovskite (S) films and possible causes responsible for the poor photovoltaic performance of the PSCs fabricated with the perovskite film 3D PVK (S) and 3D/OAI, we tested the films using KPFM. To begin with, using KPFM, we can measure the surface potential (SP) of the films in the dark and under illumination. The difference in the values of SP between dark and illumination gives us the surface photovoltage (SPV) generated in the perovskite half-solar cells. In this scenario, the terminology 'perovskite half-solar cell' is analogous to 'perovskite film' since the film contains one charge transport layer and one electrode. The generation of (SPV) can be correlated with the band bending taking place in the PSCs under illumination. The results of the KPFM measurements are shown in Figure S2 (Supporting Information). In Table S1, we summarize the information obtained from AFM and KPFM measurements.
The small fluctuations in the obtained values of SP are expected phenomena and are related to the intensity of the light source used, the exact area of illumination, and the architecture of a perovskite half-solar cell. In our case, the perovskite solar semi-cells were illuminated from above and from the side of the 3D/2D active layer and not from the semi-transparent conductive electrode as shown in Figure S1. This leads to small changes in the SP values between dark and light conditions [28]. Moreover, it is important to keep in mind that the measurements have been carried out on perovskite films without the HTL. In the absence of the HTL, there is less efficient charge separation, which is reflected in weaker SP fluctuations between dark and light measurements. Fluctuations in the measured values of the work function from the top of such a thin film stack, which indicate the absence of a charge transporting layer, which leads to a reduction in the difference of potential between dark and light experiments [29,30].
For the perovskite 3D film from stoichiometric solution PVK (S) without OAI, an increase in SP after illumination was observed. In contrast, the SP decreases under illumination when an organic layer is applied. This could imply the presence of an insulating component, which restricts the band bending in the photoactive layer.
The empirical formula for the most common 2D (Ruddlesden-Popper) perovskites is given by the empirical formula, (LC) 2 (SC) n−1 Pb n I 3n+1 where LC is a large cation such as OA + and SC is a small cation such as FA + , and n is the number of (PbI 6 ) 4− octahedral units sandwiched between the large cations. The lowest value n can have is 1, which corresponds to a pure 2D perovskite with the formula (OAI) 2 PbI 4 . This means that in order to form a 2D perovskite layer, we need PbI 2 along with OAI. Under stoichiometric conditions, the reaction leading to the formation of the FAPbI 3 black phase after spin coating and annealing will consume all the available PbI 2 in the precursor, leaving nothing behind. Therefore, when using a stoichiometric composition, there is no additional PbI 2 to form a 2D perovskite layer, and depositing OAI on the surface of the 3D film only forms a thin insulating film, which degrades the solar cell's performance even further.
In order to have additional PbI 2 to form a 2D perovskite layer, we intentionally make the perovskite precursors non-stoichiometric. The composition of non-stoichiometric precursors is (1 − x) FAI:(1 + x) PbI 2 where the amount of 'x' is varied between 0.05 and 0.15. We then fabricated PSCs as before. The spread of the PCEs in the PSCs made with different x values along with the IV curve of the best cell is shown in Figure S3 (Supporting Information). We observe the highest efficiency of 17% for the condition x = 0.1. All the photovoltaic parameters (J sc , V oc− , and FF) show a significant improvement compared to the PSC from the stoichiometric solution.
In Figure 2a, we show the morphology of the perovskite film embodying the best solar cell. Similar to the perovskite film fabricated from a stoichiometric solution, the morphology of the perovskite film made from the non-stoichiometric solution is also dense and compact. Along with the big grains of perovskite, we can also see small, bright particles scattered on the surface. These bright particles can be associated with unreacted PbI 2 (as elements with heavier atomic numbers appear brighter in the backscatter electron detection mode). Similar examples can be found in several other papers, which discuss 3D perovskites [31,32]. The crystallographic features of the Pb-rich perovskite film are shown in Figure 2b. Similar to the perovskite film made from the stoichiometric solution, the reflections in the XRD pattern of this film are also matched by FAPbI 3 , PbI 2 , and FTO using the same references. However, the main difference in this XRD pattern comes from the intensity of PbI 2 . Due to the presence of excess PbI 2 present in the films, the intensity of the peak at 14.6 • is very high in this film. From Tauc's plot shown in Figure 2c, we calculate the band gap of the film to be 1.54 eV, which is the same as the stoichiometric perovskite film, suggesting that small changes induced in the stoichiometry of the precursor do not influence the band gap of the absorber. At this point, we are still unable to answer why the photovoltaic performance of the PSCs fabricated using a stoichiometric precursor has poor efficiency. From the results we have thus far, we observe that the bulk properties of the perovskite films (microstructure, crystallinity, and band gap) are essentially the same irrespective of the precursor composition. In the literature, we often find a Pb excess composition being used for high-efficiency solar cells. There are many reports that support this claim [33][34][35]. Our results also In Figure 2d SP after illumination is observed compared to the SP measured in the dark. This implies the generation of SPV in the perovskite half solar cell, which leads to changes in the work function [30]. We also notice that the magnitude of SP in the PVK (NS) film is higher when compared to the SP of the film from the stoichiometric solution PVK (S). Comparing the photovoltaic performance of the PSC made from different precursors, we are inclined to believe that perovskite films made under Pb-rich conditions are less defective or more defect-tolerant than the stoichiometric films. A higher SPV in the PVK (NS) films suggests optimal band bending at the perovskite/HTL interface, leading to the efficient transport of carriers at the interface and better photovoltaic performance.
At this point, we are still unable to answer why the photovoltaic performance of the PSCs fabricated using a stoichiometric precursor has poor efficiency. From the results we have thus far, we observe that the bulk properties of the perovskite films (microstructure, crystallinity, and band gap) are essentially the same irrespective of the precursor composition. In the literature, we often find a Pb excess composition being used for high-efficiency solar cells. There are many reports that support this claim [33][34][35]. Our results also favor the same observation. However, in order to look for specific differences among the perovskite films that could explain the photovoltaic performance of the PSCs presented till now, we studied the surface of the bare perovskite films using XPS.
With XPS, we measured the total concentration of the elements at the sample surface (Survey spectra shown in Figure S4) and the deconvoluted absolute spectrum of every element measured in the survey spectra. Table S2 (Supporting Information) shows the percentage of all the elements measured in the survey spectra. The XPS survey spectra are sensitive to the surface composition of the perovskite films. With the given beam energy, information about the composition is obtained from the first 5-10 nm of the surface. Since the films differ from each other in the amount of PbI 2 , we first looked at the differences in the Pb 4f spectra and I 3d from the survey spectra for the different perovskite films.
The film '3D PVK (NS)' made from a non-stoichiometric precursor is Pb rich on the surface due to the presence of excess PbI 2 in the film in comparison with both stoichiometric films. The increase in the I signal for the 3D/OAI film is due to the presence of OAI at the surface of the film. In the same film, we observe small concentrations of O 2 and N 2 (shown in Table S2 in Supporting Information) at the surface. This is due to the hydrophobic nature of the long alkyl chain in OAI. The amount of Iodine measured from the surface of 3D PVK (S) is higher than that of sample 3D PVK (NS). The total mass of Iodides taking part in the formation reaction of the perovskite is the same in all three films. Therefore, this observation suggests that changing the precursor composition could actually result in perovskite films with different surface compositions or, in other words, different ways in which the surface of the FAPbI 3 perovskite film terminates.
We analyzed the absolute spectra of each element in the survey spectra in detail. The chemical states are assigned to certain peaks according to the databases [36,37]. For this discussion, we use the results from the absolute spectra of Carbon (C), Nitrogen (N), and Oxygen (O).
When we deconvolute the absolute 1s spectra of carbon (C) as shown in Figure S6 (Supporting Information), we see major differences in the peak at 286.5 eV. This peak corresponds to the bonding of C with Nitrogen (N) in the form of a C-N single bond. This bond is characteristic of amine (C-NH 2 ) or amide (NH 2 -C=O) groups. Amide groups are present in the precursor solvent DMF. DMF has a boiling point of 153 • C. We use DMSO alongside DMF in a 4:1 ratio. DMSO has a boiling point of 189 • C. The annealing temperature of the perovskite film is 150 • C. During the annealing step, rapid crystallization of the perovskite takes place, forming dense grains within the first few seconds before all of the solvents present in the wet film could fully evaporate. DMF offers strong polarity and strong intermolecular forces, which makes solvent evaporation a difficult task. As a result, it is very likely that a fraction of the solvent present in the wet film is trapped between the grains. Residues of DMF and DMSO are regularly present in perovskite films after conventional annealing [38][39][40][41]. Therefore, the primary contribution of C=O in these films is expected to come from residual DMF present in the film after annealing. In Table 2 below, we show the proportion of C=O bonds calculated from both C 1s spectra and O 1s spectra. The amount of C=O bonds is almost the same in both perovskite films from stoichiometric PVK (S) and non-stoichiometric solution PVK (NS) without the OAI layer. In the film 3D/2D PVK (S), the % of C=O is lower than in the other two films for two reasons: One is due to the presence of OAI, which covers the surface of the 3D perovskite film, limiting the depth of penetration of the probing beam, and the second is due to the low amount of O 2 adsorbed on the surface of the film, as seen in Table 2. The C-NH 2 single bond is present in both FAI and DMF. However, the difference in the peak proportions among the analyzed samples indicates that they could be influenced by the presence of FA + ions on the surface of the perovskite films. It is possible for MA + ions to contribute to this bonding because the precursor utilizes 35% MACl in the mixture. The fact that our absorbers exhibit a band gap of 1.54 eV as opposed to the widely reported 1.47 eV forces us to think that some MA + could be substituted for FA + , leading to a widening of the band gap. We believe that the adsorption of N 2 from the atmosphere did not contribute to this bond because the total amount of N 2 measured from the surface of the lead-rich perovskite film is lesser than that of the stoichiometric film.
We make the non-stoichiometric perovskite precursor FAI deficient intentionally. It contains approximately 12 wt% FAI less compared to the stoichiometric film. In spite of this considerable difference, the surface of the bare perovskite film is rich in C-N signature bonds (Table 3), which suggests that the surface of this perovskite film is possibly terminated by FA + ions. Referring to the work of Oner et al. [42], we find that in FAPbI 3 films prepared under Pb-rich conditions, surface termination of the bulk FAPbI 3 by FA + ions is very much a possibility. It is the most suitable surface considering the formation energies of defects and the surface energy. The FA + -terminated surface has the lowest surface energy and, therefore, offers the most stable surface of all the possibilities. For the same criteria, I − termination is found to be the most vulnerable surface against all types of defect formations, as well as one of the most energetically unfavorable surfaces. On the I − terminated surface, the formation energies of the possible defects are quite low, and hence they can be created easily after the film has formed. Such defects are likely to create deep transition levels in the band gap, thereby possibly degrading the photovoltaic performance. Table 3. Table showing the percentage of C-N bonds from the absolute spectra of C and N for the different perovskite films. Taking into account the results from the survey spectra, we see that a perovskite film obtained from a stoichiometric solution, PVK (S), is I − rich in comparison and would likely result in an I − terminated surface. An over-stoichiometric amount of PbI 2 would result in a Pb-rich surface due to excess PbI 2 that precipitates at the end of the reaction, but the resulting perovskite film is possibly surface terminated by FA + ions. From our findings, it is clear that the precursor stoichiometry heavily influences the surface composition of the perovskite and thereby the nature of the surface termination, which in turn affects the performance of the solar cells.
Optimization of the Concentration of the Large Cation to Improve the PCE
In order to ensure a stable solar cell over a long period of time, it is necessary to preserve the α phase and prevent the phase transition of FAPbI 3 . Therefore, we deposited OAI on the surface of the lead-rich perovskite film.
We varied the concentration of OAI from 10 mM to 40 mM, and by doing so, we varied the thickness of the 2D perovskite film, as the film thickness strongly depends on the concentration of the large cation. In Figure 3a-d, we show the morphology of the perovskite films containing different thicknesses of the OAI layer. In comparison with bare lead-rich perovskite film without the 2D layer, the perovskite films containing the 2D layer show an improvement in the grain size and display a more uniform morphology. When the surface of FAPbI 3 is treated with a solution of IPA containing OAI and subsequently annealed, recrystallization of the 3D film takes place, resulting in more uniformly shaped grains [43,44]. We also notice that as the concentration of OAI increases, the amount of PbI 2 present on the surface of the film decreases gradually (fewer bright particles), and when the concentration of OAI reaches 40 mM, we no longer observe the bright PbI 2 particles in the image. At a concentration of 40 mM, more OAI is available to react with the excess PbI 2 present on the surface, thereby forming a thick 2D perovskite layer on top of the 3D perovskite film. We also notice that perovskite films containing thicker 2D perovskite layers appear brighter than films without the 2D layer. This is because the 2D perovskite has lower conductivity than PbI 2 and FAPbI 3 , and as a result, charge accumulation takes place at the surface, making it look relatively bright [45]. consumption of uncoordinated PbI2 from the surface by the large cation also results in the formation of a graded 3D/2D heterojunction, which exhibits a different energy level at the interface and constructs a built-in electric field (Vbi), facilitating charge extraction [48], with optimal band alignment facilitating carrier transport across the junction, resulting in an improved fill factor. In Figure S8, we show the topography of the 3D/2D perovskite films along with the color plots showing the distribution of the SP from KPFM measurements. We estimated the SPV for the films containing OAI by taking into account the SP measured in the dark and under illumination. The values of the SPV for the non-stoichiometric perovskite film containing different concentrations of OAI (10 mM, 20 mM, 30 mM, and 40 mM) are 25 mV, 57 mV, 30 mV, and 13 mV, respectively. Comparing the SPV of the control sample (without OAI, which is 85 mV) with that of the perovskite films with the OAI layer, it is quite clear that the SPV decreases with an increase in the concentration of OAI, or in other words, with the formation of a 2D perovskite at the interface. A similar observation was reported in the work of Zhang et al. [30]. Taking this published work into consideration, we can postulate that with the incorporation of OAI, there is a formation of a thin 2D In Figure 3e-h, we show the JV curves in the forward scan for the corresponding champion PSCs. The PSCs with the 2D perovskite layer display an improvement in the PCE in comparison with the PSCs made without the 2D layer, primarily due to the increase in the open circuit voltage of the devices. The formation of a thin layer of 2D perovskite at the surface acts as a capping layer, protecting the film from ambient conditions. Additionally, the ammonium atoms in OAI are capable of filling in for defects [46,47] on the surface of the film, thereby minimizing the recombination losses in the absorber. As a result, the V oc of the PSC's increases contributed to an increase in the overall PCE. The consumption of uncoordinated PbI 2 from the surface by the large cation also results in the formation of a graded 3D/2D heterojunction, which exhibits a different energy level at the interface and constructs a built-in electric field (V bi ), facilitating charge extraction [48], with optimal band alignment facilitating carrier transport across the junction, resulting in an improved fill factor.
In Figure S8, we show the topography of the 3D/2D perovskite films along with the color plots showing the distribution of the SP from KPFM measurements. We estimated the SPV for the films containing OAI by taking into account the SP measured in the dark and under illumination. The values of the SPV for the non-stoichiometric perovskite film containing different concentrations of OAI (10 mM, 20 mM, 30 mM, and 40 mM) are 25 mV, 57 mV, 30 mV, and 13 mV, respectively. Comparing the SPV of the control sample (without OAI, which is 85 mV) with that of the perovskite films with the OAI layer, it is quite clear that the SPV decreases with an increase in the concentration of OAI, or in other words, with the formation of a 2D perovskite at the interface. A similar observation was reported in the work of Zhang et al. [30]. Taking this published work into consideration, we can postulate that with the incorporation of OAI, there is a formation of a thin 2D perovskite layer on the surface, which minimizes the density of the acceptor trapping states localized on the surface of 3D perovskite active layer. This possibly causes a decrease in the magnitude of the trap states on the surface [30], which leads to an improvement in the photovoltaic performance of the PSCs with the 2D perovskite layer. Overall, from KPFM, we predict a reduction of defects at the surface of the film with the formation of a 2D perovskite film. This finding is consistent with the improvement in the IV parameters of the PSCs.
In Figure 4a, we show the XRD patterns of the perovskite films with different thicknesses of the 2D perovskite layer along with the lead-rich film without the 2D perovskite layer. As shown previously, the α phase is present in all the films. As the concentration of the large cation is increased from 10 mM, we see that the intensity of the main PbI 2 peak starts to decrease, reaching a bare minimum for a concentration of 40 mM. This indicates the consumption of PbI 2 with increasing OAI concentration, thereby forming thicker 2D perovskite layers, as we pointed out in Figure 3. The decrease in the intensity of PbI 2 is accompanied by an increase in the intensity of the perovskite peak at 16.2 • . This is due to the formation of a more ordered surface as a result of the recrystallization of the 3D film during the surface OAI treatment. The band gap of the perovskite films was calculated from Tauc's plot shown in Figure 5b. In the presence of a 2D perovskite, the absorption of photons in the measured wavelength range has considerably increased. Part of this enhancement comes from the absorption of 2D perovskite, which has a band gap in the region of approximately 2.4 eV [49] as opposed to the band gap of the FAPbI 3 , which has a band gap of 1.54 eV in all cases. However, this increase in absorption does not necessarily translate to an increase in the J sc of the solar cells, as one would expect. Only for the perovskite films that contain thick 2D layers (30 mM and 40 mM) do the resulting solar cells show an increase in J sc of approximately 1 mA/cm 2 compared to the rest of the solar cells. The presence of 2D layers does not change the band gap of the 3D FAPbI 3 perovskite films; rather, the 2D perovskite merely increases the absorbance in the high-energy region of the spectrum. Such phenomena have also been previously reported [50]. We performed a detailed optical analysis on all the perovskite films discussed s using TAS. We first show the TAS spectra of the stoichiometric perovskite film and film post-treated with OAI in Figure 5. The measurements were performed with bot probe and pump exciting beams entering from the perovskite side. The spectra for films show a photobleaching peak around 780 nm, which corresponds to the black p For the n = 1 two-dimensional perovskites made with OAI, PB peaks are expected seen around 506 nm (2.45 eV) [49]. In the case of the OAI-treated film, there are n peaks in this region, which confirms the absence of any 2D perovskite in the film. I We performed a detailed optical analysis on all the perovskite films discussed so far using TAS. We first show the TAS spectra of the stoichiometric perovskite film and the film post-treated with OAI in Figure 5. The measurements were performed with both the probe and pump exciting beams entering from the perovskite side. The spectra for both films show a photobleaching peak around 780 nm, which corresponds to the black phase. For the n = 1 two-dimensional perovskites made with OAI, PB peaks are expected to be seen around 506 nm (2.45 eV) [49]. In the case of the OAI-treated film, there are no PB peaks in this region, which confirms the absence of any 2D perovskite in the film. In the set of lead-rich perovskite films, only an OAI concentration of 40 mM was bleached at 540 nm observed, as can be seen in Figure 5b. For the reference lead-rich film without OAI, there is a positive transient absorption signal in this spectral range instead (originating from the 3D phase), while for an OAI concentration of 10 mM, we can observe a signal in between, around zero, which might indicate the small presence of a 2D phase but without clear bleaching. For two other OAI concentrations (20 and 30 mM), we were not able to measure reliable TAS spectra below 600 nm, likely due to the too-high stationary absorbance there and/or too-low concentration of the 2D phase. FAPbI 3 has a very high optical absorption in the region of 550-800 nm as seen from Tauc's plot, and combined with the high thickness of the layer (approximately 750 nm as shown in the color-marked cross-section TEM image in Figure S9), most of the output signal was dominated by FAPbI 3 , and it was not possible to obtain any signal from the thin layer of 2D perovskite, which has ground-state bleaching in the low-energy part of the spectrum. Only when the thickness of the 2D phase was considerably higher (as in 40 mM PVK film), the ground-state bleaching could be recorded. We expect this signal to come from the low-dimensional n = 1, 2D perovskite [49]. However, the position of this bleachinig is shifted to a higher wavelength as opposed to the band gap of n = 1, 2D perovskite, which is approximately 506 nm (2.45 eV). This indicates that the signal could originate from higher-order 2D perovskites (like n = 2). When we look back at the formula for 2D perovskites mentioned earlier in the manuscript, it is clear that small cations such as FA + or MA + are needed to form higher-order perovskites such as n = 2, 3, etc. In the lead-rich perovskite films, we have seen that the surface is terminated by FA + ions, as evidenced by XPS. During the annealing step of the post-treatment with OAI, it is very likely that some of these FA + ions on the surface would have reacted with OAI and PbI 2 , forming a higher-order 2D perovskite such as (OA) 2 FAPb 2 I 7. The band gap of 2D perovskites increases with increasing n. Therefore, a band gap of 540 nm (2.3 eV) would correspond to a 2D perovskite with n = 2 [49]. The formation of such higher-order 2D perovskites at the interface of 3D FAPbI 3 has been reported previously [16]. The overall optical analysis by TAS correlated with the published literature reveals the exact composition of the 2D perovskite layer formed during the post-treatment with OAI.
We share more information on the dynamics of the charge carriers by comparing the TA response (∆A) of the photo-bleached peaks as a function of the pump-probe delay time ( Figure S10). We probe the kinetics at 550 nm (within the 2D phase bleach band) and around 780 nm (3D phase bleach). For stoichiometric samples, no negative amplitude due to 2D phases can be seen ( Figure S10c), and the kinetics of the bulk bleaching are similar for the samples with and without OAI ( Figure S10a). In contrast, for the lead-rich sample with an OAI concentration of 40 mM, we can clearly see 2D-phase bleaching (negative ∆A) kinetics that decay on the time scale of single and tens of ps ( Figure S10d). The decay may be due to the direct hole transfer from the 2D phase to spiro-OMeTAD and/or energy or the electron transfer to the 3D phase (e.g., [32]). The kinetics at 550 nm of the sample with 10 mM of OAI are approximately zero, which might be due to the overlap of the positive signal of the 3D phase (the sample without OAI) and the negative contribution of 2D phases ( Figure S10d). However, likely more convincing proof of the presence of the 2D phase comes from the analysis of 3D phase kinetics (at 775 nm) of non-stoichiometric samples ( Figure S10b). We can expect a delayed partial population of the 3D phase from the 2D phase, and this is indeed what is seen for the samples with 10 mM and 40 mM of OAI: The bleaching amplitude becomes larger, and the minimum is shifted towards longer times with respect to the reference lead-rich film without OAI.
Conclusions
We have successfully fabricated highly efficient FAPbI 3 PSCs with the best PCE of 22% on glass substrates. The main steps behind the fabrication of the perovskite involve adjusting the stoichiometry of the precursors (FAI and PbI 2 ), followed by optimization of the concentration of the large cation, which is needed to form a 2D perovskite layer. The primary difference between a stoichiometric perovskite film and a Pb + excess in lead-rich perovskite film is not strictly limited only to the amount of excess PbI 2 present in the films at the end of annealing but also extends to the nature of surface termination in these films. A Pb + excess composition offers a stable and defect-tolerant surface, which eventually results in improved photovoltaic performance of the PSCs compared to a stoichiometric composition. However, the presence of uncoordinated PbI 2 on the surface of the films limits the PCE considerably. Performing a post-surface treatment with OAI on the lead-rich perovskite films allows OAI to react with the excess PbI 2 already present in the film to form a 2D perovskite layer on the surface. The post-treatment also results in the recrystallization of the 3D-FAPbI 3 perovskite surface. The 3D/2D perovskite heterojunction in the posttreated films passivates the perovskite surface and enhances the optical absorption and crystallinity of the perovskite films. All these effects synergistically work together and result in an improvement in the PCE of the respective solar cells, with maximum values reached for an OAI concentration of 40 mM.
Supplementary Materials: The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/ma16155352/s1, Figure S1: Schematic representation of KPFM measurement setup; Figure S2: Topography and surface potential measured by atomic force microscopy methods; Table S1: Average surface roughness measured from AFM images and surface potential values of the perovskite films 3D PVK (S) and 3D/OAI; Figure S3: (a) Box plot showing the distribution of PCEs and (b) current-voltage characteristics of the best PSC made with a non-stoichiometry of 0.1; Figure S4: Survey spectra measured by XPS; Table S2: Table showing the percentage of elements from the surface of the different the perovskites measured by the XPS survey spectra; Figure S5: Absolute and deconvoluted spectra of the core level of C (1s) measured by XPS; Table S3: Table showing the proportion of the different bonds formed by C with other elements from the absolute spectra; Figure S6: Absolute and deconvoluted spectra of the core level of O (1s) measured by XPS; Table S4: Table showing the proportion of the different bonds formed by O with other elements from the absolute spectra; Figure S7: Absolute and deconvoluted spectra of core level of N (1s) measured by XPS; Table S5: Table showing the proportion of the different bonds formed by N with other elements from the absolute spectra; Figure S8: Topography of the perovskite films prepared from a nonstoichiometric precursor measured by atomic force microscopy and distribution of surface potential measured from the surface; Table S6: Average surface roughness measured from AFM images and surface potential values of the perovskite films PVK (NS) with different concentrations of OAI measured in dark and under white-light illumination; Figure S9 | 10,985.8 | 2023-07-30T00:00:00.000 | [
"Engineering",
"Materials Science",
"Physics"
] |
State of Information and Communication Technology (ICT) Usage in the Training of Construction Students in Ghanaian Technical Universities
In recent years, the construction sector has undergone impressive developments in the use of ICT. Thus, placing much demands on capacity building that will produce an adequate, highly skilled workforce to manage the growing technology in the industry. ICT takes the front role and essentially becomes the medium in which all the other operations are carried out. Most researchers, thinkers and educators have taken up the challenge of using ICT since the 1980s with varied successes. For this reason, this paper assesses the extent to which ICT is used in the training of construction students in technical universities in Ghana. Quantitative method (questionnaire) was employed to elicit data from construction lecturers and students in five technical universities in Ghana. Descriptive statistics and factor analysis were adopted in analyzing the data. The research discovered that ICT adoption in training construction students in Ghanaian technical universities is very low. The study also established twenty-four (24) factors as constraints to ICT usage in construction students' training in technical universities in Ghana. These come under "technology constraints", "economic constraints", "human constraints", "environmental constraints," and "administrative constraints". There is, therefore, a pressing need for a drastic shift towards integrating Construction related ICT tools in training construction students in Ghanaian technical universities.
Introduction
Education generally determines the level of advancement of a nation, society and the individual. Appropriate Education makes way for appropriate technological development (Desai, 2010). According to the World Bank (1998), tertiary education must be the vanguard for this transformation, and Information Communication Technology (ICT) must be promoted as a primary training facility. Danso (2012) also reiterated that tertiary institutions must encourage highly appropriate courses that represent the skills needed in today's working world.
Polytechnics in Ghana were converted to Technical Universities in 2016. They have a significant role in the country's required skilled labor (Aidoo-Tailor, 2009;Owusu-Agyeman & Osterkamp, 2009). For this reason, all technical universities must ensure that they adequately train their graduates to meet the demands of the job market. With ICT appearing in all areas of development, including the construction industry, construction education in technical universities cannot underestimate ICT in their training. The Skills Development Fund (2011) affirmed that Ghana's developmental success must be powered by skills, knowledge and innovation in the construction sector. This vision can be accomplished where construction educational institutions are committed to training characterized by the capacity to innovate, compete and thrive to ensure the socio-economic development of the nation. Increasingly, society advocates for technology-savvy workers to enhance creativity and efficiency (Nwosu & Ogbomo, 2011). This can be achieved by ascertaining dependable information sources and finding the best means of accessing this information for use (Nwosu & Ogbomo, 2011).
In recent years, the construction sector has experienced impressive developments in the use of ICT. (Loosemore et al., 2003). Therefore, Construction training should concentrate on creating sustainable information and professional workforce in Technology adoption to enhance efficiency and productivity within the construction industry. Most workers worldwide have taken up the challenge of using ICT since the 1980s with varied successes.
Ghana's construction sector must follow the actions of the developed countries by integrating ICT into construction activities. The adoption of ICT in the construction sector in Ghana will help to strengthen the existing limitations of the sector (Fugar, Ashiboe-Mensah, & Adinyira, 2013). Therefore, the education sector, particularly technical universities providing construction education, have a role to play.
Construction education is essential for nation-building in industrialized and emerging economies. The quality of construction professionals affects the delivery of construction projects. It is, therefore, critical to build the capacity of the construction workforce for efficient and effective project delivery and to also ensure the success of the sector. The construction sector in Ghana is low in technology adoption and more manual driven as compared to other sectors (Fugar, Ashiboe-Mensah, and Adinyira, 2013). Therefore, the continuous delivery of the obligation of the construction industry depends on its human capital development (Ofori, 2012).
Evidence abounds on the usage of computers and ICT in Construction estimating, construction planning, surveying, procurement, electronic tendering and robotics in construction intelligence building elsewhere in the world (Harris & McCaffer, 2001). Despite the forgoing, not much is encountered in the literature regarding empirically documented works on the state of ICT on construction education in Ghana. It is against this backdrop that this research sought to evaluate the state of ICT usage in training construction students in Technical Universities in Ghana.
ICT in construction education
ICTs in construction training require the implementation of general ICT elements in the process of teaching and learning. Globalization and international transformation have created a modern, technologically based, information-driven, knowledge-driven global economy. This current world economy has major consequences for the existence and function of academic institutions. ICT has transformed conventional construction approaches and doings, offering new dimensions for teachers, learners, and researchers. Prospective Technical University graduates ought to be compliant with the ever-expanding knowledge and trained with the requisite technologies.
The involvement of ICT in teaching and learning plays a critical role, particularly in empowering technology in learning programs (Kaka, 2008). ICT is considered the center of education due to its presence in mainstream and unique settings and numerous educational establishments. Numerous scholars contend that this era requires selfconfidence and efficiency in the use of ICT in all disciplines, both at industry and academia levels, to make learners productive. However, there is several roles ICT play in equipping students to function well in their world of work. ICT strives to foster a culture of diversity and cooperation, building on individual talent and unique capacities, introducing 'new worlds' to instructors exploring innovative avenues to support learners move beyond the classroom, school, and society. In Africa, evolving electronic messaging offers more guidance and opportunities for different styles and levels of cooperation between different institutions (Nwoso & Ogbomo, 2011). ICTs create interactive learning and power-sharing and influence that encourages students to study according to their cultural style. In construction education in the West, Students are responsible for researching and being open-minded about these new ICT dealings. The technology revolution gives Africa a remarkable opportunity to leapfrog into the future from centuries of sluggish growth. Therefore, to prevent marginalization, African academics and universities must seize the opportunity to implement communication technologies.
Methodology
The quantitative research design (questionnaire survey) was adopted for the study. The research population consisted of final-year students pursuing Building Technology and Civil Engineering programs and lecturers conducting these programs at the eight technical universities in Ghana. A multi-stage sampling method was employed to select the participants. First, a simple random sampling method was applied to choose five (5) out of eight (8) technical universities. Where a university runs a single construction-related program, it was selected automatically for the survey. In schools with more than one construction-related program, one program was deliberately chosen to ensure equal representation. All selected program's final year students and their lecturers were part of the study sample. The survey selected 275 final-year students and 50 lecturers.
The questionnaire utilized in this study was closed-ended and was in three sections. Part I sought to investigate the characteristics of the respondents. The second part evaluated the use of ICT software in training construction students in technical universities. The final part investigated the Constraints on ICT usage in training construction students in technical universities. The questionnaire was structured in a Likert scale form for most of the sections.
Statistical Package for Social Sciences (SPSS) was the primary tool adopted for the data analysis. The main statistical methods of analysis adopted were percentages, tables, averages and frequencies in summarizing respondent details. Factor analysis was utilized to evaluate the inherent interdependencies between the several constructs defined as the barriers to ICT software usage in technical university training. The factor analysis method helps in reducing the variables to a more functional setting. Table 2 shows the rotated component matrix. The Kaiser-Meyer -Olkin (KMO) sampling adequacy test reached 0.925 in the preliminary study. All 24 variables had 1.00 communalities or higher, suggesting their suitability for factor analysis. The 24 constructs were condensed to standard factor patterns. These were undertaken to scientifically illustrate the challenges of using ICT software in Technical Universities educating construction students. In doing this, principal component analysis with Varimax rotation and Kaizer Normalisation was used to determine which factors have empirical significance. Factor retention was by the eigenvalue 1.0 criterion, suggesting that only factors that account for variances greater than one should be included in the factor extraction. The principal component analysis, where linear combinations of observed variables are developed was the technique for extracting the variables.
Background of respondents
Concerning faculty of affiliation, it was realized that the majority (66%) of the respondents were affiliated to the school of engineering. In contrast, the remaining respondents (34%) were affiliated to the faculty of built and natural environment. In terms of departmental affiliation, 38% of the respondents were affiliated to the civil engineering department, while 62% of the respondents were affiliated to the building technology department. Furthermore, the majority (93%) of the respondents were males, while the remaining (7%) were females. This finding can be attributable to the reality that Construction is a male-dominated field.
Usage of ICT tools in training construction students
The survey examined the lecturers' and students' usage of ICT tools in training construction students in technical universities in Ghana. According to table1, the extent of use of Microsoft word by lecturers recorded a mean of 4.30, whereas students recorded a mean of 3.90, which all represented "regularly". Both participants rated it as the most-used ICT software. For Word perfect, the extent of usage by lecturers recorded a mean of 1.80 while that of students scored 1.30, which implies "very rarely" and "not at all" respectively. Education usually is information-intensive, with a vast amount of knowledge produced and used by all involved. Learners utilize Microsoft word primarily for their tasks and other academic programs. Most lecturers often use Microsoft word to take notes and process their research work. This result is anticipated as software like "Microsoft Office suite" is the most commonly known commercially existing combined software for general works like word processing (Doherty, 1997). According to Goh (2005), the typical types of Construction education software comprise "Microsoft word".
Both participants were requested to show their extent of usage of presentation software. In this regard, the students' degree of usage of "Adobe PageMaker" recorded a mean of 1.20, while the results of the lecturers recorded a mean of 1.40, both implying 'not at all.' With "MS PowerPoint", a mean score of 3.20 was recorded for lecturers and 2.40 for students, which implies "rarely" and "rarely". This result points to the fact that both students and lecturers hardly use these tools for presentation. This result differs from that of Doherty (1997) and Arif and Karam (2001), where they stated that "MS PowerPoint" has gained Much popularity in education due to multiple presentations that are done by both students and lecturers. This difference may be explained by the fact that education has reached a higher stage of development in their part of the world where the research was carried out than in Ghana In the case of the use of Architectural Design and Drawing tools like the AutoCAD, the responses of the lecturers recorded a mean of 2.80. In contrast, the responses of the students scored 2.30, meaning "rarely" and "very rarely," respectively. In terms of ArchiCAD, a mean value of 2.60 was recorded for lecturers and 2.10 for students, which implies "rarely" and "very rarely". This signified that both lecturers and students do not use these ICT tools regularly in construction education. Howard (1998) Asserted that Computer-Aided Design (CAD) software is commonly used among construction practitioners, with AutoCAD holding the CAD market's most significant share. Therefore it is imperative for Technical Universities to appropriately integrate these tools into its curriculum in training construction professionals. These CAD systems have replaced conventional drawing boards. CAD Software has vastly enhanced drawing. 3D modeling helps designers to explore the internal spatial structure and its interaction with the external world (Ozumba & Shakantu, 2008).
Concerning the usage of Project Planning software, participants were asked to show their extent of usage. With MS Project, responses from lecturers recorded an average of 2.80, meaning "rarely" used. Conversely, responses from students obtained an average of 2.20, which connotes 'very rarely' used. Concerning Primavera usage, the lecturers' response produced an average of 1.20 and those of the learners obtained an average of 1.10, which both implied "not at all" used. These are all emerging ICT tools that have been introduced to resolve a lot of challenges inherent in the industry. However, these tools are not introduced to construction students in their Education in Technical Universities in Ghana. According to Colwell (2008), Project planning based software is applied in areas such as management of portfolio, scheduling and planning, inventory management, estimating and controlling costs, project management and risk management. Management of portfolio helps organizations define, assess, pick and execute the appropriate projects. Scheduling and planning determine project tasks, their relationships, resource commitment needed and duration of project activities. Moniem (2000) opined that cash flow forecasts could be prepared by the input of cost figures. The implications are that there is the need for all Construction related curriculums in technical universities to be reviewed for these aspects of the trade to be included for an all-round training for the students. The lecturers could also be given in-service training in these areas.
Study participants were also required to identify their extent of use of Structural Design software. Regarding APM Civil Engineering usage, the usage level of the students recorded an average of 1.00. At the same time, those of the lecturers obtained an average of 1.10, which all meant "not at all" used. Concerning structural engineering software like Lira, Robot, SCAD and Monomakh, the trainers and learners' response averages were both ranging from 0 to 1.49. This finding implies that all structural engineering software was "not at all" employed by study participants.
With structural design components in the syllabus of construction-related programs in Ghanaian technical universities, it is necessary to introduce these tools to both lecturers and students. This will, in the long run, train future construction practitioners for the job market and expose them to emerging industry technologies.
Furthermore, both groups of participants were required to state their extent of usage of quantity surveying/measurement/estimation tools. The extent of usage of MasterBill by lecturers recorded a mean of 1.80, whereas that of learners recorded a mean of 1.60, which both implied "very rarely" used. In terms of Revit, the level of usage of lecturers recorded a mean of 1.80, while students' responses recorded a mean of 1.30, which represented "very rarely" and "not at all" in that order. The result signifies that these tools are 'not at all' used by the respondents. Cost control is among the construction project's most critical criteria for project success (Oladapo, 2006). Oladapo (2008) established that these tools help project managers generate the budget of projects and track project expenditure (Oladapo, 2006). Several of this software also assists in quantity taking-off and BOQ development. Therefore it is imperative to introduce construction students to these tools so that they will be capable of performing when they get to the job market.
Finally, on communication tools, both study participants were requested to show their extent of e-mail and internet usage. With e-mail usage, the degree of usage of the lecturers achieved an average value of 4.10, which represented "regularly" used. At the same time, those of the learners obtained an average value of 3.40, which implied "rarely" utilized. The learners and teachers ranked this as the second and third most-used ICT software in construction-related education in that order. For internet usage, a mean value of 3.80 was recorded for lecturers and learners, meaning it is used "regularly". The learners and teachers also ranked this in the third and second positions correspondingly. Teachers and students embraced the internet as it could be used as a useful communication tool. Tam (1999) argued that the internet's contribution to educating construction students comes primarily from its ability to quickly link globally to large volumes of information that otherwise would have taken enormous resources.
Overall, the level of usage of ICT resources by lecturers scored an average of 2.05, suggesting "very rarely" usage. At the same time, those of students reported an average of 1.75, which also implied "very rarely" utilized. Thus, there is ample evidence that the extent of usage of ICT resources in Ghanaian technical university training of construction students is really low. Constraints to ICT usage in training construction students The rotated factor matrix, as illustrated in The results suggest deficits in terms of ICT contents in training construction students in Technical Universities in Ghana. These shortfalls are due in part to the many facilities and logistical challenges facing Technical University education in Ghana (Owusu-Agyeman & Oosterkamp, 2009). To address these shortfalls, the following suggestions are proposed:- Establishing a reliable source of funding purposely for integrating ICT in Construction related programs Introducing ICT in construction education curriculum based on the knowledge and skills required of the construction industry Purchasing excellent Construction related software for lecturers to enhance construction student's knowledge and skill acquisition in ICT Building capacity of lecturers in the use of Construction related ICT tools Instructional methods should include using media technology to improve teaching and learning, including the use of audio and visual technology during teaching. Promoting research on issues relating to ICT usage in construction education. Review of the curriculum to ICT application in all the functional areas of Construction.
Conclusion
This study offers findings and insights on the usage of ICT in training construction students in Ghanaian Technical Universities. The results offer a synopsis of the scope of ICT usage and limitations on ICT adoption in training construction students in Technical Universities in Ghana. It revealed that the extent of usage of ICT resources in training construction students in Ghanaian technical universities is really low. The limitations of using ICT in construction education relate to technology, economic, human, environment, and administration challenges. The implication is that there is a pressing need for a drastic shift towards integrating Construction related ICT resources in training construction students in Ghanaian Technical Universities. | 4,289.6 | 2020-10-15T00:00:00.000 | [
"Education",
"Engineering",
"Computer Science"
] |
Sparse force-bearing bridges between neighboring synaptic vesicles
Most vesicles in the interior of synaptic terminals are clustered in clouds close to active zone regions of the plasma membrane where exocytosis occurs. Electron-dense structures, termed bridges, have been reported between a small minority of pairs of neighboring vesicles within the clouds. Synapsin proteins have been implicated previously, but the existence of the bridges as stable structures in vivo has been questioned. Here we use electron tomography to show that the bridges are present but less frequent in synapsin knockouts compared to wildtype. An analysis of distances between neighbors in wildtype tomograms indicated that the bridges are strong enough to resist centrifugal forces likely induced by fixation with aldehydes. The results confirm that the bridges are stable structures and that synapsin proteins are involved in formation or stabilization. Electronic supplementary material The online version of this article (10.1007/s00429-019-01966-x) contains supplementary material, which is available to authorized users.
Introduction
A key step in chemical synaptic transmission is exocytosis of neurotransmitter stored within vesicles in presynaptic terminals. Exocytosis occurs at the active zone region of the plasma membrane, but space is only available for docking a few of the hundreds within resting terminals. The nondocked vesicles are held in clouds in the interior (Pieribone et al. 1995;Schikorski and Stevens 1997;Rey et al. 2015). The current concept is that some of the vesicles within the clouds occupy reserve pools that can be recruited to the active zone when needed (Denker and Rizzoli 2010).
Electron-dense bridges have been detected between some pairs of vesicles within the clouds (Landis et al. 1988;Hirokawa et al. 1989). Although initially hypothesized to underlie cloud formation, each vesicle was bridged to an average of only 1.5 of the approximately 7 contiguous neighbors in Schaffer collateral terminals, and a different kind of mechanism for cloud formation has been proposed (Siksou et al. 2007;Milovanovic et al. 2018). Based on this, the idea that bridges are stable structures in vivo-as opposed to artifacts of tissue preparation-and the relevance for function, have been questioned (Wang and Kaeser 2018).
However, our own working model of rate-limiting steps in vesicle recruitment to the active zone at Schaffer collateral terminals does ascribe a critical function to the bridges that is very much in-line with the sparse distribution, and even Electronic supplementary material The online version of this article (https ://doi.org/10.1007/s0042 9-019-01966 -x) contains supplementary material, which is available to authorized users. predicts the precise value of 1.5 per vesicle (Gabriel et al. 2011). The idea is that vesicles throughout the terminals are connected into short chains consisting of only four vesicles (Fig. 1).
The chains were included in the model to explain functional results that are not related to the discussion about whether or not the bridges are involved in cloud formation. Specifically, the chains can explain evidence against the idea at the core of most other models that mass action plays a role in the timing of vesicle recruitment to the active zone (Stevens and Wesseling 1999;Garcia-Perez et al. 2008;Gabriel et al. 2011); the chains would dock to the active zone as a unit, endowing each release site with an autonomous reserve pool that would be replenished, stochastically, as an all-ornone event. We were, therefore, interested in determining if the bridges are stable enough to play a role in vivo as a consistency check for the working model.
The molecular composition of the bridges is not known, but synapsin proteins have long been hypothesized to play a role (Landis et al. 1988;Benfenati et al. 1989;Hirokawa et al. 1989;Hilfiker et al. 1999). If our working model is correct, the chains would continue to be present in knockout synapses missing synapsin 1 and 2, but would have to be shorter, and the number of links per vesicle would be reduced (Gabriel et al. 2011). A previous electron tomography study demonstrated that at least some of the bridges remain in synapsin knockouts, but did not resolve whether the number per vesicle was altered (Siksou et al. 2007). We were, therefore, additionally interested in determining if the number is reduced in synapsin knockout synapses.
Standard segmentation techniques for comparing the bridges in tomograms from wildtype (WT) and synapsin knockout synapses can be complicated by the fact that the density of vesicles within clouds is substantially lower in knockout synapses (Li et al. 1995;Rosahl et al. 1995;Gitler et al. 2004;Siksou et al. 2007). However, we have now developed a new method for quantifying the bridges that avoids this complication.
We detected a substantial reduction in the fraction of pairs of neighboring vesicles that are connected by a bridge in knockouts after fixation with aldehydes, along with other indications that the bridges are real structures in vivo, and not artifacts of tissue preparation. A follow-on analysis of the distribution of distances between neighbors indicated that the bridges are strong enough to resist centrifugal forces likely induced by vesicle shrinkage caused by the fixation procedure. Taken together, the results indicate that the bridges play a physical role in tethering vesicles together in vivo and confirm the involvement of synapsin proteins in formation or stabilization.
Tissue preparation
Age-matched 14-to 21-day old WT and synapsin DKO mice were first anesthetized with 20 μl∕g of a cocktail containing 3% ketamine and 0.08% xylazine. Brains were then fixed by transcardial perfusion with a ringer solution containing 2 UI/ ml heparin and 0.02% xylocaine followed by a fixation solution containing 2% paraformaldehyde, 2.5% glutaraldehyde, and 0.15 M cacodylate (Giepmans et al. 2005). Fixed brains were stored at 4 • C for up to 4 days in the same solution before sectioning at 100 μm , high-pressure freezing, freeze substitution with tannic acid, then osmium tetroxide in acetone, and infiltration with Durcupan resin and polymerization as described in Sosinsky et al. (2008). 300 nm sections were collected on 200 hexagonal mesh grids coated with 0.5% formvar and then stained with 2% aqueous uranyl acetate ( 20 min ) or Sato's lead ( 5 min ) (Sato 1968), or both in sequence as indicated. The grids were then coated with either 5 or 10 nm colloidal gold particles as fiducial markers, and electron microscopy images of classically defined asymmetric synapses in stratum radiatum of the hippocampus were acquired at 300 keV on an FEI Titan microscope at 37,000 × (Group 1 in Fig. 5) or 29,000 × (Groups 2 and 3) magnification.
Image acquisition
Two tilt series were recorded for each specimen from at least − 50 • to + 50 • at 1 • increments; the specimen was rotated 90 • in the horizontal plane between the first and second tilt series (Mastronarde 1997). Fig. 1 Model of cloud where synaptic vesicles are tethered together into short chains. This scenario was proposed in Gabriel et al. (2011) to explain simplifying mathematical constraints that emerged from electrophysiological studies of rate-limiting steps in synaptic vesicle cycling at Schaffer collateral synapses. Vesicles throughout the terminal are linked together in short chains. Non-docked vesicles attached to docked vesicles serve as an autonomous reserve pool that can be expended/depleted during heavy use. Docked vesicles are replaced stochastically at the slow rate of 1/min with a vesicle at the start of a full chain that was previously not docked
Tomogram reconstruction
The etomo program of the IMOD suite (Kremer et al. 1996) was used for fiducial alignment and standard data manipulations prior to back projection. Back projection reconstruction was then performed using the TxBR software package (Lawrence et al. 2006).
Segmentation
Center location and radius of vesicles was estimated by a two-step procedure. First, rough estimates were obtained by modeling each vesicle with a sphere using the 3dmod computer program of the IMOD suite. Then, a separate minitomogram containing each vesicle was extracted from the whole tomogram, and estimates were fine tuned with the graphical user interface depicted in Fig. 2.
Experimental design and statistical analyses
Experimental design is described throughout the "Results" section. Statistical analyses were performed as indicated throughout the "Results" and figure legends.
Results
We employed a combination of chemical and rapid highpressure freezing techniques to fix brains of age-matched WT and synapsin 1 and 2 double knockout mice (DKO); mice were between 14 and 21 days old to match the range Fig. 2 Graphical user interface used for fine tuning estimates of vesicle center and radius. As a first step in the analysis, rough estimates for the position and size of each vesicle within each tomogram were obtained by modeling each vesicle with a sphere using the 3dmod computer program of the IMOD suite of software (Kremer et al. 1996). A separate mini-tomogram containing each vesicle was then extracted from the whole tomogram. The three images in the user interface are all the means of virtual slices from the same mini-tomogram spanning the estimated center point (i.e., spanning a total of 0.5 median vesicle diameters). The center image is the mean of vir-tual slices parallel to the horizontal plane without rotating the minitomogram, the leftmost image is after rotating the mini-tomogram 45 • about the y-axis, and the rightmost image is the mean after rotating about the x-axis. The white circle is the perimeter of the sphere calculated from the center and radius estimates and will match all three images when the estimates are correct. The horizontal sliders directly below the leftmost image determine contrast. The bottom-most horizontal slider allows the observer to adjust the radius estimate. The vertical sliders allow the user to adjust the estimated location of the center point used in previous functional studies ( Garcia-Perez et al. 2008;Gabriel et al. 2011). The fixation procedure has previously been shown to preserve labile tissue ultrastructure for electron tomography at a level that is comparable to rapid high-pressure freezing alone (Sosinsky et al. 2008), and allowed us to avoid artifacts related to dissecting and freezing living tissue, and to avoid shipping live animals between laboratories. Classically defined asymmetric synapses in stratum radiatum of the hippocampus were then imaged with an electron microscope from multiple tilt angles along two orthogonal axes and tomograms with at least 6.4 Å resolution were generated as described in "Methods" (view the supplementary Video for an example). Bridges between some vesicles were obvious, particularly in WT tomograms (Fig. 3a). See below for multiple types of evidence that the bridges were real structures and not artifacts of chemical fixation or freezing.
Detection procedure
Vesicles within synapsin DKO terminals were less crowded together than vesicles in WT terminals owing to fewer vesicles overall as reported previously (Rosahl et al. 1995). The difference was readily apparent while browsing through the tomograms using the 3dmod computer program of the IMOD suite (Kremer et al. 1996), precluding unbiased blind comparisons at the level of whole tomograms. However, we did not use automatic machine detection algorithms because of the concern that false negatives or false positives might occlude relevant differences.
Instead, we first estimated the center location and radius of individual synaptic vesicles (see "Methods"). No difference was detected in the size of vesicles at WT vs synapsin DKO synapses; the mean, across tomograms, of the median vesicle diameter for each was 37.6 ± 0.5 nm for WT and 39.6 ± 0.5 nm for synapsin DKO. Median vesicle diameter of each tomogram was used as our standard measure of length for comparisons across tomograms throughout the study.
We then used computer programs to: (1) automatically extract a separate miniature tomogram for each vesicle pair; and (2) present images from each miniature tomogram in a random order to an observer who was asked to score for the presence or absence of a bridge on a five point scale (Fig. 3a). The observer was additionally given the option of indicating that the vesicles appeared to be touching, which would make detecting a bridge impossible, and if images were perceived to be of low quality.
We conducted an initial survey on a subset of vesicle pairs separated by up to 1.5 median vesicle diameters from 13 WT and 15 DKO tomograms. A maximum of 150 vesicle pairs from each tomogram was included in the analysis. The pairs were chosen at random, irrespective of distance to landmarks such as the plasma membrane or active zone (but see below). Pairs where neither vesicle was at least 1.5 vesicle diameters from artificial edges of the vesicle cluster created by tissue sectioning were excluded. The full analysis consisted of 1631pairs for WT and 931 for DKO because some tomograms contained fewer than 150 such pairs. Tomograms containing fewer than 50 pairs (all synapsin DKO) were excluded.
Scores of 1 ("Certain") and 2 ("Likely") were counted as bridges. The survey revealed degradation in image quality (Fig. 3b, c), and information content (Fig. 3d), for vesicle pairs that were oriented more than 45 • from the horizontal plane, likely resulting from the so-called missing wedge effect inherent to tomograms reconstructed from axial tilt series (Frank 2006). The survey additionally revealed that almost all bridges were between vesicles separated by 0.5 median vesicle diameters or less (Fig. 3e).
Fewer bridges between neighbors in synapsin DKO synapses
The results in Fig. 3e suggest that equivalently spaced vesicles are less likely to be linked together by a bridge in synapsin DKO compared to in WT synapses, in line with the model proposed in Gabriel et al. (2011). For higher resolution information, we next conducted a second random survey, again containing a maximum of 150 vesicle pairs from each tomogram, but this time the pairs were drawn from the subset that were oriented within 25 • of the horizontal plane and had separation distances between vesicles of 0.5 median vesicle diameters or less. For this second survey, the full analysis consisted of 1950 pairs for WT and 1673 for DKO; more pairs were available per tomogram despite the additional constraints on separation distance and angle because Fig. 3 Method for evaluating the presence or absence of bridges between vesicles. a Screenshot of graphical user interface for evaluating presence or absence of bridges. Mini-tomograms for each vesicle pair were rotated so that the centers of both vesicles were within the horizontal plane. Images of vesicle pairs were then presented in a random order simultaneously from three angles: not rotated; and rotated ±45 • about the central axis common to both vesicles. When the "3-D display" radio button is selected, the scroll bar to the right allows the user to scroll through the individual virtual sections. Otherwise, the mean images of all virtual sections between 0.25 median vesicle diameters above and below the central plane are displayed. The lower scroll bars below the images control contrast. b-e Initial survey of 13 WT and 15 DKO tomograms for vesicle pairs separated up to 1.5 vesicle diameters. b Z-angle for a pair of vesicles is defined as the angle between the axis passing through the centers of both vesicles and the plane that is horizontal to the tissue slice. c Fraction of WT vesicle pairs for which images were evaluated as qualitatively "Bad" or "Terrible" vs Z-angle. d Probability of detecting a bridge vs Z-angle in WT synapses. Scores of 1 ("Certain") and 2 ("Likely") were counted as bridges. e Probability of detecting a bridge vs distance between the vesicles for Z-angles ≤ 25 • (640 pairs for WT, 374 for DKO). Distance units are median vesicle diameters, which were calculated across all vesicles for each tomogram independently ◂ pairs were included where neither vesicle was less than 0.5 vesicle diameters from artificial edges of the vesicle cluster, compared to the 1.5 vesicle diameters excluded in the first survey. We found that the probability of detecting a bridge between neighboring vesicles was more than twofold lower for synapsin DKO synapses (Fig. 4a, b), whereas no difference was detected between WT and synapsin DKO in the fraction of pairs judged to be ambiguous (Fig. 4c). Similar results were obtained for specimens post-fixed with Uranyl Acetate with or without Sato's lead staining (Groups 1 and 3 in Fig. 5). Sato's staining without Uranyl Acetate yielded similar results for WT tissue, but only a single tomogram of a synapsin DKO synapse surpassed the lower limit of 50 pairs meeting criteria (Group 2 in Fig. 5) precluding any test for statistical significance.
The mean distance between vesicles within the pairs was slightly greater for synapsin DKO synapses (i.e., even though all were within the 0.5 median vesicle diameter limit established by the criteria). However, the lower probability of detecting a bridge between neighboring synapsin DKO vesicles could not have been simply because the vesicles were farther apart because the relative difference between WT and synapsin DKO was even greater when the analysis was restricted to the shortest distances. For example, bridges were only detected in 3 of 42 synapsin DKO pairs (7.1%) for separation distances of less than 0.1 median vesicle diameters, but 41 of 112 WT pairs (36.6%; p < 0.001, 2 with Yate's correction).
These results provide strong evidence that most of the bridges are real structures and not artifacts of tissue preparation that would pertain equally to WT and synapsin DKO tissue; this is the first of multiple types of evidence against fixation artifacts noted above. No differences were detected in the probability of finding a bridge between neighbors located proximal vs distal to the active zone for either genotype (Fig. 6). However, the analysis was specifically designed to detect bridges between pairs of vesicles; the increased amount of synapsin seen at locations distal to the active zone using immunochemical techniques in previous studies might reflect soluble synapsin in a liquid phase or bridges to elements of the cytoskeleton which would not have been detected here (Evergren et al. 2007).
Quantification of WT bridges is consistent with new model
As noted in the "Introduction", our working model predicts that vesicles in WT synapses are connected into short chains of four vesicles (Gabriel et al. 2011). The technical limitations documented above caused by the missing wedge phenomenon (i.e., Fig. 3c-e) precluded identification of full chains in most cases. However, we did find clear examples of chains with four vesicles while browsing through the full tomograms using the 3dmod computer program of the IMOD suite (Kremer et al. 1996) (Fig. 7a, b). In addition, information about the number of neighbors per vesicle was available from the initial step of the analysis where the center location and diameter of vesicles throughout the tomograms were estimated (Fig. 7c). This information could be combined with the probability of detecting a bridge in Fig. 4b to extrapolate the mean number of bridges from each vesicle to neighbors (Fig. 7d). The resulting value of 1.3 ± 0.1 for WT tomograms was close to the value estimated in a previous study where tomograms were acquired from tissue that was rapidly frozen without first fixing with aldehydes (Siksou et al. 2007). The close agreement provides quantitative support for the conclusion in Sosinsky et al. (2008) that the combination of chemical fixation followed Fig. 4 Lower probability of detecting a bridge between neighboring vesicles in synapsin DKO synapses. a Images of pairs of vesicles with and without a bridge. Magenta arrows indicate bridges. Scale bars are 30 nm. Image thickness was 8.3 nm (i.e., average of 13 virtual sections with a voxel spacing of 6.4 Å ). Post-fix staining for these examples was with uranyl acetate, but no lead, which is referred to as "Group 1" below. b Probability of detecting a bridge per pair of neighbors vs genotype. Values were estimated individually for each tomogram by dividing the number of bridges detected by the number of pairs that were analyzed. Circles are the median values across tomograms, boxes delineate the middle two quartiles, ***p < 0.001 (Kolmogorov-Smirnov; n = 13 tomograms for WT, 15 for DKO). c No difference in scores of "3 Can't tell" 1 3 some time later by rapid high-pressure freezing can preserve the ultrastructure of labile tissue. In addition, the values are in line with the prediction of our working model that vesicles are connected together in short chains consisting of four vesicles; if the model is correct, the mean number n would be n = (2⋅r)−2 r , where r is the number of vesicles in a chain and n would be 1.5 for WT because r = 4.
Synapsin DKO bridges are less detectable
However, we detected only 0.24 ± 0.04 bridges per synapsin DKO vesicle (Fig. 7d), which is less than predicted by the model; i.e., chains in synapsin DKO synapses were predicted to contain r = 2.6 vesicles, on average, in which case each vesicle would be connected to 1.2 others. The result does not support the model per se, but does not rule it out either because some bridges in synapsin DKO synapses may have escaped detection owing to lower contrast compared to WT bridges. And indeed, bridges were more difficult to detect in synapsin DKO tomograms; only 1.8% (3 of 165) of all bridges were scored as 1's (as opposed to 2's), whereas 4.8% (18 of 372) of WT bridges were scored as 1's ( p < 0.05, 2 with Yate's correction). Or, some of the bridges may have broken during fixation (see below). In any case, synapsin DKO vesicles had 3.2 ± 0.4 (s.e.m.) neighbors on average (Fig. 7c), which is safely more than the minimum of 1.2 required of the model.
No evidence for lateral bridges between docked vesicles
Next, an additional analysis focused on bridges between pairs of vesicles of which one or both were docked (Fig. 8).
The probability of finding a bridge between pairs when both were docked was lower compared to when neither was docked (first vs third box in Fig. 8c) or compared to when only one of the pair was docked (fifth vs third box). For this analysis, we compared binomial proportions calculated directly from the data sets as a whole instead of the median values calculated individually for each tomogram used above. The two types of comparison yielded similar results (compare leftmost 2 bars in Fig. 8c to Fig. 4b), but comparison of binomial proportions was more appropriate in this case because of the near absence of bridges between pairs of docked vesicles (i.e., only 68 docked pairs met the criteria established above, and of those, we detected only 4 putative bridges in all 28 tomograms). Indeed, the 4 may well have been false positives, which would be in-line with our working model because docked vesicles are predicted to be connected to vesicles within the interior, not laterally to other docked vesicles.
Spatial inhomogeneity within vesicle clouds
Finally, an analysis of vesicle locations within the clouds confirmed that the bridges are strong enough to bear force.
That is, non-docked vesicles were not randomly distributed in space as would be expected of independently mobile constituents of a liquid phase. This can be seen by comparing histograms of the distances between neighbors to the analogous histograms calculated from: (1) the same tomograms after simulating diffusion of the individual vesicles using a random walk algorithm as depicted in Fig. 9a (compare bars to green squares in Fig. 9b); or, (2) simulations of Fig. 4b for each protocol ( n = 5 and 10, respectively, for WT and DKO tomograms for Group 1; 5 and 1 for Group 2; and 3 and 4 for Group 3; * p < 0.05 , Kolmogorov-Smirnov) cloud formation generated by randomly placing spheres in an initially empty space (compare bars to orange squares in Fig. 9b).
For this analysis, simulations were simplified by limiting the spatial bounds to hyperrectangles rather than by approximating the irregular borders of real vesicle clouds. Random walks were applied to hyperrectangular subsets from the interiors of vesicle clouds (e.g., red box in Fig. 9a), and cloud formation was simulated within initially empty 300 × 300 × 300 nm cubes. The simplification was appropriate because limiting spatial bounds of simulations to hyperrectangles did not alter the distribution of distances between neighbors (compare black circles to bars in Fig. 9b).
Randomly distributed vesicles touched in some cases (non-zero values of leftmost green and orange points in Fig. 9b), but neighbors in actual tomograms rarely came close (leftmost bar in Fig. 9b is close to zero), indicating that some mechanism kept vesicles apart. In principle, the absence of touching could be caused by mutually repellent forces between vesicles, but was more likely to have occurred during tissue fixation because aldehydes are known to cause organelles to shrink-likely owing to extrusion of water-which would open spaces between vesicles (Murk et al. 2003); a similar mechanism may account for aldehyde induced flattening of vesicles in Gray's type 2 synapses (Gray 1969;Rapisardi and Lipsenthal 1984;Korogod et al. 2015). And indeed, vesicles did appear to touch in tomograms of tissue that was fixed by rapid freezing without aldehyde fixation (leftmost bar of Fig. 9c); and, the diameters of aldehyde fixed vesicles are consistently less than when unfixed tissue is frozen rapidly. We estimated a reduction in diameter of 10-13%; the 13% value was obtained by comparing the mean of vesicle diameters here (39.6 nm) to 45.6 nm in Imig et al. (2014), whereas the 10% value emerged from the modeling study documented next.
Simulating shrinkage or mutual repulsion after cloud formation did prevent vesicles from touching, but could not be tweaked to reproduce the peak in the WT distribution occurring at about 0.25 median vesicle diameters that can be seen in Fig. 9b. Instead, the peak aligned closely with that of the distribution of lengths of the bridges between Magenta arrows indicate bridges. Models and images correspond to the same tomogram, but the tilt angles are slightly different to better illustrate the docked vesicles. Scale bars are 100 nm and pertain to both the model and corresponding image slice. a WT. b DKO. c Fractions of pairs with a bridge vs distance between the geometric center of the vesicle pairs and the center of the closest docked vesicle. Same data set as Figs. 4 and 5; every tomogram had at least two docked vesicle because only synapses where the synaptic cleft was visible were selected for imaging. Mean number of docked vesicles/tomogram was 13.7 ± 1.8 for WT and 10.3 ± 1.5 for DKO (not significant, Kolmogorov-Smirnov). Probability of detecting a bridge was calculated for each range of distances for each tomogram separately and then averaged across tomograms; n = 13 tomograms for WT and ≥ 14 for DKO instead of the 15 in previous figures because no pairs were present farther than 4 median vesicle diameters from the active zone for one of the DKO tomograms ▸ vesicles (Fig. 9d), suggesting that the bridges restricted mobility. And indeed, simulating cloud formation by placing chains of vesicles-rather than individuals-at random locations in an initially empty space, and subsequently reducing the diameter of each vesicle 13% to simulate shrinkage, did generate distributions with peaks that matched actual tomograms (blue circles in Fig. 9e).
A perfect match was only obtained when the density of vesicles in the simulation was lower than the density in actual tomograms. That is, density was quantified as the volumetric fraction by dividing the sum of the volumes of all vesicles in the hyperrectangles by the volume of the hyperrectangles. Volumetric fraction in WT tomograms was 0.17 ± 0.01 , but simulations only matched actual measurements when the volumetric fraction was lowered to 0.06 (compare black and green circles in Fig. 9f).
However, the remaining mismatch at a volumetric fraction of 0.17 could be eliminated by introducing a small amount of order to the placement of chains, for example by combining the random placement with the regular placement depicted in Fig. 9g at a ratio of 85 to 15 (red circles in Fig. 9e). In this case, the best match was obtained when vesicle shrinkage was 10% rather than 13%.
A key point is that the distribution of distances between neighbors could only be reproduced if bridges were modeled as rigid structures that were strong enough to drag vesicles closer together to compensate for the decreases in diameter. Also relevant, bridge lengths in simulations were chosen by randomly sampling from the distribution of bridge lengths measured in tomograms, whereas the distribution of distances between neighbors could not be reproduced if bridge lengths were instead drawn from wider or narrower distributions, or from distributions with substantially different means. Moreover, the distribution of distances between neighbors within DKO synapses, where bridges were fivefold less frequent (Fig. 7d), could be fit by simulations where individual vesicles were placed at random locations and then shrunk by between 10 and 13% (Fig. 9h).
Taken together, these results suggest strongly that the bridges in WT synapses provide force for holding vesicles together. If our working model is correct, the placement of the chains would not be completely random in space, consistent with an additional level of organization (Siksou et al. 2007;Fernández-Busnadiego et al. 2010;Schrod et al. 2018), or simply crowding/compaction by a still unidentified mechanism.
Tethered pairs overlap after reversing shrinkage
The conclusion that bridges provide physical force for holding vesicles together was strengthened further by the observation that neighboring vesicles connected by bridges were closer together than neighbors that were not connected ( 0.19 ± 0.004 median vesicle diameters versus 0.31 ± 0.005 ). And, simulating the reverse of shrinkage by increasing the diameter of each vesicle by 15%-i.e., 45.6 nm 39.6 nm -often caused neighbors to overlap if they were connected by a bridge ( 23% ; 71 of 306), but almost never if they were clearly not connected (3%; 16 of 496). The difference was statistically significant ( p < 1E − 5 ; 2 ), and was not simply because the pairs without a bridge were farther apart because the fraction of overlaps only rose to 11% (14 of 132; p < 0.003 ) for pairs separated by up to only 0.23 vesicle diameters; i.e., for which the mean separation distance ( 0.19 ± 0.002 ) matched the mean for neighbors that were connected by bridges.
Discussion
Here we report: (1) a new method for detecting bridges connecting neighboring vesicles in electron microscopy tomograms; that we have used to (2) demonstrate a reduction in synapsin knockouts in the probability that any two neighbors are connected. Previous doubts about the relevance of the bridges for synapse structure and function are countered by multiple independent indications that the bridges are real structures that can bear force rather than artifacts of sample preparation.
The bridges have long been suspected of containing synapsin proteins (Landis et al. 1988;Benfenati et al. 1989;Hirokawa et al. 1989). However, our results do not support the long-standing synapsin hypothesis where vesicles would be heavily interconnected in a dense network (rev. in Hilfiker et al. 1999). Instead, individual vesicles were only connected to an average of 1.3 of 7 neighbors (i.e., ∼ 20%).
Nor do our results prove that synapsin proteins are constituents of the bridges. Instead, synapsins might be involved in bridge creation or stabilization by a more indirect mechanism (Orenbuch et al. 2012a). And indeed, we found that at least some bridges continued to be present in synapsin DKO terminals. The sparse distribution of bridges and the observation that at least some remain in synapsin knockouts was reported previously (Siksou et al. 2007), but, to our knowledge, the present study is the first to demonstrate that the probability of finding a bridge between neighbors is altered in synapsin knockouts.
Our results do not argue against the recent proposal that clouds of vesicles are held together by a liquid phase of quickly associating and dissociating synapsin molecules (Milovanovic et al. 2018). If so, individual vesicles would not be free to move around independently, but individual chains of vesicles might be. However, our results are equally consistent with the idea that short chains of vesicles are fixed within a larger cytoskeletal structure (Siksou et al. 2007;Fernández-Busnadiego et al. 2010;Schrod et al. 2018), in which case motion would instead be directed by motor proteins.
Model where vesicles are linked into short chains
Our initial motivation was to conduct a consistency check for our working model of rate-limiting steps in synaptic vesicle trafficking (Fig. 1), which was originally proposed to explain purely functional results that are incompatible with the premise of conventional models that the timing of vesicle recruitment to a readily releasable pool is influenced by mass action of vesicles in reserve pools (Stevens Fig. 8 Comparison of bridges between non-docked and docked vesicles. The analysis included all vesicle pairs where at least one of the two vesicles was docked, along with a randomly selected subset of vesicle pairs where neither vesicle was docked (Z-angle was ≤ 25 • for all pairs). a, b Images of WT and DKO pairs where one of the vesicles is docked and is bridged to the other in the cytoplasm (left panels), and where both vesicles are docked (right panels). Scale bars are 30 nm, thickness of virtual section was 8.3 nm. Magenta arrows indicate bridges. c Quantification: horizontal lines within boxes are fraction of pairs judged to be connected by a bridge across the entire analysis, which is a different type of quantification than used above (see "Results"). Boxes delineate the 90% confidence interval for binomial proportions using the Wilson score interval with correction for continuity (Newcombe 1998); ***p < 0.001 , * p < 0.05 ; both using 2 with Yate's correction. Raw values are listed in the table and Wesseling 1999;Garcia-Perez et al. 2008;Gabriel et al. 2011). The present results provide qualitative support for the new model by confirming that, at least in Schaffer collateral synaptic terminals: (1) vesicles throughout the clouds are tethered together; (2) synapsins are involved; and, (3) docked vesicles are tethered to vesicles in the interior.
The results are in line with the quantitative prediction that vesicles are each tethered to an average of 1.5 others in WT Schaffer collateral terminals. However, quantitative aspects of the results must be taken with caution because of an unknown number of false negatives. And, indeed, the probability of finding a bridge between a given pair of vesicles in synapsin DKO terminals was substantially less than predicted, which might simply be because bridges in DKO terminals were more difficult to detect, or were weak enough to be broken during fixation.
Finally, Fig. 1 depicts vesicles arranged in linear chains, but the missing wedge phenomenon documented in Fig. 3b-d prevented us from ruling out branching arrangements. In any case, the working model was generated to explain functional results that suggested that vesicles in WT Schaffer collateral synapses are held together in groups of 4, which would be compatible with both branched and linear arrangements. A key point, however, is that, while compatible with the working model, the results do not rule out other possible arrangements such as much longer chains.
New method
A key methodological complication was the well-established observation that synaptic terminals from synapsin knockouts contain fewer vesicles than WT and that the spatial distribution is less dense (Li et al. 1995;Rosahl et al. 1995;Gitler et al. 2004;Siksou et al. 2007). The lower density is noticeable when tomograms are viewed as a whole, precluding blind comparisons using standard segmentation techniques. An alternative strategy for avoiding biases produced by human judgments would be to develop a computer program for detecting the bridges automatically (e.g. Fernández-Busnadiego et al. 2010). However, automatic detection can be less sensitive than the human eye, meaning that the algorithms may produce more false positives, more false negatives, or both, which could have occluded key results. In addition, automatic algorithms sometimes require human input-such as setting threshold levels-which can also introduce judgment biases.
Instead, we partitioned the initial tomograms into mini tomograms containing slices through the center of neighboring pairs of vesicles. At this scale, it was no longer possible to distinguish individual WT and synapsin DKO samples. We then presented an extended, randomized series of the mini tomograms to an observer who evaluated the space between each pair of vesicles for the presence or absence of a bridge (see Fig. 3a). The procedure allowed us to take advantage of the facility of human vision for pattern detection to produce quantitative estimates of the probability of finding bridges between pairs of neighboring vesicles that were not compromised by biases caused by preconceptions of the observer.
Why eliminating synapsin decreases the number of vesicles is not known. One possibility would be that vesicles that are not incorporated into the chains hypothesized by our working model are targeted for degradation simply because they are free floating (Li et al. 1995;Orenbuch et al. 2012b), but other mechanisms continue to be possible as well (Sun et al. 2006;Gaffield and Betz 2007).
Note that our analysis was designed specifically to evaluate whether the bridges between vesicles seen in WT synapses are reduced in number after knocking out synapsins. A different method would be needed to evaluate whether synapsins are additionally involved in tethering vesicles to cytoskeletal elements (Evergren et al. 2007).
Controls for fixation artifacts
The tissue used for the present study was prepared first by chemical fixation followed by rapid freezing up to several days later, which allowed us to avoid shipping live animals between laboratories, and to study synapses that were more deeply embedded in tissue than allowed by rapid freezing alone (Korogod et al. 2015 reported ice damage at depths of greater than 10-20 μm ). We are aware that non-native structures produced by chemical fixation may have complicated the interpretation of some previous electron microscopic studies and that omitting this step before freezing can reduce the impact of such factors in some cases (Wolosewick and Porter 1979). However, we are confident that most of the bridges detected in our tomograms were native structures, and not the product of fixation, because: (1) the combination of chemical fixation with freezing used here has previously been shown to preserve labile tissue ultrastructure for electron tomography at a level that is comparable to rapid high-pressure freezing alone (Sosinsky et al. 2008); (2) the probability of finding a bridge between neighboring vesicles is greatly reduced in synapsin DKO synapses; (3) if present, fixation artifacts would have additionally been found between neighboring pairs of docked vesicles, but this was almost never seen; and (4) an analysis of local inhomogeneities in the positions of vesicles within clouds indicated that the bridges were strong enough to resist centrifugal forces likely produced by vesicular shrinkage during the chemical fixation stage. In any case, quantitatively similar numbers of bridges between vesicles were seen in tomograms from the same brain region when unfixed tissue was frozen rapidly (Siksou et al. 2007).
Synapsin 3
The synapsin family contains three genes, but we analyzed synapsin 1 and 2 double knockouts rather than triple knockouts so that results could be compared directly to our previous functional study (Gabriel et al. 2011). However, we would not expect triple knockout synapses to be substantially different because synapsin 3 is not thought to be expressed in Schaffer collateral terminals during the 14-21 day range analyzed here (Hosaka and Südhof 1998;Ferreira et al. 2000), and because synapsin 3 appears to play a functional role that is qualitatively different compared to synapsins 1 and 2 (Feng et al. 2002). In addition, the previous tomography study has already demonstrated the continued presence of bridges between vesicles in synapses from the triple knockouts (Siksou et al. 2007). Fig. 9 Inhomogeneities in spatial distribution are predicted by short chains connected by force-bearing bridges. a Example of hyperrectangular subset of vesicles before and after simulating a random walk in space. b Bars are fraction of WT vesicle pairs vs separation distance between vesicles in each pair in hyperrectangles before the random walk. Black circles are the same measurement for the entire WT data set, which includes vesicles inside and outside the hyperrectangles. Green squares represent pairs in hyperrectangles after the random walk. Orange squares represent pairs from simulations of cloud formation where vesicles that were not connected into chains were placed at random locations in a cube. c Analogous histogram for tomograms from Siksou et al. (2009) where tissue was frozen rapidly without chemical fixation ( n = 316 pairs from seven tomograms containing a mean of 36.7 ± 6.3 vesicles/tomogram; one of eight tomograms was excluded owing to too few pairs). d Histogram of measurements of bridge lengths from the WT data set used in simulations. The blue line is a normal distribution with mean of 0.18 and standard deviation of 0.06. e Left panel is a snapshot of a simulated cloud where vesicles connected into chains of four were placed at random locations in a cube. Each vesicle was shrunk by 13% after cloud formation to simulate the shrinkage that likely occurs during aldehyde fixation. Bridges are represented by the sparsely distributed brown rods. The bars in the right panel are identical to in b; blue circles are the analysis of simulations matching the snapshot to the left; red circles are after adding a small amount of order as described in "Results". f Analogous histogram for simulations where vesicles were connected into chains of four for a variety of volumetric fractions. The blue circles are the same as in e; the simulations of lower volumetric fractions were then generated by eliminating randomly selected chains. Native data are the same as in b. g Left panel is a snapshot of a simulated cloud where vesicles connected into chains of four were placed in an ordered rather than random formation; volumetric fraction was 0.17. Right panel is histogram of separation distances between vesicles in each pair. h Analogous histogram for DKO tomograms; compare to bars in b. Magenta squares are from simulations of cloud formation where vesicles were not connected into chains, followed by shrinkage by 12%; volumetric fraction was 0.07 for both tomograms and simulation ◂ | 9,746.8 | 2019-10-30T00:00:00.000 | [
"Biology",
"Physics"
] |
Using Fully Homomorphic Encryption to Secure Cloud Computing
The concept of cloud computing receiving a great deal of attention both in publication and among users. Cloud computing is the delivery of computing services over the Internet. Cloud services allow individuals and businesses to use software and hardware resources that are managed by cloud providers at remote locations. The distance between the client and the physical location of his data creates a barrier because this data can be accessed by a third party and this would affect the privacy of client’s data. The using of traditional encryption schemes to encrypt the remoted data before sending to the cloud provider has been most widely used technique to bridge this security gab. But, the client will need to provide the private key to the server to decrypt the data before perform the calculations required. Homomorphic encryption allows to perform computations on encrypted data without decryption. This paper deals with the use of homomorphic encryption to encrypt the client’s data in cloud server and also it enables to execute required computations on this encrypted data.
Introduction
Cloud computing is a hot topic in the information technology field. It enables users to get almost unlimited computing power and it offers potential benefits to these users in terms of instant availability, scalability and resource sharing. Examples of cloud services offers by cloud providers include online file storage (e.g. Dropbox), social networking sites (e.g. Facebook), webmail (e.g. Gmail), and online business application (e.g. Brokerage).
The essential characteristics of cloud computing include on-demand self-service, broad network access, resource pooling, rapid elasticity and measured service. On-demand self-service means that clients (users or organizations) can request and manage their own computing resources. Broad network access allows services to be offered over the Internet or private networks. Pooled resources mean that customers draw from a pool of computing resources, usually in remote data centers. Rapid elasticity means that services can be scaled larger or smaller. And use of a service is measured and customers are billed accordingly [1].
Although cloud computing has become a mature service model, the adoption of its services by customers (businesses, consumers, etc.) is limited by concerns about the loss of privacy of their private data. Encryption of data could solve this issue, but if the clients want to manipulate their encrypted data in the cloud, they have to share the secret key with cloud provider to decrypt it before execute the required operations [2].
Homomorphic encryption is the appropriate solution to solve cloud computing security issues, since its schemes enable to perform computations on encrypted data without sharing the secret key needed to decrypt the data.
In 2009, Craig Gentry [3] introduced the first fully homomorphic encryption (FHE). In 2010, M. Dijk, C. Gentry et el. [4] presented a second fully homomorphic encryption. In the march 2010, Gentry [5] proposed a homomorphic encryption scheme (called Gen10), heading toward widespread use of cloud computing. Unfortunately these schemes are insecure to use in cloud computing. In 2012, Jian Li, Danjie Song et al. [6] proposed a simple FHE derived from Gentry cryptosystem to ensure the privacy in cloud storage, namely SDC scheme. In 2014, Chen and Zhao [7] proposed an improvement to the second scheme of Gentry to make the application of FHE in cloud is more secure.
Many researchers proposed different applications of homomorphic encryption in cloud computing. In 2014, Yan Zhang et el. [8] proposed a secure image retrieval method for cloud computing based on homomorphic properties of Paillier scheme. In 2015, Kocabas and Soyata [9] presented a method for privacy-preserving medical cloud computing using fully homomorphic encryption. In January 2016, Shu Qin Ren et el. [10] proposed an XOR homomorphism encryption scheme to support secure keyword searching on encrypted data foe cloud storage.
This paper addresses the security and confidentiality of user data in cloud computing. The main aim of the paper is to introduce the concepts of homomorphic encryption and how to exploit these concepts to secure cloud computing data.
The rest of the paper is organized as follows: section 2, describes the cloud computing concepts and its deployment and services models. Section 3 discusses the security issues of cloud computing with related solutions. Section 4 provides the definition of homomorphic encryption and discusses some examples of existing homomorphic schemes. Section 5 describes how to use homomorphic encryption to secure cloud computing data. Finally, our conclusions are drawn in section 6.
Cloud Computing
Cloud computing can significantly reduce the cost and complexity of owning and operating computers and networks. If an organization uses a cloud provider, it does not need to spend money on information technology infrastructure, or buy hardware or software licenses. Cloud services can often be customized and flexible to use, and providers can offer advanced services that an individual company might not have the money or expertise to develop.
Cloud computing is a subscription-based service where the user can obtain networked storage space and computer resources. To illustrate the idea, cloud computing is similar with the dealing with email client (e.g. Gmail, Yahoo, and so on), the provider of email client provides all of the hardware and software necessary to support the email account. When the user wants to access his email he opens the web browser, and goes to the email client, and makes log in. The most important part of the equation is that the user must have internet access. The email is not housed on physical computer of the user, he accesses it through an internet connection from anywhere. The email client is different than the software installed on the user's computer, such as a word processing program. When the user creates a document using word processing software, that document stays on the device used to make it, unless the user physically moves it. An email client is similar to how cloud computing works. Except instead of accessing just the email services, one can choose what information he have access to within the cloud. The main structure of cloud computing is illustrated in figure 1.
Definition of Cloud Computing
The following definition [1] of cloud computing has been provided by National Institute of Standards and Technology (NIST) of U.S.
"Cloud computing is a model for enabling convenient, ondemand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. This cloud model promotes availability and is composed of five essential characteristics, three service models, and four deployment models".
Types of Deployment Models
There are four deployment models of cloud computing as follows [11]: The Public Cloud: A public cloud can be accessed by any subscriber with an internet connection and access to the cloud space. The Private Cloud: A private cloud is established for a specific group or organization and limits access to just that group. The Community Cloud: A community cloud is shared among two or more organizations that have similar cloud requirements (e.g. security requirements). The Hybrid Cloud: A hybrid cloud is a combination of at least two clouds, where the clouds included are a mixture of public, private, or community.
Types of Service Models
There are three service models of cloud computing as follows [1]: Software as a Service (SaaS): A SaaS provider gives subscribers access to both resources and applications. SaaS makes it unnecessary for the user to have a physical copy of software to install on his devices. SaaS also makes it easier to have the same software on all of user's devices at once by accessing it on the cloud (e.g. web-based email),). In a SaaS agreement, user have the least control over the cloud. Platform as a Service (PaaS): A PaaS provider gives subscribers access to the components that they require to develop and operate applications over the internet. Like using the programming languages, libraries, services, and tools supported by the provider. Infrastructure as a Service (IaaS): An IaaS agreement deals primarily with computational infrastructure. In an IaaS agreement, the subscriber completely outsources the storage and resources, such as hardware and software (e.g. host firewalls).), that they need.
Security of Cloud Computing
The use of cloud computing has increased rapidly in many organizations. Concomitantly, the problems of third party data security and securely outsourcing computation become increasingly prominent. There is the risk that personal information sent to a cloud provider is often seen as valuable to individuals with malicious intent and might be kept indefinitely or used for other purposes. Also, such information could also be accessed by government agencies, domestic or foreign and this might affect the privacy of user.
There are some security issues in cloud computing such as data security, third-party control, and privacy. If all data stored in cloud were encrypted using traditional cryptosystems, this would effectively solve the three above issues [12].
To perform a required computation on encrypted data stored in cloud, a user must share the secret key with cloud provider. First, cloud provider decrypts the data to execute necessary operations then sends the result to the user. To solve this issue, it is necessary to use a cryptosystem based on homomorphic encryption to encrypt the data. Since these cryptosystems allow to do computation on encrypted data.
Homomorphic Encryption
Homomorphic encryption is a type of encryption that allows particular computations to be conducted on ciphertext and return an encrypted result, the decrypted of result is equal the result of conducting the operation on the plaintext. The property of homomorphic is useful to develop a secure evoting system with high privacy data retrieving scheme, also it makes the use of cloud computing by ensuring the privacy of processed data. An example for its mathematical consistency, if there are two numbers 10 and 20 then both are encrypted to 56 and 69 respectively, the addition operator gives a number with value 125, the decrypted of this value is 30.
History of Homomorphic Encryption
The concept of homomorphic encryption was suggested in 1978 by Ronald Rivest and Leonard Adleman [13]. But for 30 years the progress is very slow. In 1982, Shafi Goldwasser and Silvio Micali [14] proposed their encryption system that able to encrypt one bit in additive homomorphic encryption. Pascal Paillier 1999 [15] suggested another additive homomorphic encryption. In 2005, Dan Boneh, Eu-Jin Goh and Kobi [16] invented a security system of encryption which conduct only single multiplication but large number of additions. In 2009 [3], Craig Gentry construct a fully homomorphic encryption based system that able to conduct both of addition and multiplication in the same time.
Categories of Homomorphic Encryption
There are two main categories of homomorphic encryption schemes: Partially Homomorphic Encryption (PHE) and Fully Homomorphic Encryption (FHE) schemes. PHE schemes, such as RSA, ElGamal, Paillier, Etc., allow to perform either addition or multiplication on encrypted data. Construction of scheme supporting both operations simultaneously was elusive. Although Boneh et al. [8] came closest, allowing unlimited additions and a single multiplication, It was not until 2009 that the three decade old problem was solved in seminal work by Gentry [3], where he showed that performing both addition and multiplication simultaneously are possible in fully homomorphic encryption.
Partially Homomorphic Encryption
A. Multiplicative Homomorphic Schemes A Homomorphic Encryption is multiplicative, if there is an algorithm that can calculate Enc(x × y) from Enc (x) and Enc (y) without knowing x and y [17]. Such as RSA and ElGamal Algorithms. Figure 2 illustrates the RSA algorithm as an example of multiplicative homomorphic schemes [18]. The multiplicative homomorphic property of RSA scheme is as follows [19]: (mod n)= .
B. Additive Homomorphic Schemes
A Homomorphic Encryption is additive, if there is an algorithm that can calculate Enc(x + y) from Enc(x) and Enc (y) without knowing x and y [17]. Such as Paillier and Goldwasser-Micali algorithms. Figure 3 illustrates the Paillier algorithm as an example of addtive homomorphic schemes [15]. The homomorphic property of Paillier scheme can be shown as follows [12]:
Fully Homomorphic Encryption
All of PHE schemes allow homomorphic computation of only one operation, either addition or multiplication, on encrypted date, except the Boneh-Goh-Nissim scheme which supports performing unlimited number of addition operation but only one multiplication. The constructing of a scheme that allows one to compute arbitrary computation (a scheme should allow an unlimited number of both addition and multiplication operations) over encrypted data has remained a central open problem in cryptography for more than 30 years and thought to be impossible until 2009, when Craig Gentry proposed the first plausible construction of a fully homomorphic scheme [3].
Gentry's work is supporting multiplication and addition in the same time, correspond to AND (∧) and XOR (⊕) in Boolean algebra. The remarkable value of supporting these two Boolean functions is that any computation can be converted into a function that contains only (∧) and (⊕). In algebra, there are several techniques can be used to convert a function into more simple. By using this techniques can be convert a function to use only specific Boolean operation (e.g. ∧ or ⊕). For example ¬A can be expressed as A ⊕ 1, another example is A ∨ B, this can be converted into (¬A) ∧ (¬B), then converted into (A ⊕ 1) ^ (B ⊕ 1). By utilizing such techniques, all functions can be converted into a series of (∧) and (⊕) operations. This is the basis of Gentry's work [19]. Figure 4 illustrates the difference between the conventional encryption schemes (not PHE) and fully homomorphic scheme.
Gentry is using lattice-based cryptography. His proposed fully homomorphic encryption consists of several steps: start from what was called a somewhat homomorphic encryption scheme using ideal lattices that is limited to evaluating lowdegree polynomials over encrypted data. It is limited because each ciphertext is noisy in some sense, and this noise grows as one adds and multiplies ciphertexts, until ultimately the noise makes the resulting ciphertext indecipherable. Next, it squashes the decryption procedure so that it can be expressed as a low-degree polynomial which is supported by the scheme. Finally, it applies a bootstrapping transformation, through a recursive self-embedding, to obtain a fully homomorphic scheme [21].
FHE in Cloud Security
The security issues of data stored in cloud can be solved by using Fully Homomorphic Encryption (FHE) schemes. To secure it, the data should be encrypted with FHE before being sent to the cloud. First, the user login and uses the keygeneration provided by the server to generate the secret key, the user is the only holder of this secret key. Then, the user encrypts the data that wants to send it to the cloud. During transmitting, the integrity and non-repudiation can be assured by applying other cryptographic technologies such as digital signature. When the user want the server to execute some computations on these encrypted data (such as search), he can send encrypted request to the cloud server. The server performs the required operations and sent the encrypted result to user. Finally the user decrypt the data with his secret key to retrieve the correct result [22]. Figure 5 illustrates the process of using FHE to cloud computing.
In 2010, M. Dijk, C. Gentry et el. [4] presented a second fully homomorphic encryption (called DGHV). In their scheme, somewhat homomorphic encryption uses addition and multiplication over the integers rather than working with ideal lattices over a polynomial ring. This scheme is conceptually simpler than Gentry's scheme based on ideal lattice, but has similar properties of homomorphic operations and efficiency. This scheme is insecure in cloud computing because the cipher retrieval algorithm R needs to transfer the private key p to the server [6].
In the march 2010, C. Gentry [5] proposed a homomorphic encryption scheme (called Gen10), heading toward widespread use of cloud computing. This scheme is also insecure in cloud computing since its cipher retrieval algorithm R asks to submit q (a random number) to the server, yet utilizing c (ciphertext) and q, the plaintext leaks out [6].
Where r is a random number of R-bit and q is a constain Q-bit big integer.
Decrypt (p, c): Output (c mod p). Retrieval(c): When the user wants to retrieve contents m index , he encrypts the Keywords And delivers c index to the server. On receiving c index , server reads the ciphertexts, computing once R =0, ciphertext retrieval succeeds, and ci is the desired result.
In the SDC scheme, transferring q to the server merely, the server can complete the process of ciphertext retrieval successfully, without plaintext leak out, because the process of decryption uses the private key p while the retrieval process uses the integer q, which is entirely different. Thus satisfies both the demand of ciphertext retrieval and data security [6]. Jian Li, Danjie Song et al. provided a complete proof of correctness of the scheme as follows: Given two messages and . The ciphertext of these messages after encryption: = + p + * p * q.
Then SDC scheme has additively homomorphic property.
Then SDC scheme has multiplicatively homomorphic property.
Conclusion
The security issues are a big problem for cloud computing development. To preserve the privacy of his data, the user must encrypt data before being sent to the cloud. Cloud computing security based on homomorphic encryption schemes, because these schemes allow to perform computations on encrypted data without the need to the secret key. Partially Homomorphic Encryption (PHE) such as RSA and Paillier schemes are insufficient to secure cloud computing because these schemes allow to perform only one operation (either addition or multiplication) on the encrypted data of client. Fully Homomorphic Encryption is the best solution to secure the client data in cloud computing because its schemes enable to perform arbitrary computations on encrypted data without decrypting. DGHV and Gen10 schemes of FHE are insecure when they be used in cloud computing to secure data of client. SDC is a simple and considered efficient scheme to secure data in cloud computing. This paper analyzed some of the existing homomorphic encryption schemes and discussed the use of the most efficient one, SDC scheme, to secure cloud computing data. Future work will focus on implementation of SDC scheme in cloud computing and analysis the complexity of the scheme. | 4,246.2 | 2016-05-12T00:00:00.000 | [
"Computer Science",
"Mathematics"
] |
Prediction Model of Strip Crown in Hot Rolling Process Based on Machine Learning and Industrial Data
: The strip crown in hot rolling has the characteristics of multivariablity, strong coupling and, nonlinearity. It is difficult to describe accurately using a traditional mechanism model. In this paper, based on the industrial data of a hot continuous rolling field, the modeling dataset of a strip steel prediction model is constructed through the collection and collation of the on-site data. According to the classical strip crown control theory, the important process parameters that affect the strip crown are determined as input variables for the data-driven model. Some new intelligent strip crown prediction models integrating the shape control mechanism model, artificial intelligence algorithm, and production data are constructed using four machine learning algorithms, including XGBoost, Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). The overall performance of the models is evaluated using error indicators, such as Mean Absolute Percentage Error (MAPE), Root Mean square Error (RMSE), and Mean Absolute Error (MAE). The research results showed that, for the test set, the determination coefficient (R 2 ) of the predicted value of the strip crown model based on the XGBoost algorithm reached 0.971, and the three error indexes are at the lowest level, meaning that the overall model has the optimal generalization performance, which can realize the accurate prediction of the outlet strip crown in the hot rolling process. The research results can promote the application of industrial data and machine learning modeling to the actual strip shape control process of hot rolling, and also have important practical value for the intelligent preparation of the whole process of steel.
Introduction
The production of hot-rolled strip occupies an important position in the modern iron and steel industry system. Strip shape, mainly including crown and flatness, is a key quality indicator in the production process of hot-rolled strip, and strip shape control is also a challenge in the production process of strip steel [1,2]. Although scholars and enterprises have conducted a large amount of research on this issue, up to now, there are still prominent issues of poor strip shape in the production process of hot-rolled strip. A large number of thin strip coils are rolled with product defects, such as mid wave, side wave, quarter wave, wedge shape, and unqualified crown hit, which seriously affect the yield of hot-rolled strip and bring incalculable economic losses to the enterprise. Poor shape quality of hot-rolled strip not only affects the smooth progress of the rolling process, but also adversely affects the smooth progress of subsequent processes, such as cold rolling, shearing, and other processes [3]. This inevitably requires accurate control of the shape accuracy of hot-rolled strip. The preset model in the shape control model calculates the adjustment amount of each relevant shape adjustment mechanism using a mathematical model based on the relevant parameters, such as the thickness, width, material, reduction, the two prediction models, it was found that the neural network model was more effective and accurate based on microstructure and casting temperature parameters.
In addition, machine learning also has a wide range of applications in the prediction of key parameters in the rolling process and process optimization. Nandan et al. [15] proposed a multi-objective optimal control strategy based on genetic algorithms and applied it to the identification of hot-rolled strip shape parameters and the setting of rolling schedules, which successfully obtained the optimal setting values for the crown and flatness, and also improved the control accuracy of strip shape. Chakraborti et al. [16] used genetic algorithms and ant colony algorithms to optimize the crown of hot-rolled strip, demonstrating the practicality of evolutionary algorithms in optimizing rolling process parameters. John et al. [17] directly used a combination of neural networks and genetic algorithms to establish a relationship model between input parameters and strip shape to predict the minimum strip shape value of hot-rolled strip. Liu et al. [18] established a transfer matrix between the flatness error characteristic parameter and the flatness adjustment parameter using a genetic algorithm to optimize the BP neural network method, and successfully applied the transfer matrix to the flatness adjustment mechanism of a 900 mm six-high HC rolling mill. Peng et al. [19] proposed a new method for recognizing the shape pattern of cold-rolled strip and verified it on an 8000 kN HC mill. The results show that the method can effectively reduce the shape deviation of the strip. Yang et al. [20] established an intelligent collaborative control model for cold-rolled strip flatness control mechanisms based on the combination of a flatness control matrix and differential algorithm optimization extreme learning machine (ELM). Zhang et al. have, respectively, used support vector machines (SVMs) [21], T-S cloud inference neural networks [22], PID neural networks [23], and radial basis networks [24] to study the problem of strip shape pattern recognition in the cold rolling process. Simulation results under various methods show that the various models proposed can identify common defects in strip shape with high accuracy. Deng et al. [25,26] constructed a hybrid model based on the combination of hot strip rolling production data and deep learning networks to predict strip outlet crown, achieving an absolute error of less than 5 µm for 97.04% of the predicted data in the modeling data. Song et al. [27] used machine learning algorithms to establish an accurate prediction model for the strip crown of hot rolling. The model prediction results showed that 97.83% of the data had a difference of less than 4 µm between the actual value and the predicted value. Through the above analysis, it can be found that combining big data technology with artificial intelligence modeling methods is a new trend in studying how to further improve the accuracy of shape control in rolling process [28,29]. Based on the massive industrial data accumulated in the process of hot strip rolling, this paper establishes an accurate prediction model for strip crown through in-depth mining of industrial data, combined with advanced machine learning algorithms, and then constructs a new intelligent shape control preset model that deeply integrates the crown control mechanism model, artificial intelligence algorithms, and production data. As such, it solves the problem that it is difficult to further improve the shape control accuracy when relying on traditional mechanism models, achieves the goal of effectively improving the flatness control accuracy, and obtains high-quality strip steel with good flatness.
The paper is organized as follows. The definition of the strip crown in hot rolling and the influence factors of the strip crown are briefly introduced in Section 2. In Section 3, the process of collecting and processing modeling datasets and the selection of model input variables are described in detail. Section 4 briefly introduces the four machine learning algorithms used for modeling and determines the main parameters of each model. The discussion of the strip crown forecasting results is described explicitly in Section 5, and Section 6 concludes this paper.
Definition of Strip Crown
Strip shape intuitively refers to the degree of warpage of the strip, and essentially refers to the distribution of residual stress inside the strip. The measurement of strip shape usually includes both longitudinal and transverse indicators. The longitudinal direction is represented by flatness, which refers to the flat degree of the strip along the length direction. The transverse direction is expressed by crown, referring to the cross-sectional shape of the strip steel. The standard definition of strip crown is the difference between the center thickness of the strip and the specified edge thickness. To eliminate the effect of strip edge thinning, the edge reference point is usually located 40 mm from the strip edge. The schematic diagram of crown definition is shown in Figure 1. Equation (1) is as follows: where h c is the thickness at the center of the strip, h i is the thickness at the distance operation side at 40 mm, and h i is the thickness at the distance drive side at 40 mm.
Definition of Strip Crown
Strip shape intuitively refers to the degree of warpage of the strip, and essentially refers to the distribution of residual stress inside the strip. The measurement of strip shape usually includes both longitudinal and transverse indicators. The longitudinal direction is represented by flatness, which refers to the flat degree of the strip along the length direction. The transverse direction is expressed by crown, referring to the cross-sectional shape of the strip steel. The standard definition of strip crown is the difference between the center thickness of the strip and the specified edge thickness. To eliminate the effect of strip edge thinning, the edge reference point is usually located 40 mm from the strip edge. The schematic diagram of crown definition is shown in Figure 1. Equation (1) is as follows: where c h is the thickness at the center of the strip, i h is the thickness at the distance operation side at 40 mm, and i h′ is the thickness at the distance drive side at 40 mm.
Influence Factors of Strip Crown
The crown of the strip can be seen as an image of the cross-sectional shape of the roll gap at the exit of the deformation zone. Therefore, all factors that can affect the crosssectional shape of the roll gap at the exit of the deformation zone are factors that affect the strip crown. The strip width, rolling force, work roll diameter, backup roll diameter, crown of work roll, entrance strip crown, and roll bending force are listed as the main factors affecting the strip crown [30,31]. The interrelationship of various influencing factors is shown in Figure 2. The impact effects of these factors are different, and under certain conditions, a certain influencing factor may have a relatively significant effect. As can be seen from Figure 2, there is a non-linear and strong coupling characteristic between the influencing factors affecting the strip outlet crown.
Influence Factors of Strip Crown
The crown of the strip can be seen as an image of the cross-sectional shape of the roll gap at the exit of the deformation zone. Therefore, all factors that can affect the crosssectional shape of the roll gap at the exit of the deformation zone are factors that affect the strip crown. The strip width, rolling force, work roll diameter, backup roll diameter, crown of work roll, entrance strip crown, and roll bending force are listed as the main factors affecting the strip crown [30,31]. The interrelationship of various influencing factors is shown in Figure 2. The impact effects of these factors are different, and under certain conditions, a certain influencing factor may have a relatively significant effect. As can be seen from Figure 2, there is a non-linear and strong coupling characteristic between the influencing factors affecting the strip outlet crown.
Data Collection
The process layout of the hot continuous rolling production line is shown in Figure 3. Generally, the hot continuous rolling production line consists of a heating furnace, roughing mill, finishing mill, laminar cooling device, and coiling. The roughing mill unit includes a high-pressure water descaler, a vertical mill (E1), and a roughing mill (R1). The vertical mill (E1) controls the width of the strip, and the roughing mill controls the thickness of the strip when it enters the finishing mill unit. The semi-finished products from the rough rolling area are sent to the finishing mill area through the roller table. The finishing mill unit includes a high-pressure water descaler, a flying shear, and a seven-stand finishing mill (F1~F7). After the strip comes out of the finishing mill, it enters the laminar cooling device, which controls the cooling speed and final coiling temperature of the strip. The strip is rolled into coils by a coiler. During the actual production process, the automatic control system monitors and records the relevant data of each equipment and strip steel in real time. The data recorded in the database is divided into original data, calculated value data, and actual value data.
Data Collection
The process layout of the hot continuous rolling production line is shown in Figure 3. Generally, the hot continuous rolling production line consists of a heating furnace, roughing mill, finishing mill, laminar cooling device, and coiling. The roughing mill unit includes a high-pressure water descaler, a vertical mill (E1), and a roughing mill (R1). The vertical mill (E1) controls the width of the strip, and the roughing mill controls the thickness of the strip when it enters the finishing mill unit. The semi-finished products from the rough rolling area are sent to the finishing mill area through the roller table. The finishing mill unit includes a high-pressure water descaler, a flying shear, and a seven-stand finishing mill (F1~F7). After the strip comes out of the finishing mill, it enters the laminar cooling device, which controls the cooling speed and final coiling temperature of the strip. The strip is rolled into coils by a coiler.
Data Collection
The process layout of the hot continuous rolling production line is shown in Figure 3. Generally, the hot continuous rolling production line consists of a heating furnace, roughing mill, finishing mill, laminar cooling device, and coiling. The roughing mill unit includes a high-pressure water descaler, a vertical mill (E1), and a roughing mill (R1). The vertical mill (E1) controls the width of the strip, and the roughing mill controls the thickness of the strip when it enters the finishing mill unit. The semi-finished products from the rough rolling area are sent to the finishing mill area through the roller table. The finishing mill unit includes a high-pressure water descaler, a flying shear, and a seven-stand finishing mill (F1~F7). After the strip comes out of the finishing mill, it enters the laminar cooling device, which controls the cooling speed and final coiling temperature of the strip. The strip is rolled into coils by a coiler. During the actual production process, the automatic control system monitors and records the relevant data of each equipment and strip steel in real time. The data recorded in the database is divided into original data, calculated value data, and actual value data. The original data includes strip steel grade, material, slab size, and finished product size. These data are necessary parameters for setting calculation and automatic control. The During the actual production process, the automatic control system monitors and records the relevant data of each equipment and strip steel in real time. The data recorded in the database is divided into original data, calculated value data, and actual value data. The original data includes strip steel grade, material, slab size, and finished product size.
These data are necessary parameters for setting calculation and automatic control. The calculated value data includes various process parameters and model calculation data during the production process, which are important for the accuracy and quality of the finished product and for the analysis of the process. The actual value data includes the measured data of various pieces of testing equipment during the production process, which is an important basis for adaptive learning and the dynamic correction of the rolling mathematical model. The data collection process is shown in Figure 4. calculated value data includes various process parameters and model calculation data during the production process, which are important for the accuracy and quality of the finished product and for the analysis of the process. The actual value data includes the measured data of various pieces of testing equipment during the production process, which is an important basis for adaptive learning and the dynamic correction of the rolling mathematical model. The data collection process is shown in Figure 4.
Modeling Datasets
Roll
Data Preprocessing
The data were collected on a real 1780 mm hot strip mill process production line of HBIS Group Co., Ltd., which is located in Hebei Province, China. Pre-processing of collected data included deleting sample data with missing values and eliminating outlier data. After data processing, 1809 pieces of strip steel sample data were finally obtained to form a modeling dataset. Part of the modeling data is shown in Table 1. The modeling dataset is divided into the training set and test set based on the sampling balance. The training set proportion was 70%, and the test set proportion was 30%. In order to eliminate the impact of large differences in magnitude in different dimensions of sample data on model accuracy during modeling, the data of all input variables were standardized [32,33]. The formula for standardization processing is as follows: x are the maximum and minimum numbers of data sequences, respectively.
Data Preprocessing
The data were collected on a real 1780 mm hot strip mill process production line of HBIS Group Co., Ltd., which is located in Hebei Province, China. Pre-processing of collected data included deleting sample data with missing values and eliminating outlier data. After data processing, 1809 pieces of strip steel sample data were finally obtained to form a modeling dataset. Part of the modeling data is shown in Table 1. The modeling dataset is divided into the training set and test set based on the sampling balance. The training set proportion was 70%, and the test set proportion was 30%. In order to eliminate the impact of large differences in magnitude in different dimensions of sample data on model accuracy during modeling, the data of all input variables were standardized [32,33]. The formula for standardization processing is as follows: where max(x i ) and min(x i ) are the maximum and minimum numbers of data sequences, respectively.
Determination of the Input and Output Parameters of the Models
Based on the analysis of the influencing factors of strip outlet crown in the theory of strip shape control, the input variables of the machine learning model were finally determined according to the plate shape control theory. These input variables included strip width (W), slab thickness (H), exit thickness (H1~H7), entrance temperature (T1~T7), exit temperature (t1~t7), rolling force (F1~F7), rolling speed (V1~V7), strip yield strength (Q1~Q7), bending force (W1~W7), rolling shifting (S1~S7), roll diameter (D1~D7), roll thermal expansion (C1~C7), and roll wear (M1~M7). The specific input variables are shown in Figure 5. The strip crown is used as the model output variable.
Determination of the Input and Output Parameters of the Models
Based on the analysis of the influencing factors of strip outlet crown in the theory of strip shape control, the input variables of the machine learning model were finally determined according to the plate shape control theory. These input variables included strip width (W), slab thickness (H), exit thickness (H1~H7), entrance temperature (T1~T7), exit temperature (t1~t7), rolling force (F1~F7), rolling speed (V1~V7), strip yield strength (Q1~Q7), bending force (W1~W7), rolling shifting (S1~S7), roll diameter (D1~D7), roll thermal expansion (C1~C7), and roll wear (M1~M7). The specific input variables are shown in Figure 5. The strip crown is used as the model output variable.
Experimental
ANNs are one of the most classic machine learning algorithms, widely used in various fields due to their strong nonlinear fitting ability and their ability to map any nonlinear relationship. Compared with ANNs, SVMs have a more solid mathematical theoretical foundation and can effectively solve the problem of constructing high-dimensional data models under limited sample conditions. Separate ANNs and SVMs are typical individual learners. The method of constructing and combining multiple individual learners to complete learning tasks is called ensemble learning. Random forest (RF) and XGBoost are two different ensemble learning strategies. RF is an ensemble learning method based on the idea of bagging, which obtains multiple parallel individual learners through effective resampling, and then calculates the model mean as the final model result. XGBoost is an ensemble learning method based on Boosting's idea, which continuously adjusts the weights of individual learners according to the error rate during the training process, so that individual learners with low error probability can obtain larger weights. The above four models are the most representative methods of machine learning. Therefore, these
Experimental
ANNs are one of the most classic machine learning algorithms, widely used in various fields due to their strong nonlinear fitting ability and their ability to map any nonlinear relationship. Compared with ANNs, SVMs have a more solid mathematical theoretical foundation and can effectively solve the problem of constructing high-dimensional data models under limited sample conditions. Separate ANNs and SVMs are typical individual learners. The method of constructing and combining multiple individual learners to complete learning tasks is called ensemble learning. Random forest (RF) and XGBoost are two different ensemble learning strategies. RF is an ensemble learning method based on the idea of bagging, which obtains multiple parallel individual learners through effective resampling, and then calculates the model mean as the final model result. XGBoost is an ensemble learning method based on Boosting's idea, which continuously adjusts the weights of individual learners according to the error rate during the training process, so that individual learners with low error probability can obtain larger weights. The above four models are the most representative methods of machine learning. Therefore, these four methods are used to establish a strip crown prediction model based on the dataset in this article and to conduct comparative research to obtain the best strip crown prediction model.
MLP-Based Method
The multilayer perceptron model (MLP) is an ANN. An important feature of the MLP is that it has multiple neural layers. The MLP model consists of an input layer, hidden layer, and an output layer, with each layer consisting of multiple neurons [34,35]. According to the modeling dataset, the number of neurons in the input layer of the constructed MLP crown prediction model is equal to the number of input feature variables in the dataset. Therefore, the number of neurons in the input layer is 79, and the number of neurons in the output layer is equal to the number of output variables in the dataset, which means that the number of neurons in the output layer is 1. After testing the hidden layer's activation function of 'identity', 'logistic', 'tanh', and 'relu', it is found that the 'logistic' function has the best prediction accuracy. In addition, the number of neurons in the hidden layer [36] and the regularization term coefficients are important parameters that affect the accuracy of the MLP model. The MLP model adopts 'Python sklearn.neutral_network. MLPRegressor' generation, using 'GridSearchCV' to optimize the number of hidden layer neurons and regularization term coefficients (alpha). The optimal range of the number of neurons in the hidden layer is [80,100,200], and the optimal range of the 'alpha' is [0.01, 0.1, 0.5]. Under the dataset in this article, after optimizing the model parameters, the final parameters of the MLP regression model are shown in Table 2.
SVM-Based Method
The SVM is a new type of machine learning algorithm based on statistical learning theory, which is an approximate implementation of structural risk minimization methods [37,38]. It is suitable for small sample data. Through some nonlinear mapping, the input variables are dimensioned up to a high dimensional space. An SVM has the advantages of fast solution speed and strong generalization ability, and it is widely used in many fields. An SVM can ultimately be transformed into dual optimization forms, as follows: The constraints are as follows: where n is the number of samples; α i α * i are the Lagrange multipliers; K x i , x j is the kernel function, and the commonly used kernel functions include linear kernel, polynomial kernel, and radial basis functions (RBF). Due to the strong nonlinear mapping ability of RBF, this paper uses polynomial kernel as the kernel function. Polynomial kernel is expressed as follows: where a is kernel function parameter. The number of polynomial kernel functions 'degree', penalty factor 'C', and the coefficient of the kernel function 'gamma' of the polynomial kernel function in SVM models have important impacts on the prediction accuracy of the model. The SVM model adopts 'Python sklearn.svm import SVR' generation, using 'GridSearchCV' to optimize the 'degree', 'C', and 'gamma'. The optimal range of the 'degree' is [1,2,3], the 'C' is [20,30,40], and the 'gamma' is [0.1, 0.2, 0.23]. After testing the kernel function type of 'linear', 'poly', 'rbf', and 'sigmoid', it is found that the 'poly' function has the best prediction accuracy. Under the dataset in this article, the parameters of the SVM finally determined after parameter optimization are shown in Table 3.
RF-Based Method
RF is a parallel integrated learning algorithm based on decision trees. Compared to individual base learners, it achieves randomness in sample selection and feature selection through bagging and random subspace ideas [39,40], enhancing the generalization ability of models. According to the idea of bagging, the RF model obtains m sets of sampled training sets with the same capacity as the original training set through independent random sampling with m times of return, and then uses these sampled training sets to train in order to obtain corresponding m base learners. Due to the independence of sampling, the sampled training set obtained from each sampling is different from the original training set and other sampled training sets, which can effectively avoid the occurrence of local optimal solutions from the perspective of training sample selection and can also ensure a low correlation between each decision subtree.
The number of base learners 'n_estimators', the depth of decision tree 'max_depth', and the number of features to consider when limiting branching 'max_features' in the RF model have important impacts on the prediction accuracy of the model. The RF model adopts 'Python sklearn.ensemble import RandomForestRegressor' generation, using 'GridSearchCV' to optimize the 'n_estimators', 'max_depth', and 'max_features'. The optimal range of the 'n_estimators' is [60, 80, 100], the 'max_depth' is [10,30,50], and the 'max_features' is [6,9,12]. Under the dataset in this article, the parameters of the RF finally determined after parameter optimization are shown in Table 4.
XGBoost-Based Method
XGBoost is a flexible, efficient, and convenient optimal distributed algorithm proposed by Chen et al. [41][42][43], based on the Gradient Boosting Decision Tree (GBDT). XGBoost adopts the idea of boosting ensemble learning to combine multiple decision trees to achieve better results and make the combined model more generalized. XGBoost is composed of multiple decision trees. Each decision tree learns the residual between the target value and the predicted value, where the predicted value is the sum of the predicted values of all previous decision trees. After the training of all decision trees is completed, a common decision is made. After obtaining the corresponding prediction values on each tree, the samples are accumulated as their final prediction results. During the training phase, each new tree is trained on the basis of the completed tree. Each decision tree is a weak learner. Boosting technology is used to upgrade all weak learners into strong learners. To avoid model overfitting and enhance generalization capabilities, XGBoost adds regularization terms to the loss function of the GBDT model. The traditional GBDT calculates the loss function using a first order Taylor expansion, using a negative gradient value to replace the residual for fitting. XGBoost adds a second order Taylor expansion to the loss function, using second order derivatives to collect gradient direction information, thereby improving the accuracy of the model.
The loss function of the XGBoost algorithm is defined as follows: where l y i ,ŷ is the prediction residual of the ith sample to the tth iteration; X i is the ith sample; Ω( f t ) is the regularization term; ω 2 j is the score of leaf node; T is the number of leaf nodes; γ is the coefficient; λ is the coefficient of the sum of the weights regularized by the L2 regularization term of all leaf nodes.
The second-order Taylor expansion of the loss function of XGBoost algorithm is as follows: where g i is the first derivative term of the loss function; h i is the second derivative term of the loss function. The number of base learners 'n_estimators', the depth of decision tree 'max_depth', and the 'learning_rate' in XGBoost model have important impacts on the prediction accuracy of the model. The XGBoost model adopts 'Python XGBoost import XGBRegressor' generation, using 'GridSearchCV' to optimize the 'n_estimators', 'max_depth', and 'learn-ing_rate'. The optimal range of the 'n_estimators' is [60, 80, 100], the 'max_depth' is [10,30,50], and the 'learning_rate' is [0.1, 0.2, 0.3]. Under the dataset in this article, the parameters of the XGBoost finally determined after parameter optimization are shown in Table 5.
Results and Discussion
The experiments are carried out on a computer running 64-bit Windows 10 with a 3.0 GHz processor and 16 GB memory. The algorithms are implemented in the PYTHON language. This research adopted R 2 , MAE, MAPE, and RMSE as criteria for assessing the prediction performance of machine learning models. The performance criteria calculation formulas are as follows: where n denotes the number of sample data; y i and y * i are the measured value and the predicted value of the ith sample, respectively.
The modeling dataset is divided into a training set and a test set. Four machine learning algorithms, namely MLP, SVM, RF, and XGBoost, are used to train the strip crown prediction model for hot strip rolling based on the same training set data, and the performance of the established models are tested using the same test set. The regression effects of the four prediction models are shown in Figure 6. According to Figure 6 In order to visually display the comparison between the predicted values and the actual values of the four crown prediction models, Figure 7 is drawn to express the proximity of the predicted values to the actual values. Due to the large amount of modeling sample data, it is not possible to display the situation of all samples in one graph. There- In order to visually display the comparison between the predicted values and the actual values of the four crown prediction models, Figure 7 is drawn to express the proximity of the predicted values to the actual values. Due to the large amount of modeling sample data, it is not possible to display the situation of all samples in one graph. Therefore, the first 100 samples from the training set and the test set are taken for plotting. As can be seen from Figure 7a, in the training set, the predicted values of each model have a greater degree of consistency with the actual values, because all the samples in the training set participated in the training of the model during the modeling process. On the contrary, there are many samples in the test set that have significant errors with the actual values, especially in the prediction results of the MLP and SVM models, where a large number of samples seriously deviate from the actual values of the corresponding crown. There are two reasons for this result. Firstly, compared to the training set, the test set samples did not participate in the model construction, resulting in a lower accuracy for its predicted values than the training set sample. In addition, because the generalization performance of MLP and SVM models is inferior to that of XGBoost and RF models, the prediction effect shown on the test set is that there are more samples with prediction results that seriously deviate from the actual values. In order to visually display the comparison between the predicted values and the actual values of the four crown prediction models, Figure 7 is drawn to express the proximity of the predicted values to the actual values. Due to the large amount of modeling sample data, it is not possible to display the situation of all samples in one graph. Therefore, the first 100 samples from the training set and the test set are taken for plotting. As can be seen from Figure 7a, in the training set, the predicted values of each model have a greater degree of consistency with the actual values, because all the samples in the training set participated in the training of the model during the modeling process. On the contrary, there are many samples in the test set that have significant errors with the actual values, especially in the prediction results of the MLP and SVM models, where a large number of samples seriously deviate from the actual values of the corresponding crown. There are two reasons for this result. Firstly, compared to the training set, the test set samples did not participate in the model construction, resulting in a lower accuracy for its predicted values than the training set sample. In addition, because the generalization performance of MLP and SVM models is inferior to that of XGBoost and RF models, the prediction effect shown on the test set is that there are more samples with prediction results that seriously deviate from the actual values. From the perspective of quantitative analysis, the absolute error between the predicted value and the actual value of each model in the training set and test set is statistically analyzed, and the statistical results are based on the reference standard that the absolute error of the predicted crown value is less than 4 µm. The results are shown in Table 6. From the table, it can be seen that the absolute error of XGBoost model's prediction From the perspective of quantitative analysis, the absolute error between the predicted value and the actual value of each model in the training set and test set is statistically analyzed, and the statistical results are based on the reference standard that the absolute error of the predicted crown value is less than 4 µm. The results are shown in Table 6. From the table, it can be seen that the absolute error of XGBoost model's prediction results on the training set is all less than 4 µm. The proportion of samples with an absolute error of less than 4 µm on the test set reached 96.13%, with only a small percentage having an absolute error of greater than 4 µm. The sample proportions of other models within this error range are smaller than that of XGBoost model, and the order of the ratio from large to small is RF, SVM, and MLP. The absolute error frequency distribution histograms and corresponding normal distribution curves of the four models' predictions are shown in Figure 8. It can be seen from Figure 8 that compared with the other three prediction models, the prediction error of XGBoost model is more concentrated around the error of zero. With the increase in absolute error, the frequency gradually decreases and presents a normal distribution. It is fully proved that XGBoost prediction model has a good prediction effect. In order to evaluate the generalization performance more comprehensively and quantitatively, three errors, RMSE, MAE, and MAPE, are used as error indicators to analyze the above four models. Table 7 lists the calculated values of the three error indicators of each model. Figure 9 shows the error distribution histogram drawn according to the calculation results. When comparing the MLP model, RF model, SVM model, and XGBoost model, the result shows that the prediction accuracy of the XGBoost model is significantly better than the other models. In order to evaluate the generalization performance more comprehensively and quantitatively, three errors, RMSE, MAE, and MAPE, are used as error indicators to analyze the above four models. Table 7 lists the calculated values of the three error indicators of each model. Figure 9 shows the error distribution histogram drawn according to the calculation results. When comparing the MLP model, RF model, SVM model, and XGBoost model, the result shows that the prediction accuracy of the XGBoost model is significantly better than the other models. In the training set, with the results presented in Figure 9a, the XGBoost model is used to predict the hot strip crown, and the calculation results of the RMSE, MAE, and MAPE error indicators are smaller than the others. For the data-driven model, the error on the test set shows the generalization performance of the established model. Figure 9b 5.308 and 6.582, respectively. The calculation results of the three errors show the same rule, that is, that the XGBoost model shows the best prediction accuracy on the test set while maintaining the best training effect of the model. The reason for the above results is that the MLP model uses the BP algorithm for training. The traditional BP algorithm is a local search optimization method, and the weights of the network are gradually adjusted along the direction of local improvement, which can cause the algorithm to fall into local extremum [44], which often leads to the model falling into an over-fitting state, and the training time of the model will increase exponentially with the increase in the number of hidden layers and the number of neurons in each hidden layer. The time spent in model training is shown in Figure 10. Compared with the other three models, MLP has the longest training time. This seriously affects the training efficiency of the MLP model. The SVM algorithm uses quadratic programming to solve support vectors, which involves designing the cal- In the training set, with the results presented in Figure 9a, the XGBoost model is used to predict the hot strip crown, and the calculation results of the RMSE, MAE, and MAPE error indicators are smaller than the others. For the data-driven model, the error on the test set shows the generalization performance of the established model. Figure 9b 6.582, respectively. The calculation results of the three errors show the same rule, that is, that the XGBoost model shows the best prediction accuracy on the test set while maintaining the best training effect of the model. The reason for the above results is that the MLP model uses the BP algorithm for training. The traditional BP algorithm is a local search optimization method, and the weights of the network are gradually adjusted along the direction of local improvement, which can cause the algorithm to fall into local extremum [44], which often leads to the model falling into an over-fitting state, and the training time of the model will increase exponentially with the increase in the number of hidden layers and the number of neurons in each hidden layer. The time spent in model training is shown in Figure 10. Compared with the other three models, MLP has the longest training time. This seriously affects the training efficiency of the MLP model. The SVM algorithm uses quadratic programming to solve support vectors, which involves designing the calculation of m-order matrices. Therefore, when the matrix order is large, it will cause a decrease in generalization performance and consume a large amount of machine memory and computing time. This also explains why the two algorithms in Figure 6b,d have large absolute errors on the test set. RF and XGBoost all belong to the ensemble learning model. Ensemble learning integrates multiple models through a certain strategy and improves the accuracy of decision-making through group decision-making. The training of the RF model can be highly parallelized, fast, and efficient, which can be proved by Figure 10. The XGBoost model shows better comprehensive performance than RF. Based on the above analysis, it can be concluded that the XGBoost model is most suitable for establishing a prediction model of strip crown in hot rolling under the dataset of this research.
Conclusions
In this paper, some new data-driven strip crown prediction models integrating the shape control mechanism model, artificial intelligence algorithms, and production data are constructed using the XGBoost, RF, SVM, and MLP algorithms. Through the analysis and evaluation of the prediction results of strip crown for each model, the following main conclusions can be drawn. Using four machine learning algorithms and combining indus trial data, the prediction models for hot-rolled strip crown are constructed. Under the same dataset, the XGBoost model has the highest coefficient of determination (R 2 ) for the prediction results, reaching 0.971 on the test set, and the MLP model has the lowest coef ficient of determination for the prediction results, which is 0.860 on the test set. The RF and SVM models are between the XGBoost model and MLP model, with 0.945 and 0.909 on the test set, respectively. The comprehensive performance of four crown prediction models is evaluated using MAE, RMSE, and MAPE. The results show that the prediction model based on the XGBoost algorithm has the smallest errors under the same modeling dataset, showing the best prediction performance and the best generalization perfor mance. Based on these advantages, the combination of the XGBoost algorithm and indus trial data can be used to effectively predict the strip crown. By generating corresponding training data during the rolling process, those data-driven prediction methods can easily be extended to predict and optimize other parameters. The research in this paper provides a new method for solving complex industrial problems with multiple variables, strong coupling, and nonlinearity that cannot be handled by traditional mathematical models and also provides technical support for the effective utilization of massive data and shape control in hot strip rolling.
Author Contributions: Z.W. contributed to the conception of the study, design of experiments, anal ysis, and manuscript preparation; Y.H. helped perform the analysis with constructive discussions Y.L. contributed to write, review, and edit; T.W.'s contribution was investigation, supervision, and
Conclusions
In this paper, some new data-driven strip crown prediction models integrating the shape control mechanism model, artificial intelligence algorithms, and production data are constructed using the XGBoost, RF, SVM, and MLP algorithms. Through the analysis and evaluation of the prediction results of strip crown for each model, the following main conclusions can be drawn. Using four machine learning algorithms and combining industrial data, the prediction models for hot-rolled strip crown are constructed. Under the same dataset, the XGBoost model has the highest coefficient of determination (R 2 ) for the prediction results, reaching 0.971 on the test set, and the MLP model has the lowest coefficient of determination for the prediction results, which is 0.860 on the test set. The RF and SVM models are between the XGBoost model and MLP model, with 0.945 and 0.909 on the test set, respectively. The comprehensive performance of four crown prediction models is evaluated using MAE, RMSE, and MAPE. The results show that the prediction model based on the XGBoost algorithm has the smallest errors under the same modeling dataset, showing the best prediction performance and the best generalization performance. Based on these advantages, the combination of the XGBoost algorithm and industrial data can be used to effectively predict the strip crown. By generating corresponding training data during the rolling process, those data-driven prediction methods can easily be extended to predict and optimize other parameters. The research in this paper provides a new method for solving complex industrial problems with multiple variables, strong coupling, and nonlinearity that cannot be handled by traditional mathematical models, and also provides technical support for the effective utilization of massive data and shape control in hot strip rolling. | 9,990.4 | 2023-05-06T00:00:00.000 | [
"Computer Science"
] |
From Gigabyte to Kilobyte: A Bioinformatics Protocol for Mining Large RNA-Seq Transcriptomics Data
RNA-Seq techniques generate hundreds of millions of short RNA reads using next-generation sequencing (NGS). These RNA reads can be mapped to reference genomes to investigate changes of gene expression but improved procedures for mining large RNA-Seq datasets to extract valuable biological knowledge are needed. RNAMiner—a multi-level bioinformatics protocol and pipeline—has been developed for such datasets. It includes five steps: Mapping RNA-Seq reads to a reference genome, calculating gene expression values, identifying differentially expressed genes, predicting gene functions, and constructing gene regulatory networks. To demonstrate its utility, we applied RNAMiner to datasets generated from Human, Mouse, Arabidopsis thaliana, and Drosophila melanogaster cells, and successfully identified differentially expressed genes, clustered them into cohesive functional groups, and constructed novel gene regulatory networks. The RNAMiner web service is available at http://calla.rnet.missouri.edu/rnaminer/index.html.
Introduction
Transcriptome analysis is essential for determining the relationship between the information encoded in a genome, its expression, and phenotypic variation [1,2]. Next-generation sequencing (NGS) of RNAs (RNA-Seq) has emerged as a powerful approach for transcriptome analysis [3,4] that has many advantages over microarray technologies [5,6,7].
A RNA-Seq experiment typically generates hundreds of millions of short reads that are mapped to reference genomes and counted as a measure of expression [5]. Mining the gigabytes or even terabytes of RNA-Seq raw data is an essential, but challenging step in the analysis.
In order to address these challenges, RNAMiner has been developed to convert gigabytes of raw RNA-Seq data into kilobytes of valuable biological knowledge through a five-step data mining and knowledge discovery process. RNAMiner integrates both public tools (e.g., TopHat2 [8], Bowtie2 [9], Cufflinks [10], HTSeq [11], edgeR [12], and DESeq2 [13]) with our in-house tools (MULTICOM-MAP [14,15,16]) to preprocess data and identify differentially expressed genes in the first three steps. In the last two steps, RNAMiner uses our in-house tools MULTI-COM-PDCN [17,18] and MULTICOM-GNET [19,20] to predict both functions and gene regulatory networks of differentially expressed genes, respectively.
As proof of principle, we have applied the RNAMiner protocol to RNA-Seq data generated from Human, Mouse, Arabidopsis thaliana, and Drosophila melanogaster cells. The data mining process successfully produced valuable biological knowledge such as differentially expressed genes, cohesive functional gene groups, and novel hypothetical gene regulatory networks by reducing the size of the initial data set over a thousand-fold.
Methods
Some RNA-Seq data analysis pipelines (e.g. Galaxy [21], KBase [22], iPlant [23]) provide users with a convenient and free platform for RNA-Seq data analysis by combing public tools, such as TopHat [24], Bowtie [25], Cufflinks [10], Cuffmerge [10], and Cuffdiff [10]. As with these pipelines, RNAMiner combines these public tools such as TopHat2 [8], Bowtie2 [9], Cufflinks [10], Cuffdiff [10], and it is free. However, there are several differences between RNAMiner and other pipelines. First, RNAMiner integrates more tools, such as HTSeq [11], edgeR [12], DESeq2 [13], and our in-house MULTICOM-MAP [14,15,16], to calculate gene expression values and identify differentially expressed genes. These tools can generate more accurate consensus results. For example, RNAMiner uses Cuffdiff, edgeR, and DESeq2 to identify differentially expressed genes based on TopHat mapping results and gene expression values calculated by HTSeq and MULTICOM-MAP. RNAMiner generates up to five distinct lists and one consensus list of differentially expressed genes, which usually produces more accurate results. Second, RNAMiner predicts functions of differentially expressed genes and builds gene regulatory networks by integrating our in-house tools MULTICOM-PDCN [17,18] and MULTICOM-G-NET [19,20]. These analyses provide more biological information. Other pipelines (e.g. Galaxy and iPlant) do not provide these analyses. Another software package-KBase-contains a service to predict gene functions, but the service only provides GO annotation for plant genomes. Third, without requirements for user registration and selection of many parameters, RNAMiner is easier to use than other pipelines. Compared to running each tool separately, users can easily run all these tools integrated in RNAMiner at one time and download results generated by all the tools at the RNAMiner web site.
The five data analysis steps of the RNAMiner protocol (Fig 1) are described individually in sub-sections below. Tables 1 and 2 list the versions and the parameters of all the public tools used in RNAMiner and describe the meanings of the parameters.
Mapping RNA-Seq reads to a reference genome
We use two public tools, TopHat2 [8] and Bowtie2 [9], to map RNA-Seq reads to reference genomes in the UCSC genome browser [26] in conjunction with the RefSeq genome reference annotations [27]. The workflow of mapping RNA-Seq reads to a reference genome and calculating gene expression values is illustrated in Fig 2. It is worth noting that, since the RefSeq genome reference annotations contain information about some non-coding small RNAs, the reads of the non-coding RNAs are mapped and counted in addition to regular protein coding mRNAs. MULTICOM-MAP [14,15,16] is used to remove reads mapped to multiple locations in a reference genome from the mapping data in BAM/SAM format [28] generated by TopHat2 and Bowtie2. Only reads mapped to a unique location on the genome are retained to calculate the read counts of the genes. We use MULTICOM-MAP to analyze the mapping results to The RNAMiner protocol for big transcriptomics data analysis. Five blue boxes denote five data analysis steps, i.e. mapping RNA-Seq reads to a reference genome, calculating gene expression values, identifying differentially expressed genes, predicting gene functions, and constructing gene regulatory networks. The tools used in each step are listed inside each box. The external input information is represented by brown boxes and the final output information is represented by green boxes. The information flow between these components is denoted by arrows. doi:10.1371/journal.pone.0125000.g001 Mining Large RNA-Seq Data obtain baseline information, such as the total number of reads, the number of reads mapped to a unique location, and the number of reads mapped to multiple locations. This mapping process can generally reduce the size of datasets by several orders of magnitude. TopHat2 -read-mismatches 2 The maximum number of mismatched nucleotides between a read and a reference allowed for a valid mapping.
-read-gap-length 2 The maximum number of gaps in the alignment between a read and a reference genome allowed for a valid mapping.
-splicemismatches 0 The maximum number of mismatches allowed in the "anchor" region of a spliced alignment.
-segmentmismatches 2 Read segments are mapped independently, allowing up to this number of mismatches in each segment alignment.
-segment-length 25 A read is cut into segments each having at least this length. These segments are mapped independently.
Bowtie2
-end-to-end In this mode, Bowtie2 requires that the entire read to be aligned from one end to the other, without any trimming (or "soft clipping") of characters from either end. Local alignment is not used in Bowtie2.
-sensitive This option generally balances speed, sensitivity and accuracy.
-frag-len-std-dev 80 The standard deviation of fragment lengths.
-max-intron-length 300000 Ignore alignments with gaps longer than this.
Cuffdiff -min-alignmentcount 10 Minimum number of alignments in a locus for testing.
-FDR 0.05 The maximum false discovery rate allowed after statistical correction.
-frag-len-std-dev 80 The standard deviation of fragment lengths.
HTSeq -a 10 Skip all reads with alignment quality lower than this minimum value.
-t Exon Feature type (3rd column in GFF file) to be used. All the features of other types are ignored.
-i gene_id GFF attribute to be used as feature ID.
-m Union Mode to handle reads overlapping more than one feature.
DESeq2
Test LRT Use the likelihood ratio test on the difference in deviation between a full and reduced model formula (nbinomLRT).
fitType parametric The type of fitting of dispersions to the mean intensity. Parametric: fit a dispersion-mean relation via a robust gamma-family GLM.
Calculating gene expression values
For RNAMiner, MULTICOM-MAP [14,15,16] and two public tools: HTSeq [11] and Cufflinks [10] are used to calculate gene expression values according to the genome mapping output and the RefSeq genome reference annotation [27]. MULTICOM-MAP and HTSeq produce raw read counts, while Cufflinks generates normalized values in terms of FPKM, i.e., fragments per kilobase of exon model per million mapped fragments. The normalized gene expression values generated by Cufflinks are used to identify differentially expressed genes in the next step. The read counts generated by MULTICOM-MAP and HTSeq are fed separately into two R Bioconductor packages, edgeR [12] and DESeq2 [13], to identify differentially expressed genes. The estimated probability that the reads were derived from the isoform. In contrast, MULTICOM--MAP distributes the total count of such reads to each isoform, while HTSeq discards the reads without counting them for any isoform. This analysis step generates the overall expression profile of most genes in a transcriptome and can reduce the size of data from Step 1 by~one thousand-fold, from gigabytes to several megabytes.
Identifying differentially expressed genes
We use Cuffdiff [10] and two R Bioconductor packages, edgeR [12] and DESeq2 [13] to identify differentially expressed genes separately (see Fig 3 for the workflow). EdgeR and DESeq2 identify differentially expressed genes based on the raw read counts calculated by MULTI-COM-MAP and HTSeq, resulting in four lists of differentially expressed genes (i.e., edgeR+-MULTICOM-MAP, edgeR+HTSeq, DESeq2+MULTICOM-MAP, and DESeq2+HTSeq). In contrast Cuffdiff identifies differentially expressed genes directly from the genome mapping outputs containing only reads mapped to a unique location on the genome, resulting in one list of differentially expressed genes. Cuffdiff, edgeR and DESeq2 further adjust p-values by multiple testing using Benjamini and Hochberg's approach, which controls the false discovery rate (FDR) [10,12,13]. Usually, the cut-off of p-value (or q-value) is set to 0.05. Based on the five lists of differentially expressed genes generated by Cuffdiff, edgeR+MULTICOM-MAP, DESeq2+MULTICOM-MAP, edgeR+HTSeq, and DESeq2+HTSeq, a consensus list of differentially expressed genes is generated as the final output which usually comes from the overlap of at least three lists of differentially expressed genes. This step generates valuable information that may play an important role in the biological experiment. For example, the significantly differentially expressed genes identified by RNAMiner could be the targets for new biological experiments. This analysis step can generally reduce the size of data of the previous step by a couple orders of magnitude, condensing the data set size to several hundred kilobytes.
Predicting gene functions
We use MULTICOM-PDCN [17,18], a protein function prediction method ranked among the top methods in the 2011-2012 Critical Assessment of Function Annotation (CAFA) [29], to predict functions of differentially expressed genes (see Fig 4 for the workflow). MULTI-COM-PDCN integrates sequence-profile and profile-profile alignment methods (PSI-BLAST [30] and HHSearch [31]) with protein function databases such as the Gene Ontology database [32], the Swiss-Prot database [33], and the Pfam database [34], to predict functions of proteins in Gene Ontology [32] terms in three categories: biological process, molecular function, and cellular component. MULTICOM-PDCN also provides some statistical information about the predicted functions, such as the number of differentially expressed genes predicted in each function. We then use the Cochran-Mantel-Haenszel test implemented by R program mantelhaen.test [35,36] to check if predicted function terms are good for Fisher's exact test to identify the significantly enriched GO function terms. A p-value from the MH test lower than 0.05 suggests the two nominal variables (e.g., two function terms) are conditionally independent in each stratum [35,36]. We then calculate a p-value of enrichment for each predicted function using R function fisher.test [35,36,37,38,39,40,41,42] and sort the predicted functions by their p-value in ascending order, from the most significant ones to the least significant ones. The list of the most significantly enriched functions can provide an overview of the biological processes differentially perturbed in two biological conditions. Although the physical size of the data and knowledge generated in this step is comparable to the size of the data in the previous step, the differentially expressed genes can be organized in three functional perspectives: biological process, molecular function, and cellular component.
Constructing gene regulatory networks
We use MULTICOM-GNET [19,20] to construct gene regulatory networks based on differentially expressed genes and transcription factors in a genome (see Fig 5 for the workflow). MUL-TICOM-GNET firstly clusters differentially expressed genes with similar expression patterns into functional clusters using the K-means clustering algorithm. Secondly, it builds a binary decision tree to represent potential regulatory relationships between several selected transcription factors (TFs) and the genes in each cluster. Thirdly, it re-assigns differentially expressed genes into clusters whose gene regulatory tree best explained the expression patterns of the genes. The last two steps are repeated until the maximized likelihood of the gene expression data is reached [19,20]. We also use a R network analysis and visualization package "igraph" [43] to visualize gene regulatory networks by linking the regulatory relationships between and within all the gene regulatory modules predicted by MULTICOM-GNET together. The regulatory network construction step generates a comprehensive understanding of underlying mechanisms controlling the expression of a transcriptome and can significantly reduce the size of data. The hundreds of kilobytes of the biological network data provide a system view of the cellular systems, which can be more readily utilized to generate valuable hypotheses for biological experiments.
For replicates from RNA-Seq experiments, RNAMiner maps reads of the replicates to reference genomes and calculates gene expression values separately. The gene expression values of the replicates of two samples are combined into a profile (i.e. a vector of the expression values of a gene in each replicate of each condition), which is input into edgeR and DESeq2 to identify differentially expressed genes. Additionally, the TopHat mapping results of the replicates of two samples are input into Cuffdiff to identify differentially expressed genes. EdgeR [12], DESeq2 [13], and Cuffdiff [10] handle the replicates by modeling the variance (dispersion) in counts across the replicates as a function of the mean count of the replicates. EdgeR [12] estimates the variance by conditional maximum likelihood conditioned on the total count for the gene. DESeq2 [13] uses a flexible and mean-dependent local regression to estimate the variance between the replicates by pooling genes with similar expression levels to enhance the variance estimation. Cuffdiff [10] estimates the variance based on a negative binomial model and uses ttest to calculate the test statistics. Cuffdiff can make a model on each condition with replicates, or use a global model for all conditions together.
Before calling a tool to do data analysis, RNAMiner checks whether the data is appropriate to the tool. For example, MULTICOM-GNET [19,20] is not applied if no transcription factors exist in differentially expressed genes because MULTICOM-GNET needs at least one transcription factor to build gene regulatory networks. Another example is, for some special datasets, overexpression of some treatments in some regions of the genome in one condition leads to very large read counts of some genes in this condition, and dramatic differences of gene expressions between two conditions. This violates the assumption of edgeR's normalization method [12] that the majority of the genes should have similar expression levels. Therefore, calculating a normalization factor across all loci is difficult. RNAMiner will check this assumption and will not call edgeR if it is violated.
Evaluation and Discussion
We tested the RNAMiner protocol on six sets of RNA-Seq data generated from Human, Mouse, Arabidopsis thaliana and Drosophila melanogaster cells in order to evaluate its effectiveness. The details such as organisms, biological conditions, and experimental settings about the six sets of RNA-Seq data were reported in Table 3. The results of each of the five analysis steps are described and discussed as follows.
Results of mapping RNA-Seq reads to a reference genome
RNAMiner used TopHat2 [8] and Bowtie [44] to map RNA-Seq reads in the first and second data sets to the Mouse reference genome (mm9) in the UCSC genome browser [26] in conjunction with the RefSeq genome reference annotation (mm9) [27], map RNA-Seq reads in the third and fourth data sets to the Drosophila melanogaster reference genome (dm3) in the UCSC genome browser [26] in conjunction with the RefSeq genome reference annotation (dm3) [27], map RNA-Seq reads in the fifth data set to the Arabidopsis thaliana reference genome (ftp://ftp.arabidopsis.org/home/tair/Sequences/whole_ chromosomes/) in conjunction with the Arabidopsis thaliana genome reference annotation (ftp://ftp.arabidopsis.org/home/ tair/Genes/TAIR10_genome_release/TAIR10_gff3/), and used TopHat2 [8] and Bowtie2 [9] to map RNA-Seq reads in the sixth data set to the Homo sapiens reference genomes (hg19) in the UCSC genome browser [26] in conjunction with the RefSeq genome reference annotation (hg19) [27]. Tables 4-9 show the mapping statistics of six sets of RNA-Seq data. Overall, more than 70% of reads were mapped to the genome successfully. Particularly, a very high mapping rate (~97%) was reached on the sixth data set. These mapping success rates were within the reasonable range, suggesting the good quality of the data and the correctness of the mapping process. This reads mapping process reduced the size of data by several orders of magnitude.
Gene expression values calculated from the reads mapping data
RNAMiner removed reads that mapped to multiple locations on a reference genome from the mapping data. The gene expression values were calculated by Cufflinks [10], MULTICOM--MAP [14,15,16], and HTSeq [11] on the remaining RNA-Seq reads mapped to unique locations on the genome. Compiling reads mappings into gene expression values generates an overall profile of the expression levels of most genes in a transcriptome, which can reduce the size of dataset by about one thousand-fold (i.e., from gigabytes to megabytes) in our experiments. The compilation process transforms the raw data into meaningful expression profiles of genes. For example, three gene expression plots for comparisons between Control and each treatment in mutant mouse in the first data set are shown in Fig 6, order to make these figures readable. Usually, the points beyond the diagonal are candidates of differentially expressed genes. MULTICOM-MAP and HTSeq were used to calculate the raw read counts in the third and fourth sets of data. The counts were normalized by dividing them by the total number of uniquely mapped reads in the replicate. The normalized count of a gene was an indicator of the relative expression level of the gene in the replicate. The normalized counts of a gene in multiple replicates of a sample were further averaged and used as the measure of the relative expression level of the gene in the sample. Fig 10. The left plot in each figure was generated from all the genes, and the right one was generated from differentially expressed genes. The gene expression values were calculated by MULTICOM-MAP and normalized by log 2 . According to the two plots in Figs 9 and 10, the distribution of expression values of differentially expressed genes is quite different than that of the rest of the genes.
Differentially expressed genes identified from the RNA-Seq data
RNAMiner identified differentially expressed genes between control and each treatment using Cuffdiff [10], edgeR [12], and DESeq [13]. The threshold of p-value was set to 0.05 to select differentially expressed genes. For example, the number of differentially expressed genes for each comparison and their overlaps in both mutant mouse and wild-type mouse in the first data set are shown in Fig 11. The number of differentially expressed genes for each comparison in the second data set is shown in Fig 12. These differentially expressed genes were derived from the overlaps of three sets of differentially expressed genes separately identified by Cuffdiff, MULTI-COM-MAP+edgeR, and MULTICOM-MAP+DESeq. As shown in Fig 12, the number of differentially expressed genes increased with the increase of FruHis concentration in the absence or presence of 4 μM lycopene. The number of differentially expressed genes for two comparisons between Col (Wild-Type) and hae-3 hsl2-3 (mutant), between Col_qtrim (Wild-Type) and hae-3 hsl2-3_qtrim (mutant), and their overlaps in the fifth data set are shown in Fig 13. These differentially expressed genes were derived from the overlaps of three sets of differentially expressed genes generated separately by Cuffdiff, MULTICOM-MAP+edgeR, MULTICOM-MAP+DESeq. We also identified differentially expressed genes for two comparisons: between 2pc3 and 1Sfesrrb, between 4ctrl and 3DY131 in the sixth data set using edgeR based on read counts calculated by MULTICOM-MAP. EdgeR identified 6,210 differentially expressed genes for the comparison between 2pc3 and 1Sfesrrb, and 590 differentially expressed genes for the comparison between 4ctrl and 3DY131. On the RNAMiner web service, users can select different p-value (or qvalue) thresholds to select a specific number of differentially expressed genes according to their needs. In addition to generating the testable biological hypotheses (e.g. gene targets for experimental testing), differential gene expression analysis generally reduces the size of data by about two folds, shifting point of interest from almost all the genes in a genome to a small portion of genes most relevant to the biological experiment. in the top 10 biological processes. In these figures, the number besides each column is p-value of the enrichment of each predicted function. Although the step of gene function analysis does not substantially reduce the size of data physically, it can logically summarize hundreds of differentially expressed genes into a small number (i.e., tens) of biological processes activated or deactivated in the biological experiment which sheds light into the potential biological mechanism relevant to the experiment.
Constructed gene regulatory networks
RNAMiner used MULTICOM-GNET [19,20] to construct gene regulatory networks based on differentially expressed genes and transcription factors. For example, a repression gene regulatory module with expression correlation 0.85 in mutant mouse in the first data set is illustrated in Fig 17. This module was comprised of 21 differentially expressed genes. Three transcription factors: Tgfb1i1, Htatip2, and Jun, were predicted to collaboratively regulate this group of genes. An activation gene regulatory module for the comparison between Col (Wild-Type) and hae-3 hsl2-3 (mutant) with expression correlation 0.85 in the fifth data set is shown in Fig 18. This module was comprised of 35 differentially expressed genes. Four transcription factors, AT3G59580, AT1G56650, AT1G28050, and AT1G52890, were predicted to collaboratively regulate this group of genes.
RNAMiner also used a R package "igraph" [43] to visualize gene regulatory networks by linking the regulatory relationships between and within all the gene regulatory modules The step of gene regulatory network reconstruction condenses hundreds of differentially expressed genes and their expression data into dozens of valuable gene regulatory modules, which may reveal the underlying biological mechanism controlling the expression in the biological experiment. The network modules not only provide the human comprehensible interpretation of the gene expression levels, but also the important transcription factors and their target genes that are very valuable for generating hypotheses for new biological experiments.
Use of the RNAMiner Web Service
The RNAMiner web service (Fig 20) is available at http://calla.rnet.missouri.edu/rnaminer/ index.html. Users can submit requests on the home page and receive an email with a link to the data analysis results. 8. Upload reads files. The last three categories request users to upload reads files for both two samples. Users can upload more than one reads files for each sample.
After a request is submitted successfully, one web page (Fig 21) will be shown saying the data is in process. If one user submitted one request to the RNAMiner web service and it is running or it is in the waiting queue, he/she cannot submit another request.
Receive the results
When the data analysis is finished, users will receive an email with a link to one web page (Fig 22) with the data analysis information and a result link. The result page (Fig 23) will be shown by clicking the result link. Users can view and download the analysis data on the result page. The time expense of analyzing a set of RNA-Seq data by RNAMiner depends on how big the data is, how many reads files there are in the data set, and how many jobs there are in the waiting queue. Normally a data analysis can be finished by RNAMiner in several hours. However, the time expense will be longer if there are a lot of jobs in the waiting queue. Our server cannot handle too many jobs at the same time because of CPU and space limitations.
Conclusions
The RNAMiner protocol and pipeline can progressively reduce the size of large datasets to produce valuable and comprehensible biological knowledge of manageable size, ranging from gene expression values, differentially expressed genes, gene function predictions, and gene regulatory networks. The test results on six RNA-Seq datasets of four different species help demonstrate its utility and versatility.
In order to further improve the quality of RNA-Seq data analysis, additional tools can be plugged into the RNAMiner protocol. In the future, we will add a high-speed RNA mapping tool-Gsnap [25] and a high-accuracy RNA mapping tool-Stampy [45] into the pipeline to map RNA reads to reference genomes. For identifying differentially expressed genes, we will include baySeq [46], ShrinkSeq [47], and NOISeq [48] into the pipeline in order to handle various sources of noise in RNA-Seq data even better. Furthermore, we will include an in-house tool of constructing biological networks from a group of co-expressed genes to reconstruct highly valuable metabolic networks and signal transduction networks for gene clusters identified by the RNAMiner protocol. Moreover, we will add the capability of analyzing the function of non-coding small RNAs into RNAMiner and use the information during the reconstruction of biological networks. The new improvements will be incorporated into the RNAMiner web service for the community to use. | 5,839.4 | 2015-04-22T00:00:00.000 | [
"Biology",
"Computer Science"
] |
Theoretical Determination and Experimental Verification of Transport Index of Rectangular Shaped Radioactive Containers
Ionizing radiations are widely used to sustain and enhance our quality of life in the areas such as medical diagnosis, therapy, scientific research and industry etc. Ionizing radiations are available from radioactive sources which are made of radioactive materials. The radioactive materials are produced in either nuclear power or research reactors or nuclear accelerators or extracted from the naturally found radioactive ores. These radioactive sources and radioactive materials need to be transported from their places of production to the places of applications and finally to waste repositories. The radioactive materials are transported in well designed packages having various shapes and sizes. In the field of radioactive transport, it is a mandatory to find the Transport Index (TI) to be mentioned on each package for transportation. This research is focused on the determination of the maximum γ-ray radiation dose at one meter from the surface of cubic and rectangular shaped package or containers. A computer code “Solid Angle for Transport Index” (SAFTI) has been developed using MATLAB to determine the location of maximum value of the radiation dose rate from the surface of a rectangular or square container. This maximum dose rate is used to determine the transport index. Some of the results of the code have been compared with the experimental results. The results of this research are useful not only to determine TI for individual packages but also to find the TI of the vehicles carrying the transport packages.
Transport Index
Introduction
The worldwide use of radioactive materials and radiation sources in the fields of industry, energy production, medicine, agricultural sciences, scientific and technological research are very rewarding for sustaining and enhancing our quality of life [1]- [3].While these radiations are useful their detrimental effects cannot be ignored.All radioactive materials emit radiations which are very harmful for the surrounded people if exposed.So, a great care is required to handle radioactive materials.Transport of radioactive sources and materials is often required but transportation of these radioactive materials and sources through public domain may cause hazards to the general public.Some 800,000 shipments of hazardous material are carried out on daily basis in USA that also includes radioactive materials [4]- [7].It is the responsibility of the agencies transporting the radioactive materials to guarantee the safety of the general public.
Safety is defined, in one way, as taking all measures to minimize the occurrence of the accidents and if accident does take place then taking all measures to minimize the consequence of the accidents.To minimize the consequences of the accident the Transport Index (TI), just a number, helps the rescue team while taking actions after the accidents [8].Transport Index (TI) is a number that is written on each package or container of radioactive material.The transport index is related to the maximum dose equivalent rate measured in mSv/h at one meter from the surface of a package containing radioactive material.The radiation dose determined in this way is multiplied by 100 and after rounding a whole number (without any unit) is used as Transport Index.Radioactive transportation packages are divided into different categories for shipment according to the "TI" that range generally from 0 to 10.
The maximum dose equivalent rate can be located only by a radiation detector scanning all surface of the package of container.In this research, we have developed a computer code to locate the position on the surface of a package or container where maximum dose may occur.In this research two types of popular geometry have been considered, i.e. square and rectangular.Some of the points of our theoretical model have been experimentally verified.
Theoretical Model
In this research, we have employed the basic concept of radiation dose from a point isotropic γ-ray source to a rectangular surface detector.
The radiation dose rate is given by d where, ϕ = Photons flux (#/cm 2 -sec) = Linear mass absorption co-efficient (cm 2 /g) dΩ = Solid angle that the point source subtend with the rectangular area 0.57 = the conversion factor It was further assumed for simplicity that ϕ = One photon/cm 2 sec where k is taken as 0.57.One can find the dose rate in real cases, just simply, by multiplying Equation ( 2) with the known flux (#/cm 2 -sec), γ-ray energy, E (MeV) and linear mass absorption coefficient µ ρ (cm 2 /g) of the dose receiving material.Equation (2) shows that the maxima in the dose rate can occur only when the solid angle subtended by the rectangular surface of the package with point isotropic source has the maxima.The solid angle subtended by any area at a distance r from a point isotropic source can be given by [9] 2 ˆd d n r where da is a small area of any shape, and n is the unit normal vector to the surface and r is distance from the origin to observing point on the differential surface area.Now, consider an isotropic rectangular source and a point detector at height "h" as shown in Figure 1.Thus the solid angle dΩ subtended by the rectangle OABC at the point detector ( ) 0, 0, P h can be determined as [10] ( ) The solid angle subtended by the shaded area as shown in Figure 2 can be determined as Equation ( 6) covers the general situation for all square and rectangular packages or containers where the detector positions can vary as shown in Figure 3.One can move the detector position on the surface of the detector and determine the value of the solid angle.A computer code "Solid Angle For Transport Index" (SAFTI) has been developed using MATLAB to determine the location of maximum value of the radiation dose rate from the surface of a rectangular or square container.The SAFTI code is appended at the end of this paper to be used by the readers for required rectangular dimensions.
Experimental Verification
A sheet of paper of size 100 × 100 cm 2 was placed on the table with a grid of 10 × 10 cm 2 .The experimental setup is shown in Figure 4.A 1.08 mCi, Co-60 gamma-ray source was placed at heights 50 cm, 75 cm and 100 cm along the z-axis in one set of experiment whereas the survey meter (PRM 300, Sr. No, 00528, 00529) were placed in different squares of dimension 10 cm × 10 cm.The survey meters were exposed for 72 hours.
Results and Discussion
We assumed a surface area of 1 m × 1 m.Making a grid of 10 × 10 cm 2 , solid angles were computed for different grids located on the plane.For this purpose a self developed code Solid Angle For Transport Index (SAFTI) was written in MATLAB.The computed dose rates are shown in the Tables 1-3 in the third column along with the computed solid angle in the forth column with respect to the positions of the source along z-axis.The experimental and theoretical results are in concordance.The small difference in experimental and computed results might be due to approximation of a point source as can be more pronounced at close distance such as the grid (2, 2) at 50 cm.However, the result shows that in air the maximum of the dose occurs if solid angle at the point detector has the maxima.For a square region 100 cm × 100 cm and 100 cm normal to the surface the maximum solid angle has been computed to be 0.805 steradian at the center of the square.These results are valid for all square shaped containers irrespective of the dimensions of the square.The minimum solid angle has been determined theoretically as 0.524 steradian on all corners of the square.Figure 5 shows the graphical representation of solid angles and thus the locations from where the transport Index may be determined.
Table 4 shows the computed values of solid angles with different surface size of the containers.The results show that the maxima of dose rate occur at the point where the solid angle is maximum provided uniform radiation are emitted from the surface of the container.As the size of the radioactive container increases the solid angle approaches the maximum value of 2π steradian.In this research, both the experiment and theory were conducted for a point source and a planer detector surface whereas determination of TI requires a planer source and a point detector.As a matter of fact, the two cases are related with a constant and the positions of the maximum radiation dose depend up on the maxima of the solid angle.The position of the maximum radiation dose lies of the intersection of the diagonal of the square or rectangular surface of the radioactive source package.
Conclusion
The result of this research shows that the maxima of the dose at one meter above the surface can be located with maxima of the solid angle.Since there is no instrument that can find the solid angle with a given plane, therefore, computational values can be employed to find the location of maximum dose rate and thus maximum value of the transport index can be found otherwise underestimated TI value will be quoted.The results show that the locations for the maximum dose rate are independent of the size of square.The location for maximum value of the solid angle remains the same for rectangular shaped surfaces as long as the ratio of the length to width is constant.
Figure 1 .
Figure 1.The solid angle subtended by the rectangle surface OABC at point P.
Figure 3 .
Figure 3. Solid angle subtended by the shaded rectangle ABCD at the general point P.
Figure 4 .
Figure 4. Experimental set up to measure the dose rate received by the detectors from a point source at different heights along z-axis.
Figure 5 .
Figure 5. Variation of solid angles with detector positions normal to the surface on a square radioactive container (120 cm × 120 cm).
Figure3depicts even a very general case where point detector is placed at a point ( )
Table 1 .
It shows the experimental and theoretical values of dose rates and computed values of solid angles at different height (h) of the point detector along z-axis.
Table 2 .
It shows the experimental and theoretical values of dose rates and computed values of solid angles at different height (h) of the point detector along z-axis.
Table 3 .
It shows the experimental and theoretical values of dose rates and computed values of solid angles at different height (h) of the point detector along z-axis.
Table 4 .
It shows the minimum and maximum value of the solid angle with reference to the rectangular surface area of the radioactive source package. | 2,482.2 | 2014-04-14T00:00:00.000 | [
"Physics"
] |
SCALE method for single-cell ATAC-seq analysis via latent feature extraction
Single-cell ATAC-seq (scATAC-seq) profiles the chromatin accessibility landscape at single cell level, thus revealing cell-to-cell variability in gene regulation. However, the high dimensionality and sparsity of scATAC-seq data often complicate the analysis. Here, we introduce a method for analyzing scATAC-seq data, called Single-Cell ATAC-seq analysis via Latent feature Extraction (SCALE). SCALE combines a deep generative framework and a probabilistic Gaussian Mixture Model to learn latent features that accurately characterize scATAC-seq data. We validate SCALE on datasets generated on different platforms with different protocols, and having different overall data qualities. SCALE substantially outperforms the other tools in all aspects of scATAC-seq data analysis, including visualization, clustering, and denoising and imputation. Importantly, SCALE also generates interpretable features that directly link to cell populations, and can potentially reveal batch effects in scATAC-seq experiments.
A ccessible regions within chromatin often contain important genomic elements for transcription factor binding and gene regulation 1 . Assay for Transposase-Accessible Chromatin using sequencing (ATAC-seq) is an efficient method to probe genome-wide open chromatin sites, using the Tn5 transposase to tag them with sequencing adapters 2 . In particular, single-cell ATAC-seq (scATAC-seq) reveals chromatin-accessibility variations at the single-cell level, and can be used to uncover the mechanisms regulating cell-to-cell heterogeneity 3,4 . However, in an scATAC-seq experiment, each open chromatin site of a diploidgenome single cell only has one or two opportunities to be captured. Normally, only a few thousand distinct reads (versus many thousands of possible open positions) are obtained per cell, thus resulting in many bona fide open chromatin sites of the cell that lack sequencing data signals (i.e., peaks). The analysis of scATACseq data hence suffers from the curse of "missingness" in addition to high dimensionality 3 .
Many computational approaches have been designed to tackle high-dimensional and sparse genomic sequencing data, especially single-cell RNA-seq (scRNA-Seq) data. Dimensionality reduction techniques such as principal component analysis (PCA) 5 and t-distributed stochastic neighbor embedding (t-SNE) 6 are frequently employed to map raw data into a lower dimensional space, which is particularly useful for visual inspecting the distribution of input data. Clustering based on the full expression spectrum or extracted features can be performed to identify specific cell types and states, as well as gene sets that share common biological functions [7][8][9][10] . The imputation of missing expression values is also often carried out to mitigate the loss of information caused by dropouts in scRNA-seq data 11,12 .
Direct applications of these scRNA-seq analysis methods to scATAC-seq data, however, may not be suitable due to the closeto-binary nature and increased sparsity of the data (Supplementary Fig. 1). A recent method specifically developed for scATACseq data analysis, chromVAR 13 , evaluates groups of peaks that share the same motifs or functional annotations together. Another method, scABC, weighs cells by sequencing depth and applies weighted K-medoid clustering to reduce the impact of missing values 14 . To refine the clustering, it then calculates a landmark for each cluster and assigns cells to the closest landmarks based on the Spearman correlation. However, each method suffers notable caveats: chromVAR only analyzes peaks in groups and lacks the resolution of individual peaks, whereas scABC heavily depends on landmark samples with high sequencing depths, and the Spearman rank can be ill-defined for data with many missing values (in particular for scATAC-seq data). Recently a newly developed method called cisTopic applied latent Dirichlet allocation to model on scATAC-seq data to identify cisregulatory topics and simultaneously cluster cells and accessible regions based on the cell-topic and region-topic distributions 15 .
Expressive deep generative models have emerged as a powerful framework to model the distribution of high-dimensional data. One of the most popular of such methods, the variational autoencoder (VAE), estimates the data distribution and learns a latent distribution from the observed data through a recognition model (encoder) and a generative model (decoder) 16 . It does this by maximizing the similarity of the calibrated data (output by the decoder) with the input data and minimizing the Kullback-Leibler divergence of the approximate from the true posteriors 16 . VAE could be applied to data embedding and clustering based on the low-dimensional latent representation of the input highdimensional data 17 . Recently, a method called scVI adapted VAE for scRNA-seq data analysis 18 . However, the standard VAE employed by scVI uses a single isotropic multivariable Gaussian distribution over the latent variables and often underfits sparse data 19 . A tighter estimation of the posterior distribution could greatly improve the power of VAE in fitting and analyzing sparse data 19 . Applying Gaussian Mixture Model (GMM) as the prior over the latent variables has been used in unsupervised clustering and to generate highly realistic samples by learning more disentangled and interpretable latent representations [20][21][22] .
Here, we introduce SCALE (Single-Cell ATAC-seq analysis via Latent feature Extraction), a method that combines the VAE framework with the Gaussian Mixture Model (GMM, a probabilistic model to estimate observed data with a mixture of Gaussian distributions). We validated the effectiveness of SCALE in extracting latent features that characterize the distributions of input scATAC-seq data on multiple different datasets generated on different platforms with different protocols, and of different overall data qualities. We then used the latent features to cluster cell mixtures into subpopulations, and to denoise and impute missing values in scATAC-seq data. We compared the performance of SCALE with other widely-used dimensionality reduction techniques, as well as with the state-of-art scRNA-seq and scATAC-seq analysis tools. We found that SCALE substantially outperforms the other tools in all aspects of scATAC-seq data analysis. It is even comparable to sophisticated experimental technologies with additional steps (e.g., Pi-ATAC 23 , which uses protein labeling as an aid in defining cell identifies) in correctly revealing cell types and their specific regulatory motifs in a tumor sample.
Results
The SCALE model and validation datasets. SCALE combines the variational autoencoder (VAE) and the Gaussian Mixture Model (GMM) to model the distribution of high-dimensional sparse scATAC-seq data (Fig. 1). SCALE models the input scATAC-seq data x as a joint distribution p x; z; c ð Þwhere c is one of predefined K clusters corresponding to a component of GMM, z is the latent variable obtained by z ¼ μ z þ σ z à ϵ, where μ z and σ z are learned by the encoder network from x, and ε is sampled from Nð0; IÞ 16 . Since z is conditioned on c, p x; z; c ð Þ can be written as p xjz ð Þp zjc ð ÞpðcÞ where p(c) is a discrete distribution of K predefined clusters, p(z|c) follows a mixture of Gaussians distribution with a mean μ c and a variance σ c for each component corresponding to a cluster c, and p(x|z) is a multivariable Bernoulli distribution modeled by the decoder network ( Fig. 1). In the SCALE framework, each cell x i is first transformed into a d-dimensional vector of latent features z i on the GMM manifold by an encoder network and then reconstructed back through a decoder network with the original dimensionality to represent the chromatin openness at each peak in each cell. The latent features that capture the characteristics of scATAC-seq data are then visualized in the low-dimensional space with t-SNE, and used to cluster single cells with various clustering methods, e.g., K-means.
We constrained the hyper-parameters of SCALE on the Leukemia scATAC-seq dataset and found SCALE is insensitive to the encoder structure and the dimension of latent features (Supplementary Table 1). The SCALE model with default parameters can be accessed in the Online Method. We then tested the model by using the GM12878/HEK293T, the GM12878/HL-60, and the InSilico datasets 3,4,24 , and two other recently published Splenocyte and Forebrain datasets 25,26 . The Leukemia dataset is derived from a mixture of monocytes (Mono) and lymphoid-primed multipotent progenitors (LMPP) isolated from a healthy human donor, and leukemia stem cells (SU070_LSC, SU353_LSC) and blast cells (SU070_Leuk, SU353_Blast) isolated from two patients with acute myeloid leukemia 24 . The GM12878/HEK293T dataset and the GM12878/ HL-60 dataset are respective mixtures of two commonly-used cell lines 22 . The InSilico dataset is an in silico mixture constructed by ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-12630-7 computationally combining six individual scATAC-seq experiments that were separately performed on a different cell line 3,11 . Note that these four datasets were the same ones used to validate scABC 14 . The more recent Splenocyte dataset 25 is derived from a mixture of mouse splenocytes (after red blood cell removal) and the Forebrain dataset 26 is derived from P56 mouse forebrain cells. The six datasets cover scATAC-seq data generated from both microfluidics-based and cellular indexing platforms, and the distributions of the number of peaks in each single cell vary substantially in different datasets ( Supplementary Fig. 1). However, they always have a high level of data sparsity compared to the aggregation of peaks from all single cells in each dataset (Supplementary Table. 2). SCALE identifies cell types by clustering on latent features. We examined SCALE's ability to uncover features that characterize scATAC-seq data distributions. By default, SCALE extracts 10 features from the input data. For comparison, we also applied PCA, scVI and cisTopic to reduce the input data to 10 dimensions. In the comparison, we also included Cicero 27 , a scATACseq data analysis tool for predicting cis-regulatory interactions and building single-cell trajectories from scATAC-seq data, and TF-IDF a transformation for performing dimension reduction and clustering 28 . We then visualized the extracted features from these tools as well as the raw data with t-SNE. In general, the feature embeddings of SCALE and cisTopic were better separated between cell types, whereas the embeddings of PCA, scVI, Cicero, TF-IDF and the raw data overlapped between some cell types ( Fig. 2a, Supplementary Fig. 2). SCALE can also reveal the distance between different cell subpopulations and sometimes suggested their developmental trajectory in UMAP visualization 29 (Supplementary Fig. 3). For example, in the Forebrain dataset the three clusters of excitatory neuron cells (EX1, EX2, and EX3) are close to each other in the latent space. For the Splenocyte dataset, the different T-cell subpopulations are in the neighborhood, the B cells form another bigger group, and the two types of natural killer cells also cluster closely. For the Leukemia dataset, Mono and LMPP cells are the most dissimilar in leukemia evolution and they were indeed the farthest separated. LSCs exhibit strong similarity to LMPPs 30 , consistent with that the LSC cells (LSC_SU070, LSC_SU353) were close to the LMPP cells in the embedding. Finally, the blast cells (Blast_SU070 and Blast_SU353) showed a bimodal distribution, with some more differentiated blasts closer to monocytes 31,32 .
We then applied K-means clustering on the SCALE-extracted latent features and assessed the clustering accuracy by comparing the results with scABC, scVI, cisTopic, and SC3 33 , another widely-used clustering method for scRNA-seq data. SCALE displayed the overall best performance on all five real experimental mixture datasets, and was nearly as accurate as scABC and cisTopic on the InSilico dataset (Fig. 2b, Supplementary Fig. 4). The newly developed cisTopic generally performed pretty well on all datasets, with the overall clustering performance only slightly lower than SCALE, but it misclassified a few clusters on the Splenocyte dataset. We also compared with TF-IDF and Cicero on clustering. TF-IDF performed well on most datasets (although not as good as SCALE) except on the Forebrain dataset. However, Cicero did not perform well on most datasets; indeed, data visualization and clustering are not major goals of Cicero. On the Forebrain dataset, cluster assignments of SCALE were the closest to the reference cell types. Due to the sparsity of data, the Pearson and Spearman correlations were both ill-defined ( Supplementary Fig. 5a), which directly led to poor clustering for SC3 where most cells collapsed into one group. Although the VAE-based method scVI did not suffer from the problem of ill-defined cell distance, it misclassified three subgroups of cells (s1, s2, s3 labeled on the confusion matrix. Supplementary Fig. 5b).
To identify the cause of the misclustering by scVI, we searched for the most similar cell types for the three subgroups (s1, s2, s3). We aggregated the peak profiles of each cell type or subgroup to form a representing meta-cell and calculated the similarities among the meta-cells. As expected, s1 is the most similar to EX2, s2 the most similar to EX3, and s3 to AC (astrocyte) in the original data ( Supplementary Fig. 5c). Both scVI and SCALE model the distribution of scATAC peak profiles to remove noise and to impute missing values (discussed in detail in the next section). We found that, consistent with the clustering results, this data calibration by scVI actually made s1, s2, and s3 cells less similar to the original cell types of EX2, EX3, and AC, respectively. On the contrary, SCALE retained the similarities of the three subgroups to their original cell types. Strikingly, when removing the GMM restriction from the overall framework but keeping the other part of the network the same, the degenerated SCALE yield performance was similar to that of a regular VAE, like scVI ( Supplementary Fig 5d). Thus, introducing GMM as the prior to restrict the data structure provides SCALE with greater power for fitting sparse data than regular VAE using single Gaussian as the prior. Finally, we tested whether SCALE is robust with respect to data sparsity by randomly dropping scATAC-seq values in the raw datasets down to zero. We compared the clustering accuracy of SCALE and other tools at different dropping rates (10-90%), measured by the adjusted Rand Index (ARI), Normalized Mutual Information (NMI) and micro F1 score (Methods). We found that SCALE displayed only a moderate decrease in clustering accuracy with increased data corruption until at about the corruption level of 0.6, and was robust for all datasets ( Supplementary Fig. 6). In general, scABC, SC3, and scVI also showed robustness to data corruption; however, the overall clustering accuracies were much lower on some datasets (e.g., SC3 failed on the Forebrain dataset and scVI failed on the GM12878/ HEK293T and the GM12878/HL-60 datasets). On the Forebrain dataset, the ARI scores of SCALE dropped from 0.668 using the raw data to 0.448 on using the data with 30% corruption, and scABC and scVI dropped from 0.315 to 0.222 and from 0.448 to 0.388, respectively.
Finally we also provide a method to help users choose the optimal number of clusters based on the Tracy-Widom distribution 34 (Methods), which could often produce an estimate of the number of clusters close to that of the references ( Supplementary Fig. 7) and generate clustering results similar to the reference sets ( Supplementary Fig. 7).
SCALE reduces noise and recovers missing peaks. An important feature of SCALE is the ability to accurately estimate the real distribution of scATAC-seq data, which usually contains both noise and a large number of missing values. The estimate could be used to remove noise and restore missing data (Fig. 1). We evaluated the calibration efficiency of SCALE on both real and simulated datasets. Since no such tool is currently available for scATAC-seq data, we compared SCALE with scImpute, SAVER, MAGIC, and scVI, four state-of-the-art scRNA-seq imputation methods (Fig. 3a).
We first evaluated the ability of SCALE to remove noise and to recover missing values on real scATAC-seq datasets. A challenge of analyzing real data is that the ground truth data without any corruption is unknown. However, if we average all single cells of the same biological cell type, the resulted meta-cell will be a good approximate to those single cells. SCALE performed better than all scRNA-seq imputation methods in all scATAC-seq datasets, in that it achieved the highest correlation of the single cells with the corresponding meta-cell for each cell type (Fig. 3a, Supplementary Fig. 8), indicating that it obtained a better estimate of the real scATAC-seq data distribution. For most cases, scImpute was very stable and among the best comparing with other scRNA-seq imputation methods, and SAVER performed well on denser datasets (InSilico, Splenocyte) but deteriorated on sparser datasets. MAGIC and scVI might have underfit the sparse input data and the imputed data substantially deviated from it ( Supplementary Fig. 9), which may reflect that the two powerful tools that are optimized to scRNA-seq data imputation may not fit for scATAC-seq data analysis.
It is important to note that the data calibration of SCALE was obtained at the same time of data modeling and clustering, i.e., without knowing the original type of each cell. So it could not simply average all single cells of the same cell type to reconstruct the peak so that they resemble the reference meta-cell. Also importantly, SCALE achieved a high correlation with the metacells while maintaining a similar level of variation within each cell population (see the variation of correlation coefficients in Fig. 3a and Supplementary Fig. 8). Indeed, SCALE retained the original data structure (intra-correlation within the imputed data) and recovered the original peak profiles (inter-correlation with the raw data) in the process of data regularization by GMM ( Supplementary Fig. 9).
The imputation of SCALE could strengthen the distinct patterns of cluster-specific peaks by filling missing values and removing potential noise ( Supplementary Fig. 10), which improves downstream analysis, for example the identification of cell-type-specific motifs and transcription factors by chromVAR. We demonstrated this feature with the Forebrain dataset. We first followed the method used by Cusanovich et. al. to identify ARI: For comparison, SCALE, PCA, and scVI all performed dimension reduction to ten dimensions before applying t-SNE while the raw data were directly visualized with t-SNE. b Clustering accuracy was evaluated by confusion matrices between cluster assignments predicted by scABC, SC3, scVI, cisTopic and SCALE, and reference cell types. For scABC and SC3, the cluster assignments were directly obtained from the output of the tools; for SCALE and scVI, we applied the K-means clustering on the extracted features to get cluster assignments. The Adjusted Rand Index (ARI), the Normalized Mutual Information (NMI), and the F1 scores are shown on the top differentially accessible sites with the "binomialff" test of Monocle 2 package 28 . At 1% FDR threshold, we identified 4100 differential accessible sites across the eight reference clusters of the Forebrain dataset. We then used chromVAR to search for motifs enriched in the differential sites in the raw and the imputed data, respectively. Overall, the patterns of different cell types are more distinct for these differentially accessible sites in the imputed than in the raw data ( Supplementary Fig. 11a). And embedding on the imputed data shows better-defined clusters (each well corresponds to a subtype with biological definition) than on the raw data (Fig. 3b, Supplementary Fig. 11b). We found that the imputed data can greatly improve the results of chromVAR analysis by identifying more motifs (increased from 52 motifs to 105). For example, chromVAR analysis on the imputed data, but not on the raw data, identified the motifs Mafb and Hoxd9 enriched in the MG (macroglia) cluster (Supplementary Fig. 11c-d). It was recently reported that Mafb contributes to the activation of microglia 35 . It also identified Hoxd9 enriched in IN (inhibitory neuron) from the imputed but not the raw data. Similarly, we found that Dlx2, Lhx8, Arx, and Neurog1 are much more significantly enriched in the, respectively, clusters in the imputed data ( Supplementary Fig. 11c-d). Dlx2, Lhx8, and Arx are important components in the MGE (medial ganglionic eminence) pathway of forebrain development 36 , and Neurog1 is required for excitatory neurons in the cerebral cortex 37 .
We then introduced further corruption to the real data by randomly dropping out peaks at different rates (Methods). At all corruption rates, SCALE performed the best, in that the calibrated data most closely correlated with the original meta-cells ( Supplementary Fig. 12). We observed similar trends for the other scRNA-seq imputation tools as above, confirming the effectiveness of SCALE in enhancing scATAC-seq data. We further tested the impact of missingness on generative model of imputation by calculating the confusion score (Methods) to evaluate the ability to preserve the original structure (inter and intra-correlation of meta-cells) (Supplementary Fig. 13). We found that the effect was minimal when the corruption level was lower than about 0.5, and after that threshold, the generative model was less capable of preserving the original structure ( Supplementary Fig. 13b).
We subsequently tested the calibration accuracy on a simulated dataset. We constructed the dataset by first generating reference scATAC-seq data consisting of three clusters, each containing 100 peaks with no missing values, then randomly dropping out peaks and introducing noise (Methods, Supplementary Fig. 14a). As we knew the ground truth data of each single cell, we could quantify the efficiency of all tools by calculating peak-wise and cell-wise correlations of each calibrated single cell with its original ground truth. At all corruption rates, SCALE recovered the original data most accurately (Supplementary Fig. 14b-c). On the other hand, although scImpute could also recover the missing values in most cases, it messed up two clusters at the 0.2 corruption rate and was unable to remove the noise. SAVER and scVI smoothed both the signal and noise simultaneously and only recovered missing values to some degree. MAGIC performed very well at low corruption rates, but apparently over-smoothed the data and removed true signals along with noise at high levels of data corruption.
SCALE reveals cell types and their specific motifs. Next, we used SCALE to analyze a dataset generated by a recently developed technology, protein-indexed single-cell assay of transposaseaccessible chromatin-seq (Pi-ATAC), which uses protein labeling to help define cell identities 23 . Dissecting complex cell mixtures of in vivo biological samples may be challenging. By simultaneously characterizing protein markers and epigenetic landscapes in the same individual cells, Pi-ATAC provides an effective approach to tackle the problem. The Breast Tumor dataset is derived from a mouse breast tumor sample, including two plates of tumor cells (Epcam+) and another two plates of tumor-infiltrating immune cells (CD45+), isolated by protein labeling and FACS sorting. In the original study, a set of motifs was used to project the Epcam+ and CD45+ -specific chromatin features with t-SNE, and it was difficult to separate these two cell types computationally (Supplementary Fig. 15a). However, we found that SCALE was able to separate the two cell types well, better than PCA and scVI in latent embedding (Fig. 4a). On clustering, SCALE also yielded results the closest to the proteinindex labels, better than scVI and scABC, whereas SC3 poorly distinguished the two cell types (Fig. 4b). Although cisTopic grouped the cells well in the embedding, it misclassified parts of CD45+ cells into Epcam+ cells. SCALE thus can reveal cell types within complex tissues based only on scATAC-seq data, with performance comparable to sophisticated experimental technologies like Pi-ATAC. We validated the biological significance of the cell clusters based on Pi-ATAC peaks. For each cluster, we calculated the top 1000 peaks with the highest specificity score as type-specific peaks (Methods, Supplementary Fig. 15b). We then used Homer 38 to identify transcription factor binding motifs that were enriched in the type-specific peaks. We removed the common motifs enriched in both CD45+ cells and Epcam+ cells, and kept those that were enriched in only one cell type. We found that CD45+ cells were enriched for immune-specific motifs Maz, Pu.1-Irf, Irf8, Runx1, Elk4, Nfy, Elf3, and SpiB binding motifs. These findings are consistent with the role of Runx1 in maintenance of haematopoietic stem cells (HSC) and that knockout of Runx1 results in defective T-and B-lymphocyte development 39 . Nfy promotes the expression of the crucial immune responsive gene Major Histocompatibility Complex (MHC) 40 . Epcam+ cells were enriched for tumor-related motifs Klf14, Mitf, Ets1, Nrf2, and Nrf1 binding motifs. Ets1 is frequently overexpressed in breast cancer and associated with invasiveness 41 , whereas Nrf2 is a key signature for breast cancer cell proliferation and metastasis 42 (Fig. 4c). Thus, SCALE analysis of the Breast Tumor data revealed biologically relevant cis-elements for gene regulation.
SCALE disentangles biological cell types and batch effects.
In addition to tighter estimates of the multimodal input data, by pushing each dimension to learn a separate Gaussian distribution, GMM has another advantage in that it leads to latent representations that are more structured and disentangled, and thus more interpretable 21 . In SCALE, as each feature is directly connected with output peaks, it can be assessed by the most weighted connections (Methods, Supplementary Fig. 16a). For example, in the Leukemia dataset, dimensions 9 of the extracted features captured peaks specific to the Mono cell type and enriched regulatory elements related to immune-related "biological process" (BP, Methods) ( Supplementary Fig. 16b). In the Forebrain dataset, feature 3 characterized the AC (astrocyte) and the OC (oligodendrocyte) cell types, enriched elements related to "glial cell differentiation" (Supplementary Fig. 16c). In the Splenocyte dataset, features 4 and 7 portrayed two complementary sets of cell types ( Supplementary Fig. 16d), with feature 4 enriched with B cell-related processes like "regulation of cell morphogenesis" and "myeloid leukocyte activation and differentiation", and feature 7 enriched with T cell-related processes such as "immune response" and "regulation of cell killing" (Supplementary Fig. 16d). These data suggest that the features learned by the model of SCALE are disentangled and can shed light on the biological significance.
Most interestingly, we found that SCALE could possibly reveal features corresponding to potential batch effects in the input data. For example, the Breast Tumor dataset is derived from experiments performed separately on two plates of Epcam+ tumor cell samples and two plates of CD45+ tumor-infiltrating immune cells. Although SCALE successfully clustered the two cell types, the data structure in the low-dimensional space also revealed bias towards different plates (Fig. 5a). We carefully analyzed the SCALE-extracted features (Fig. 5b) and noticed that while some, e.g., features 1 and 6, were well-correlated to biological cell types, the others, e.g., features 2, 4, 8, and 10, more or less corresponded to independent plates, or, e.g., features 3 and 5, displayed biased distribution not related to cell types. Using the plate-related features (i.e., features 2, 3, 4, 5, 8, and 10) for data embedding, we found that the cells were separated by plates, but not by types. On the other hand, if we used the other plateindependent features (i.e., features 1, 6, 7, and 9), we found that indeed the cells of different plates of the same types more evenly distributed in the cluster (Fig. 5a). We further checked the represented peaks of these features and its biological significance ( Supplementary Fig. 16e). Most of plate-related features have no biological relevance, except for peaks of feature 8, which appeared in one plate of CD45+ cells and are enriched with biological processes such as "response to cytokine stimulus". This finding, however, suggests another possibility in interpreting the "plate bias" as a real biological difference in the two separate plates of CD45+ cells that might arise from sorting and cell culture.
We noticed that GM12878 cells in the InSilico dataset contain four replicates with many peak values much greater than 2. PCA analysis showed that replicates 1 and 3 were separated in the lowdimensional space ( Supplementary Fig. 17a), suggesting a possible batch effect. However, the differences in the two replicates disappeared after we binarized the data, by masking values greater than 1 to 1 (Supplementary Fig. 17b). On the other hand, we observed no particular features corresponding to any batch among the SCALE-extracted features ( Supplementary Fig. 17c). Consequently, in the embedding and clustering results based on the SCALE-extracted features, the cells of each replicate were distributed evenly in the low-dimensional space (Supplementary Fig. 17c). We confirmed this result by checking the top 200 specific peaks for each replicate based on raw data and found no significantly different pattern across replicates ( Supplementary Fig. 17d). The distinction may reflect the different characters of the two approaches: while PCA is a linear method and sensitive to quantitative variations, SCALE is non-linear and more stable. Lastly, we repeated the analysis on the Splenocyte and the Forebrain datasets-the other two datasets that contain different experimental batches, and found no batch-related features, and the cells of different batches were distributed indistinguishably in the low-dimensional space ( Supplementary Figs. 18, 19).
SCALE is scalable to large datasets. We further examined a mouse single-cell atlas of profiled chromatin accessibility iñ 80,000 single cells from 13 adult mouse tissues by sci-ATAC-seq 28 to investigate whether SCALE works for large datasets. The atlas study used a computational pipeline to infer 30 cell types from the dataset by graphic clustering, which were used as "reference" cell types when benchmarking SCALE. SCALE worked well on this big dataset and showed a good agreement with the reference: the overall F1 score was 0.419, and most of the major reference clusters have a corresponding one identified by SCALE. Nevertheless, some large reference clusters were split into two or three small groups (Supplementary Fig. 20). Finally, we benchmarked the running time and memory usage of SCALE on different scales of datasets by downsampling a subset of cells and peaks from the mouse atlas datasets (10,000 peaks and different cell number). We found that SCALE required a little over 1. importantly the used computational resource only increased slightly with the size of datasets ( Supplementary Fig. 21).
Discussion
Our work shows that SCALE accurately characterizes the distribution of high-dimensional and sparse scATAC-seq data by using a deep generative framework to extract latent features. SCALE is thus a powerful tool for scATAC-seq data analysis, including data visualization, clustering, and denoising and imputation. In all comparisons, SCALE performs much more favorably than scABC and scRNA-seq tools. Based on the better clustering assignments and imputation data, we can improve the discovery of cluster-specific peaks, and regulatory motifs as well, when combined with tools like Homer 38 or chromVAR 13 . The success of SCALE can be attributed to the powerful deep generative framework and the GMM to accurately model the high-dimensional, sparse, multimodal scATAC-seq data. Similar to SCALE, a recent scRNA-seq analysis tool scVI also learns latent representation of scRNA-seq data by aggregating information across similar cells using a hierarchical Bayesian model 18 . However, SCALE also applies a GMM to overcome the increased sparsity of scATAC-seq data and more tightly estimate data distribution, thus achieving higher accuracy than scVI on scATAC-seq data analysis. It highlights the necessity and advantage to develop new methods that are optimized for scATAC-seq data, but not to use scRNA-seq data analysis tools.
An attractive additional observation about SCALE is the interpretability of the GMM model. We showed that SCALE could possibly capture biological cell-type-related and potential batcheffect-related latent features in the low-dimensional space. By excluding batch-related features in embedding and clustering, we are able to reduce batch effects. Nevertheless, SCALE is not specifically designed to identify and remove these artifacts from the input data. In the future, we could improve the model to explicitly incorporate variables that are designated for the discovery and removal of batch effects and other possible technical variations.
Methods
Data and preprocessing. Data: The Leukemia dataset is derived from a mixture of monocytes (Mono) and lymphoid-primed multipotent progenitors (LMPP) isolated from a healthy human donor, and leukemia stem cells (SU070_LSC, SU353_LSC) and blast cells (SU070_Leuk, SU353_Blast) isolated from two patients with acute myeloid leukemia 24 . The GM12878/HEK293T dataset and the GM12878/HL-60 dataset are respective mixtures of two commonly-used cell lines 3 . The InSilico dataset is an in silico mixture constructed by computationally putting together six individual scATAC-seq experiments separately performed on a different cell line 3,11 . The Splenocyte dataset 25 is derived from a mixture of mouse splenocytes (after red blood cells removal) and the Forebrain dataset 26 is derived from P56 mouse forebrain cells. The Breast Tumor dataset 23 is obtained from a mouse breast tumor sample, including two plates of tumor cells (Epcam+) and another two plates of tumor-infiltrating immune cells (CD45+) from protein labeling and FACS sorting.
Preprocessing: Similar to scABC 14 , we filtered the scATAC-seq count matrix to only keep peaks in10 cells with ≥2 reads for the InSilico dataset, the GM12878/ HEK293T dataset, and the GM12878/HL-60 dataset, ≥5 cells with ≥2 reads for the Leukemia dataset, ≥50 cells with ≥2 reads for the Forebrain dataset, and ≥5 cells with ≥1 reads for the Breast Tumor dataset. We kept all the peaks for the Splenocyte dataset. We also only kept cells with read counts ≥(number of filtered peaks/50). For the InSilico dataset, there were still almost 90,000 peaks after filtering. For the efficiency of the SCALE model, similar to SC3 33 , we further removed rare peaks (reads >2 in less than X% of cells) and ubiquitous peaks (reads ≥1 in at least (100-X)% of cells).
The probabilistic model of SCALE. SCALE combines a variational autoencoder (VAE) and the Gaussian Mixture Model (GMM) to model the input scATAC-seq data x through a generative process. Given K clusters, corresponding latent variable z can be obtained through the encoder via the reparameterization then to generate sample x through the decoder. It can be modeled with a joint distribution p x; z; c ð Þ, where z is the latent variable and c is a categorical variable whose probability is Discrete (c|π) where P C ¼ j ð Þ¼π j ; π 2 R K . p(z|c) is mixture of Gaussians distribution parameterized by μ c and σ c conditioned on c. Given that x and c are independently conditioned on z, then joint probability p(x, z, c) can be factorized as: which can be transformed to maximize the evidence lower bound (ELBO). The ELBO can be written with a reconstruction term and a regularization term: The reconstruction term encourages the imputed data to be similar to the input data. The regularization term is a Kullback-Leibeler divergence, which regularizes the latent variable z to a GMM manifold. And q(z, c|x) and p(x|z) are an encoder and a decoder, respectively, which can be modeled by two neural networks.
The overall network architecture of SCALE. SCALE consists of an encoder and a decoder. The encoder is a four-layer neural network (3200-1600-800-400) with the ReLU activation function. The decoder has no hidden layer but directly connects the ten latent variables (features) to the output layer (peaks) with the Sigmoid activation function. A GMM model is used to initialize the parameters μ c and σ c . The Adam optimizer 43 with a 5e-4 weight decay is used to maximize the ELBO. Mini-batch size is 32. SCALE also provides a quick mode for large datasets with the encoder structure of two layers (1024-128), and model training with maximum iterations of 30,000 and early stopping when no improvements in 10 epochs. The GMM models are constructed with the Python "scikit-learn" package, and the neural network is implemented with the "pytorch" package.
Visualization. We used t-SNE from the Python "scikit-learn" package to project the raw data or latent features to 2-dimension with random state as 124. We used Python package "umap" to visualize the trajectory cell relationships.
Clustering. We used the K-means clustering method from the Python "scikitlearn" package to cluster the input single cells based on the extracted features. To repeat the result, we set the random seed to 18.
Evaluation of clustering results. Adjusted Rand Index: The Rand Index (RI) computes similarity score between two clustering assignments by considering matched and unmatched assignments pairs independently of the number of clusters. The Adjusted Rand Index (ARI) score is calculated by "adjust for chance" with RI by: ð ÞÀExpected RI If given the contingency table, the ARI can also be represented by: The ARI score is 0 for random labeling and 1 for perfectly matching. Normalized mutual information: NMI ¼ IðP; TÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi H P ð ÞHðTÞ p where P, T are empirical categorical distributions for the predicted and real clustering, I is the mutual entropy, and H is the Shannon entropy. F1 score: Generation and corruption of the simulation dataset. A simulation dataset consisting of 300 cells and 1000 peaks was generated. The peaks formed three clusters, with each cluster containing 100 specific peaks. These specific peaks had a value of 1 or 2 (ratio 1:4) in the cells of the corresponding clusters, and 0 in other cells. Corrupted datasets were generated by randomly dropping out values at different rates from 0.1 to 0.8, followed by introducing random noise by setting values as 1 or 2 (ratio 1:4) with the probability of 0.1. | 8,502.2 | 2019-10-08T00:00:00.000 | [
"Computer Science",
"Biology"
] |
Unlocking the Potential of Smart Security and Surveillance Technology in Prisons: A Brief Review
ABSTRACT
INTRODUCTION
In the ever-evolving landscape of criminal justice, the integration of smart technology within prison settings holds transformative potential, promising advancements in security, surveillance, and overall prison management.The concept of "smart technology in prison" involves harnessing digital innovations such as the Internet of Things, Data Mining, Cloud Computing, and Web 2.0 to enhance the understanding of environmental changes, identify public and community needs, and respond to demands promptly and efficiently [1].The infusion of smart technology into prison settings heralds a transformative era.Real-time data analytics and connectivity fortify situational awareness, promptly detect anomalies, and ensure a safer environment for inmates and staff.Advanced surveillance technologies promise robust monitoring capabilities, while streamlined administrative processes and evidence-based decision-making contribute to more efficient prison management.The integration of smart technology emerges as a strategic imperative, poised to revolutionize correctional operations and address longstanding challenges within the criminal justice system.
Despite the widespread adoption of smart governance initiatives across various sectors, the integration of smart technology in correctional facilities remains a relatively uncharted territory [2].The current state of literature on technology in prisons is characterized by dispersion, with a predominant focus on digital technology implementation.While a substantial body of work addresses the broader context of digital technology within prisons, only a limited number of articles specifically tackle the concept of smart technology.This fragmented landscape underscores the critical need for a comprehensive examination and synthesis of existing knowledge, aiming to offer a more nuanced understanding of the potential benefits and challenges associated with the incorporation of smart technology in correctional settings.This review provides valuable insights into the nuanced aspects of smart technology, addressing issues, challenges, and implementation dynamics, particularly concerning security and surveillance in prison environments.
A thorough literature study is required to comprehend better ideas of smart technology for security and surveillance in prison [3].This research fills a gap in the literature by investigating the concept of "smart technology in prison" within the context of smart governance.The study aims to comprehensively review existing literature to understand the evolution of discussions around smart technology in prisons, emphasizing distinctions from digital technology.It specifically explores the benefits and challenges associated with implementing smart technology for prison security and surveillance.
METHODOLOGY
This study refers to the Kitchenham approach.The Kitchenham comprises three stages: planning, implementation, and reporting.In the planning stage, objectives, scope, and research questions are defined, and a review protocol is developed.The implementation stage involves executing a thorough literature search, applying predefined criteria for study selection, and rigorously assessing study quality.The reporting stage includes documenting the entire process and presenting results using concise and transparent methods.The Kitchenham approach ensures a structured and rigorous approach to systematic literature review.
In the planning stage, the protocol consists of criteria and research questions (RQ).This study uses the PICOC framework (Population, Intervention, Comparison, Outcome, and Context) formula to formulate RQ.An explanation of the PICOC framework can be seen in Table 1.According to the PICOC framework, RQ was obtained regarding smart technology definition, issues, and challenges, and implementing security and surveillance technology in prison.In the implementation stage, there are two criteria for the selection process, namely inclusion and exclusion.An explanation of the inclusion and exclusion can be seen in Table 2.The search strategy is structured based on selecting key terms using alternative words and synonyms in each search string.The publication search was carried out since the study of prison technology appeared, 2013 to 2023.The data source was taken from the Scopus database and then imported into the Mendeley software.The Scopus database provides access to leading journals, conferences, and other scholarly literature.Google Scholar covers various publications from various disciplines and provides broader coverage than Scopus.At this stage, several irrelevant papers were issued based on the title and abstract.The second selection stage is performed from the appropriate paper to determine the final paper.We used item checklist to do quality test of the articles.The item checklist is provided in Table 3.In the reporting stage includes documenting the entire process and presenting results using concise and transparent methods.Data were extracted by reviewing papers with annotated bibliographies.This study uses annotated bibliographic network visualization using VosViewer, a software tool that can create, visualize, and analyze bibliometric networks [4].It uses cluster analysis and color mapping techniques to enhances the exploration of bibliographic networks.Cluster Analysis groups closely related items, revealing thematic concentrations, while color mapping assigns distinct colors to clusters or attributes.These techniques provide an intuitive visual representation of relationships, aiding researchers in identifying themes, influential works, and the evolution of research topics over time within Mendeley sources.The VOS Viewer visualization on the Kitchenham approach is a distinct technique that can be used together to enhance the effectiveness and impact of the literature review and help better understand complex relationships and patterns within the literature.The visualization of the annotated bibliography can be seen in Figure 1.
Figure 1. The visualization of the annotated bibliography
The process of analyzing the final papers uses thematic analysis techniques.Data collected from various articles were analyzed and compared systematically.The authors identified and coded the relevant data units, then grouped these codes into categories according to the scope of the research.Each theme captures the essence of the information in the data related to smart technology, the issues and challenges of smart technology, and the implementation of security and surveillance technology in prison.An explanation of the whole selection process can be seen in Figure 2.
RESULTS
This section will be organized into three sub-sections, each categorizing articles based on type, year, and country of distribution.This systematic categorization provides a structured approach to presenting and analyzing the distribution patterns within the literature, offering a comprehensive overview of the types of articles, their temporal distribution, and the geographical origins of the research.
Distribution of articles based on type
There are 40 articles included in this study.The included articles consist of journal papers (27), conference papers (9), and review papers (4).The details of article distribution can be seen in Figure 3.
Distribution of articles by country
Research on the topic of smart technology in prison was most widely published in the United States with eight articles, followed by the United Kingdom with seven articles, China with six articles, Australia with five articles, Finland with three articles, Germany and Sweden with two articles, and the rest of the articles are scattered within Brazil, Hong Kong, South Korea, Spain, Switzerland, Turkey, Zimbabwe with 1 article of each country.The details of article distribution can be seen in Figure 5. Future research in this field should prioritize addressing these geographical disparities, fostering a more globally representative body of knowledge.Comparative studies examining the impact of smart technology adoption in diverse prison systems could offer valuable insights into effective strategies and potential challenges.Moreover, there is a need for interdisciplinary research that considers the sociocultural, ethical, and legal dimensions of implementing smart technology in prisons.In conclusion, while the existing literature provides a foundation for understanding smart technology in prisons, it also reveals notable gaps.Bridging these gaps through comprehensive reviews, addressing temporal variations, and fostering global inclusivity in research can contribute to a more holistic understanding of the implications of smart technology on prison management and policymaking.This critical analysis serves as a roadmap for future research directions, emphasizing the importance of a nuanced, context-aware approach to advancing the field.
DISCUSSION
According to the results, this section will be divided into 4 (four) sub-sections related to the definition of smart technology in prison, issues and challenges of smart technology in prison, the implementation of smart technology for security and surveillance in prison, and implications of smart technology in prison.
Smart technology in prisons
Smart government entails the effective utilization of Information Technology within governmental settings.It evolved from the concept of e-government, which seeks to leverage technology to enhance the quality and efficiency of public services.These services are delivered through a wellstructured, integrated system that ensures easy access and effective monitoring.While the definition of smart government is still evolving, it essentially involves employing ICT in governmental processes to foster innovation, effectiveness, efficiency, transparency, and accountability.Gil-Garcia's study further defines smart government as a progression from e-government, incorporating elements of innovation and public engagement [5,6].
The accessibility of digital technologies and tools for inmates remains a contentious and sensitive topic.Jewkes and Johnston's research offers a historical perspective on prisons, highlighting how restricted access portrays prisoners as secondary citizens in an era dominated by information technology, likening them to cavemen in an era of light-speed technology [7].The current approach, limiting prisoners' access to digital tools and the Internet, further widens the gap between prison and society [8].In an age where access to technology and the Internet is crucial, denying prisoners these resources deprives them of essential advancements and educational opportunities available to the general populace.This deprivation not only exacerbates social isolation but also impedes their reintegration into society post-release.Knight's work [9] provided insights into how the withdrawal of technology can impact prisoners' relationships with others, giving voice to their experiences.In recent discourse, an increasing number of researchers advocate for a digital transformation within prisons, advocating for prisoners' participation in this process.In many modern societies, digital technologies serve as the primary means of communication with the outside world, highlighting their importance [10].Several countries have begun implementing and developing digital self-service systems in prisons, viewing it as part of a broader digitalization trend or as a feature of smart cities [11].
The digitization process influences the dynamics of interactions and relationships among staff and inmates, along with reshaping staff roles and the broader prison culture [5,12].It's imperative for correctional facilities to introduce targeted digital transformation initiatives, guided by thorough needs assessments and collaborative networking across organizations [12].Effective digital projects must navigate potential risks, fully leverage the benefits of digitization, and demonstrate measurable effectiveness.Aldhaheri and Xia's research [13] highlighted the anticipated hurdles in integrating smart technology within prison infrastructures and underscored the essential considerations prior to adopting intelligent technologies.They reveal that while installing smart technology in prison, the authority should consider its durability, high labor efficiency, safety, and security features [14].Kaun and Stiernstedt [15] examined the benefits and drawbacks of intelligent technology in prisons.Digital technology in smart prisons includes both systems that can help prisoners by giving them access to services and systems built into the prison's infrastructure to benefit the authorities through security features [16].However, more must be done before prisons can offer the same degree of digital services as smart government.Using the intelligent government model to develop smart prisons would be perfect.Correctional facilities should strive to keep pace with digitally advanced institutions like schools, hospitals, and the military.Authors contend that integrating smart technologies for inmates can enhance security and surveillance measures [12].Our study addresses this research gap by offering valuable insights into the complexities and obstacles associated with implementing security and surveillance technologies within prisons.
A novel approach termed "smart" seeks to enhance daily experiences by merging physical and digital components through embedded sensors that interface with intelligent devices [17].This paradigm shift, as described by IBM, involves a thorough examination of our activities using big data to promote productivity and enrich our lifestyles [17].The integration of supercomputers and cloud computing facilitates this connectivity, linking everything through the Internet of Things.Within correctional settings, smart technology represents a pioneering concept that emphasizes digital services aimed at rehabilitation, education, and social reintegration.This approach acknowledges the importance of rehabilitation initiatives that strive to disrupt the cycle of crime by reshaping offenders' mindsets, attitudes, and behaviours [18].According to Lindström and Puolakka [19], smart prison is defined as cutting-edge technology to reduce recidivism rates and facilitate successful reintegration while being costeffective and therapeutic.Some correctional industries try to define the concept of a smart prison by approaching smart prison as a method for increasing operational efficiency, visibility to critical data, and leveraging technology to improve organizational operations dramatically.Smart prisons use digital technology to improve management, reduce costs, and increase safety and security [20].Smart prisons use technologies such as biometrics, radio frequency identification (RFID), and artificial intelligence (AI) to manage prisoners, reduce the workload of prison staff, and improve safety and security.It may also use digital technologies to educate and train prisoners and help them prepare for reintegration into society after release.The concept of a smart prison revolves around establishing a secure and humane setting for inmates, personnel, and visitors alike.Smart prisons seek to enhance the correctional system by leveraging technology and data to gain deeper insights into and effectively manage the dynamics among prisoners, staff, and visitors.This approach aims to not only minimize costs and enhance public safety but also to bolster security measures to deter repeat offenses, facilitate successful reintegration into society, and foster a shift away from recidivism culture.
Issues and challenges on smart technology in prisons
In recent years, the perception of prisons has shifted towards emphasizing their role in separating individuals from society rather than solely as places of punishment.The primary aims of prisons now include correcting behavior, preparing inmates for successful reintegration into society, and reducing recidivism rates.There is a growing movement advocating for the use of digital technology to afford inmates access to education, communication, leisure activities, and opportunities for personal growth, including technological resources.Such technology aids in the delivery of rehabilitation programs, enhances access to healthcare services, and facilitates secure communication with family members outside the prison walls.However, despite the availability of various technological solutions, the adoption of technology in prisons remains relatively slow and often premature.Many current implementations fall short of achieving true digital transformation within correctional systems.Given the crucial role prisons play in society, further research is needed to explore how prison staff can effectively leverage technology to improve their operations.Details outlining the challenges associated with implementing smart technologies in prisons can be found in Table 4.
Ensuring prison security is a challenging matter.There are several challenges associated with implementing smart technology in prisons.Some of these challenges are highlighted below: (1) Cost: Integrating smart technology into prisons can pose substantial financial obstacles, particularly for developing nations.The expenses associated with acquiring and sustaining these technologies often present a formidable challenge [13,15].
(2) Technological Constraints: Smart technology in prisons may be limited by the availability of the necessary technological infrastructure.The need for appropriate technology infrastructure, including power, connectivity, and hardware, can affect the implementation of smart technology in prisons [13,26].
(3) Privacy and Ethical Issues: Using technology in prisons raises concerns about privacy and ethical issues.Surveillance and monitoring devices can infringe on prisoners' privacy rights and dignity, leading to ethical concerns [16,28].
(4) Cybersecurity threats: Smart prison technology can be vulnerable to cyber-attacks, leading to data breaches and security threats.It is necessary to implement appropriate cybersecurity measures to protect the prison's critical data and systems [29].
(5) Resistance to change: Implementing smart technology in prisons can be met with resistance from prison staff and management.The staff may fear job loss or reduced job security due to automation, while management may be hesitant to invest in new technologies [12,15,27,30,31].
Table 5 provides a comprehensive overview of the hurdles encountered in the deployment of smart technologies within prisons.
Challenges
References Cost efficiency for smart technology [13,15] The compatibility of existing buildings with smart technologies [13] Regulatory policy and local standards for smart technology [13,26] Integration of the prison system into smart technology [2,13,16,19,[22][23][24][25][32][33][34][35]] Automated administration [15] Technology ethics [16,28] Changing the behavior of prisoners [12, 22-24, 28, 33-36] Technology acceptance for officer and prisoner [16,22,25] Cybersecurity threats [37] In the United States, Europe, and Australia, there is a burgeoning interest in modernizing prison environments through the integration of digital technology and tablet computers [38,39].Discussions have revolved around the potential benefits of introducing digital self-service devices to empower prisoners with more control over their lives while incarcerated.Over the years, sensor network solutions have been employed for monitoring and security purposes in prisons.A variety of technologies are employed across administration, security, and prisoner education, showcasing the adoption of intelligent technology in prison operations [40].Examples include the utilization of IoT in monitoring systems, which integrate data into global monitoring platforms via wireless sensor networks and cloud computing, as well as the deployment of video surveillance and analytics systems, digital self-service devices, among others.These advancements underscore the increasing utilization of smart prison technology to streamline existing processes and address challenges within prison settings [41].
The predominant focus of applications within prison environments revolves around security and surveillance [42][43][44], particularly evident in Asian countries like China.Here, sophisticated surveillance systems are implemented to continuously monitor inmates, even within their cells [45][46][47].Through a network of cameras and sensors, inmates' movements are tracked in real-time and integrated into an artificial intelligence framework, employing facial recognition and motion analysis technologies.This enables the generation of daily activity reports for each prisoner and the identification of any aberrant behavior [18].Similar surveillance systems are also deployed in Hong Kong, utilizing cameras equipped with analytical monitoring capabilities to detect anomalies such as self-harm or loss of consciousness among inmates, with the capability to promptly alert officers [45].Another application makes use of data obtained from tracking bracelets.This wearable bracelet will track prisoners' vital signs and whereabouts, including biometric data and early warning signs, such as signs that they may have injured themselves [47].Robots monitor prisoner violence and suicide risk in South Korea to help reduce the workload of human guards.Robotic arms that move on the feces of prisoners are also used in Hong Kong prisons to inspect contraband such as drugs.An additional instance of advanced security technology within prisons is exemplified by a comprehensive monitoring system for inmates' phone communications in US prisons [18].This system utilizes speech recognition, semantic analysis, and machine learning algorithms to compile a database of keywords capable of identifying potentially concerning calls, such as discussions involving criminal activities like conspiracy or contraband smuggling.Furthermore, technologies like image and pattern recognition, exemplified by AI-driven video analysis platforms currently in use in select prisons, are increasingly integrated into CCTV surveillance systems.Notably, Liverpool prison incorporates security cameras equipped with artificial intelligence to deter illicit activities such as drug, contraband phone, and weapon smuggling, as well as to identify other suspicious behaviors [48].Additional insights into the implementation of smart security and surveillance technologies in prisons are detailed in Table 6.
Smart Technology Country Security
Face recognition [16] Australia Smart sensor [15] Sweden Fire sensor [13,14] United Arab Emirates Building automation system [14,15] Sweden, United Arab Emirates Robot wardens [16] Australia Recommender system [15,28,49] Sweden, United Kingdom, United States Cybersecurity [29,50] United Kingdom, Australia Surveillance Advanced information system [14,15,16] United Arab Emirates, Sweden, Australia Healthcare system [15,23] Sweden, Australia Smartphone [22,23,26] Nigeria, Australia, Namibia Digital self-service [15,16,19] Sweden, Australia, Finland Digital participation [27] Germany Alert system [15,28,49] Sweden, United Kingdom, United States Digital rehabilitation [16,19,33,34] Australia, Finland, United States, Sweden Artificial Intelligence [15] Sweden Building management [14] United Arab Emirates Video surveillance [14,16,28,51] United Arab Emirates, Australia, United Kingdom Tracking wristbands [15,16] Sweden, Australia Heat, ventilation, and air-conditioning system [14] United Arab Emirates Energy and sustainability system [14] United Arab Emirates VR-technology [32] Greece Data sharing [15] Sweden Phone calls analytics [15] Sweden Video analytics [15,16,36,52,53] Sweden, Australia, Germany, China, India The use of smart technology has brought revolutionary changes in security and surveillance fields.The use of smart cameras and smart sensors can help increase the effectiveness of surveillance [15].With motion detection and facial recognition capabilities, smart cameras can automatically identify suspicious activity or unwanted individuals and immediately notify security guards [16].This enables a fast and timely response to potential threats.Furthermore, intelligent data analysis systems can integrate and analyze data from multiple sources, such as cameras, sensors, and other security systems.Using artificial intelligence, the system can identify suspicious behavior patterns, highlight vulnerable points, and provide early warning of potential security risks [15].Internet of Things (IoT) networks can enhance security and surveillance.Connecting security devices, such as motion sensors, alarms, and smart door locks, to a centralized network allows systems to be managed and monitored remotely [16].It allows users to easily control and monitor the security of their home or workplace through a mobile app or other device.The biometric recognition technology can also be integrated into security and surveillance systems [16].Using fingerprints, retina scanning, or facial recognition can help ensure more secure and controlled access to restricted areas.With this technology, individual identities can be verified accurately, reducing security risks related to stolen access cards or passwords.Finally, strong cybersecurity must complement intelligent security and surveillance systems [37].Protecting data and networks from cyber threats like hacking or malware is critical in this digital era.Therefore, sophisticated, and constantly updated cybersecurity technologies must be integral to any implemented smart technology solution.In conclusion, smart technology offers great potential to enhance security and surveillance systems.By leveraging smart cameras, smart data analytics, IoT networks, biometric recognition, and cybersecurity, we can create a safer and more secure environment in prison.However, it is essential to remember that these technologies must also be managed with proper ethics and privacy, so that we can harness their potential without compromising fundamental human rights values.
Implications of smart technology in prisons
The exploration of the ramifications of smart technology in prison will occur within three primary areas: Ethical Contemplations, Prison Administration, and Policy Development.The infusion of smart technology into prison environments introduces a nuanced spectrum of ethical considerations that demand careful examination [15,54].A critical aspect involves intricately addressing privacy concerns intertwined with surveillance and data collection.Achieving a delicate balance between the imperatives of security and the preservation of individual privacy rights emerges as a pivotal challenge.Additionally, the proactive ethical stance required for deploying biometric systems aims to identify and mitigate potential biases, ensuring fair and equitable treatment of inmates.
The transformative impact of smart technology on prison management unfolds across multiple facets, with security enhancement as a foundational element.Real-time monitoring, automated threat detection, and responsive systems collectively bolster the safety and security of prisons [45,46,53,55,56].Simultaneously, smart technology acts as a catalyst for resource optimization, automating routine tasks, streamlining operations, and empowering prison staff to engage in more intricate aspects of their roles.This not only elevates overall operational efficiency but also fosters a judicious and strategic use of personnel and resources.
The integration of smart technology necessitates a dynamic and adaptive approach to policymaking [13,57].Policymakers grapple with the challenge of aligning legal frameworks with the rapid evolution of technology, necessitating continuous updates to address emerging challenges and evolving ethical considerations.The data-rich environment fostered by smart technology becomes a cornerstone for evidence-based policymaking.Policymakers can leverage these insights to craft informed and responsive policies, spanning correctional practices, rehabilitation initiatives, and holistic inmate management.The overarching goal is to strike a delicate equilibrium, harnessing the transformative potential of technology while upholding ethical principles and ensuring the resilience and efficacy of correctional policies.
CONCLUSIONS
This study significantly enriches the corrections field by conducting a thorough review of security and surveillance technology in prisons.It articulates the concept of a smart prison, aligning the contemporary demand for smart technology with overarching objectives encompassing security, and surveillance.The analysis uncovers multifaceted challenges inherent in implementing smart technology in prisons, spanning from cost considerations and technological design inadequacies to issues of expertise, facility management, staff training, and the strain of accommodating expanding prisoner populations.Policy creators and professionals can glean valuable understandings from this research to guide strategic decision-making and resource distribution for incorporating intelligent technology in prisons.The results endorse crafting policies that carefully address cost-effectiveness, ethical concerns, and the smooth integration of technology.Policymakers are encouraged to focus on crafting regulatory frameworks that not only facilitate the adoption of technology but also uphold ethical principles and ensure robust cybersecurity measures.Additionally, investing in comprehensive staff training programs and cultivating technological expertise within prison administrations emerges as crucial for the successful deployment of smart technology.While this study offers valuable insights, it acknowledges certain limitations.The selection process for articles may have benefited from a more rigorous approach, potentially impacting the comprehensiveness of the findings.It is essential to recognize that the chosen papers and references, while meeting required standards, may not capture the entire spectrum of the evolving landscape of smart technology in prisons.Future research in this domain should delve deeper into the policy and ethical dimensions surrounding smart technology in prisons.Exploring critical factors influencing the acceptance of smart technology within prison environments presents an avenue for further investigation.
1 .
OR intelligent prison OR digital prison OR smart correctional facility OR smart detention center OR smart penal institution OR e-carceration OR e-prison OR eprison OR smart jail OR digital jail OR intelligent jail) Inclusion Criteria Ranging since the study of prison technology appeared on 2013 to 2023 2. Define the term of smart technology prison 3. Explain the issue and challenge in implementing smart technology in prison 4. Outline the purposive smart technology in security and surveillance.Exclusion Criteria 1.Not journal or proceeding 2.Not full-paper version 3.Not written in English
Figure 2 .
Figure 2. The selection process of Kitchenham's approach
Figure 3 .
Figure 3. Articles distribution according to article type 3.2 Distribution of articles by year Between 2013 and 2023, there is only 1 article in 2023, 14 in 2022, 12 in 2021, 5 in 2020, 4 in 2019, 2 in 2018, and only 1 in 2016 and 2015.The details of article distribution can be seen in Figure 4.
Figure 4 .
Figure 4. Articles distribution by year of publication
Figure 5 .
Figure 5. Articles distribution by country The presented literature analysis, encompassing 40 articles on smart technology in prisons, provides valuable insights into the current state of research within this domain.The distribution across publication types reveals a predominant focus on journal papers, indicating a scholarly emphasis on indepth exploration and analysis.However, the limited number of review papers suggests potential opportunities for more comprehensive and synthesized reviews of the existing body of literature.The temporal distribution highlights a recent surge in interest, particularly in 2022, indicating a growing momentum in smart technology research within prison settings.Nevertheless, the substantial gaps in certain years, such as the single article in 2015 and 2016, warrant attention.Further investigation into the factors influencing these temporal variations could illuminate evolving trends or gaps in the research landscape.Geographically, the concentration of research in the United States, the United Kingdom, and China underscores a Western-centric perspective.While these regions may
Table 3 .
Item checklist represent prominent adopters of smart technology in prisons, there is a noticeable dearth of representation from various parts of the world.Research gaps exist in understanding the application, challenges, and implications of smart technology in diverse cultural and institutional contexts, emphasizing the need for a more globally inclusive research agenda.The implications of these findings for prison management and policymaking are noteworthy.The concentration of research in certain countries suggests that smart technology adoption and its associated challenges might differ significantly across jurisdictions.Policymakers can benefit from a nuanced understanding of regional variations to tailor effective and context-specific policies for implementing smart technology in prisons.Additionally, the emphasis on technological adoption in the United States, the United Kingdom, and China may influence the development and sharing of best practices internationally.
Table 4 .
Issues on smart technologies in prison | 6,258 | 2024-06-21T00:00:00.000 | [
"Computer Science",
"Law",
"Political Science"
] |
In Vivo Evaluation of the Antitumor and Immunogenic Properties of Silver and Sodium Dichloroacetate Combination against Melanoma
Posgrado Conjunto de las Facultades de Agronomía y Medicina Veterinaria y Zootecnia, Universidad Autónoma de Nuevo León, Ave. Universidad S/N, Cd. Universitaria, San Nicolás de los Garza, N. L., CP 66455, Mexico Laboratorio de Inmunología y Virología, Unidad C, Facultad de Ciencias Biológicas, Universidad Autónoma de Nuevo León, Ave. Universidad S/N, Cd. Universitaria, San Nicolás de los Garza, N. L., CP 66455, Mexico
Introduction
Targeted therapies have increased the chances of survival for people with melanoma [1]; however, cancer cells present within the tumor favor different metabolic pathways [2]; as a consequence, the tumor eventually becomes resistant to targeted therapies, especially the ones designed against a single target [3].
The development of silver-based therapies is a promising tool in cancer treatment. Silver ions and silver nanoparticles induce oxidative stress, mitochondrial membrane dysfunction, DNA damage, and cytokines upregulation [4]. The exact action mechanism varies depending on the physical and chemical properties of the nanoparticle and the type of cancer [5]. Furthermore, the clinical use of colloidal silver for bactericidal and antiviral purposes proves that this treatment is safe [6,7].
Sodium dichloroacetate (DCA) is a pyruvate analog which interferes with tumor-associated glycolysis (Warburg effect), decreases cancer malignancy, and reduces lactate production by altering cancer cell metabolic pathways [8]. A decrease in lactate counteracts the acidic state of tumoral microenvironment, reducing tumor growth and metastasis [8]. WZB117, a bis-hydroxybenzoate, 2-deoxy-d-glucose, metformin, and DCA reduce glycolysis and block glucose uptake in cancer cells. Under low intracellular glucose levels, biosynthetic pathways, such as nucleotides and amino acids genesis, are interrupted due to a shortage of intermediate molecules, putting a break on cell proliferation. Despite of its use as monotherapy or combined with chemotherapy, few or none adverse effects have been reported [9].
Because of these activities, we evaluated immunogenic cell death as a possible action mechanism, owing to the increasing number of studies that demonstrate that cellular and mitochondrial danger-associated molecular patterns (DAMPs) can be actively released when exposed to external stimuli [10]. The release of alarmins (Hsp70, HSP90, calreticulin, HMGB1, ATP, DNA, and RNA) and tumor neoantigens induce a tumor-specific immune response that eliminates live cancer cells and residual tumor tissue, avoiding cancer recurrence [11].
The main focus of this study was to use silver and DCA as dual-function agents that affect the DNA integrity and mitochondria activity in order to increase the antitumor response in melanoma treatment. Furthermore, this study could serve as a starting point for the next level developmental stage of dichloroacetate-loaded silver nanoparticles targeted pharmacological formulation. 2.3. Cell Viability. Cells (5 × 10 3 cells/well) were plated on 96 flat-bottom well plates and incubated for 24 h at 37°C in 5% CO 2 atmosphere. After incubation, culture medium was removed, and Ag (0.8mM to 6:5 × 10 −5 mM) or DCA (75mM/ml to 750mM/ml) diluted in the same medium were added. The plates were then incubated for 4h at 37°C and 5% CO 2 atmosphere. Thereafter, the supernatant was removed, and cells were washed twice with DMEM/F-12 medium. Cell viability was determined by the resazurin (Alamar Blue) method, and cytotoxicity was expressed as the concentration of 50% cell growth inhibition (LD 50 ). Results were given as the mean ± standard deviation (SD) of three independent experiments. The LD 50 of each treatment was used in further experiments.
2.5. Animals. Female C57BL/6 mice aged 6 and 10 weeks with a body weight around 23 (±2) g were purchased from the Harlan Laboratories (Mexico City, Mexico). The mice were kept at 25-29°C and a 12h light to 12h dark cycle. Food and water were provided ad libitum. The experimental protocol was approved by the Ethics Review Committee for Animal Experimentation of the Biological Sciences Faculty, Autonomous University of Nuevo Leon (San Nicolas de los Garza, Mexico).
Tumor Implantation and Treatment Administration.
Tumors were induced subcutaneously by injecting 1 × 10 6 B16F10 cells in 200 μl of phosphate buffered saline (PBS) solution. Seven days after B16F10 cell transplantation, a noticeable tumor mass appeared, and mice were distributed randomly into four groups (five mice per group). The control group received only saline solution. The DCA group received 50mg/kg of DCA, whereas the Ag group received 28mg/kg of Ag. Saline solution, Ag, and DCA were administrated by peritumor route, daily, for 21 days. Finally, the Ag + DCA group received the same of doses Ag and DCA by peritumor route every other day, alternating between treatments. Tumor length and width were measured weekly, and tumor volume was determined using the equation: L × W 2 , where L is the longest side and W is the shortest side. Animals were euthanized at the study endpoint (21 days), and the tumors were excised for further experiments.
2.7. ELISA for Active NF-κB p65 Subunit. To measure NF-κB p65 subunit activation, nuclear extracts were prepared from 3 × 10 6 tumor cells, using a Nuclear Extract Kit according to the manufacturer's protocol. Levels of nuclear p65 concentrations were determined by a sensitive ELISA assay (TRANS-AM, Active Motif, Rixensart, Belgium).
TNF-α and NO Production.
Tumors were macerated with RPMI and the supernatant collected and adjusted at a concentration of protein by BSA and stored at -20°C for evaluation. TNF-α was measured in the tumor supernatant by enzyme-linked immunosorbent assays (TNF alpha Mouse ELISA Kit; Invitrogen; Thermo Fisher Scientific; Viena, Austria). All assay procedures were performed according to the manufacturer's protocol.
Nitrate/Nitrite assay kit was used to measure the levels of NO in the tissue homogenates of the tumor followed the protocol established by the manufacturer (Nitrate/Nitrite colorimetric assay kit; Cayman Chemical, USA).
Journal of Nanomaterials
where V is the tumor volume, L is the tumor length, and W is the tumor width, same formula as previously used by Rodríguez-Salazar et al. [12].
Journal of Nanomaterials
Lower doses of Ag and DCA were required in the combinatorial setting to achieve DL 25 and DL 50 (Figure 1(c)).
Ag-DCA Induced Tumor
Regression. The administration of Ag, DCA, and Ag + DCA induced tumor volume regression (p < 0:05) in a time-dependent manner, observing a better effect in mice treated with Ag + DCA treatment (Figure 2(a)).
Calreticulin
Exposure in Ag, DCA, or Ag + DCA Treated Cells. Ag, DCA, and Ag + DCA treatments do not induce calreticulin surface exposure in B16F10 cells, as compared to the control (B16F10 untreated cells) (Figures 4(a) and 4(b)).
Discussion
The cytotoxic effect of colloidal silver (Ag), sodium dichloroacetate (DCA), and their combination was evaluated against B16F10 murine melanoma cells. Our results show that Ag has antiproliferative effects against B16F10 cells, as previously reported by our research group [13]. Further reports of the cytotoxic activity of silver against melanoma cells refer to silver nanoparticles, although the proposed toxicity 5 Journal of Nanomaterials mechanism remains the same [14]. DCA also exhibited an antiproliferative effect against melanoma cells. In a similar manner, Rivera-Lazarín et al. reported a dose-dependent viability decrease in B16F10 cells treated with DCA [15].
The cytotoxic activity of Ag and DCA increased when used as a combined treatment. This was expected since the combination of two or more agents is a cornerstone for cancer treatment; it allows to target key pathways simultaneously, achieving an efficacy increase [16].
After observing the increased cytotoxic effect, we evaluated whether our results correlated with an in vivo antitumor activity. At the tumor level, the generation of necrosis was noted; it is worth mentioning that the lesions completely healed in all cases of tumor elimination. Skin lesions can occur due to the overexpression of tumor necrosis factor alpha [17]. Our results revealed higher levels of TNF-α in untreated melanomas and a significant decrease of this factor in response to all of our treatments. TNF-α correlates with melanoma aggressiveness and metastatic potential in vivo [18], and its overexpression has been reported in advanced primary melanomas by Rossi et al. [19].
It is important to mention that TNF-α is a pleiotropic cytokine, and its proapoptotic effects against cancer cells have been widely described [20]; however, melanoma cells resist TNF-α-induced apoptosis through NF-κB and nitric oxide [21].
In this study, our results showed NF-κB and nitric oxide decrease in mice treated with Ag, DCA, or the combination of both, correlating with tumor regression. Wang et al. reported that NF-κB suppresses TNF-α-mediated apoptosis through the activation of the antiapoptotic proteins TRAF1, TRAF2, c-IAP1, and c-IAP2 [22]. On the Journal of Nanomaterials other hand, Salvucci et al. reported nitric oxide production in human melanoma cells, and blocking this production induces cell death in human melanoma [23]. Specifically, nitric oxide inhibits at least seven caspases trough snitrosylation [24]. Despite our observations, it is important to mention that TNF-α, NF-κB, and NO have pleiotropic effects, and their role in melanoma is not well understood. However, we emphasize that our results indicate that these molecules decrease in correlation with tumor regression and wound healing.
Many anticancer therapies have the potential to induce cancer cell death, resulting in tumor elimination and a patient free of malignancy. However, only immunogenic cell death inducers can prevent cancer recurrence. Therefore, drugs that induce immunogenic cell death represent a recent innovation in the field of onco-immunotherapy [25], such is the case of the use of immunomodulator IMMUNEPOTENT CRP, that recently demonstrated this capacity [12]. We set out to evaluate whether our treatments were capable of inducing an immunogenic cell death.
The presence of alarmins in vitro indicates the potential to induce immunogenic cell death [26]. But, despite treatment-dependent increase of HMGB1, HSP70, and HSP90 (but not calreticulin), the vaccination of mice with B16F10 cells lysed with Ag, DCA, or the combination Ag + DCA did not induce immunogenic cell death as evidenced by the tumor appearance in all mice (vaccinated or unvaccinated) after challenge with viable B16F10 cells. Tumor growth indicates a specific immune response was not induced by the vaccines. A reason for this could be that DCA, Ag, and DCA + Ag do not induce the release of DAMPs in a coordinated spatiotemporal pattern; therefore, they lack the capacity to induce cytokines and efficient antigen presentation [27].
In conclusion, the combination of Ag and DCA has potential antitumor properties against melanoma cells; however, the in vivo antitumor mechanism is not immunogenic cell death. Further studies to elucidate the cell death mechanism are important in order to design strategies and combinations with clinical efficacy against melanoma.
Data Availability
Data associated with the manuscript is available upon reasonable request.
Conflicts of Interest
The authors declare that there is no conflict of interest. | 2,462.6 | 2020-11-07T00:00:00.000 | [
"Biology"
] |
Weathering Ambivalences: Between Language and Physics
The chapter engages the nature–culture divide with the generative ambivalences of weathering in both language and physics. Taking the different uses of the enantiosemic and ambitransitive verb as indicative of the human’s fraught relationship with its environment and itself, it analyses multiple ways in which ‘weathering’ can involve subject–object relations, objectless subject–predicate relations, or even subjectless processes, and proposes to think them with mechanics, thermodynamics, and chaos theory. The ICI Berlin Repository is a multi-disciplinary open access archive for the dissemination of scientific research documents related to the ICI Berlin, whether they are originally published by ICI Berlin or elsewhere. Unless noted otherwise, the documents are made available under a Creative Commons AttributionShareAlike 4.o International License, which means that you are free to share and adapt the material, provided you give appropriate credit, indicate any changes, and distribute under the same license. See http://creativecommons.org/licenses/by-sa/4.0/ for further details. In particular, you should indicate all the information contained in the cite-as section above. Weathering Ambivalences Between Language and Physics CHRISTOPH F. E. HOLZHEY
INTRODUCTION
Proverbially unpredictable, the weather presents many challenges and ambivalences both on the level of (human) language and of (natural) science. Many languages have a particular, rather peculiar class of 'weather verbs', which are arguably as 'ill-behaved' as the weather itself. Nothing seems more banal than to speak about the weather, wondering whether it will be warm or cool today, whether it will rain or storm. Yet, linguists still discuss the 'it that does the raining in English and many other languages' . 1 They ask, for instance, whether 'it' is a dummy, expletive pronoun or whether 'it' refers rather to an 'allencompassing', 'total environment' . 2 Perhaps this wavering between 4 WEATHERING AMBIVALENCES all and nothing should be taken as an indication that the subject of the weather is an ill-posed question. Indeed, what Noam Chomsky calls the 'weathering-it' conjures up the kind of 'grammatical habit' and 'seduction of language' that Friedrich Nietzsche insistently considered as the source of misguided beliefs. 3 Following Nietzsche's reasoning, the question of the (grammatical) subject of weather verbs can indeed be said to transport an erroneous and detrimental dualism that takes all change, alteration, and becoming to be conditioned and caused by a radically different, separate subject -essentially the Cartesian ego conceived as substance, being, and free will.
Correcting the Cogito by an 'it thinks' is insufficient for Nietzsche, insofar as the doing remains doubled by a fictive doer -leading to 3 Noam Chomsky, Lectures on Government and Binding, Studies in Generative Grammar, 9 (Dordrecht: Foris Publications, 1981) Good and Evil,§17,p. 19 and Genealogy of Morality, i-13, p. 236. To quote the passages to which I will return more fully, Nietzsche writes in Beyond Good and Evil (1886): 'it is a falsification of the facts to say: the subject "I" is the condition of the predicate "think." It thinks: but that this "it" is precisely that old famous "ego" is only an assumption, an assertion, to put it mildly, and by no means an "immediate certainty." In fact too much is already claimed with this "it thinks": even this "it" contains an interpretation of the process and doesn't belong to the process itself. Here the concluding is done according to grammatical habit, namely "thinking is an activity, to every activity belongs something that is active, therefore-." Following basically the same scheme, the older atomism looked at every effective "force" for that little particle of matter in which it resides, and from which it produces effects, that is, the atom; more rigorous minds finally learned to do without this "earth residuum," and perhaps someday we will even accustom ourselves, logicians included, to doing without this little "it" (into which the honest old ego has vanished)' ( §17, p. 19). In On Genealogy of Morality (1887), Nietzsche further expands on this 'grammatical habit', referring to the 'seduction of language (and the basic errors of reason petrified in it), which understands and misunderstands all effecting as conditioned by something that effects, by a "subject." For instance, just as ordinary people separate lightning from its flashing and take the latter as its doing, as the effect of a subject that is called lightning, so too popular morality separates strength from the expressions of strength, as if behind the strong one there were an indifferent substratum free to express strength or not to. But there is no such substratum; there is no "being" behind the doing, effecting, becoming; the "doer" is merely tacked on as a fiction to the doing -the doing is everything. The people basically double the doing when they have the lightning flashing; this is a doing-doing: it posits the same occurrence once as cause and then once more as its effect. Natural scientists do no better when they say "force moves, force causes" and so on -despite all its coolness, its freedom from affect, our entire science still stands under the seduction of language and has not gotten rid of the false changelings foisted upon it, the "subjects" (the atom for instance is such a changeling, likewise the Kantian "thing in itself")' (pp. 236-37).
CHRISTOPH F. E. HOLZHEY 5 a redundant 'doing-doing' as he notes in a related text, which takes the common separation of 'lightning from its flashing' as example for language foisting in subjects everywhere. Suggesting that the natural sciences succumb to the same seduction when they speak of forces that move and are to be localized in matter, Nietzsche claims that 'more rigorous minds finally learned to do without this "earth-residuum"', and looks forward to the day when we 'accustom ourselves, logicians included, to doing without this little "it" (into which the honest old ego has vanished [zu dem sich das ehrliche alte Ich verflüchtigt hat]) ' . 4 However, such attempts at overcoming the last anthropomorphism -even in impersonal weather verbs with expletive or all-encompassing subjects -risk reinforcing anthropocentric oppositions and generating new human, or indeed overhuman, figures. If talking about the weather is a well-tried manner of establishing sociality, the unpredictable power of the elements, weather conditions, or atmospheric agencies conjure up an abyss of chaos as the sublime ground for (re)newed constitutions of the human.
As I shall suggest, such an anthropogenic function of the weather is sedimented in the verb 'to weather' insofar as this verb seems to be the precise obverse of impersonal weather verbs: in its manifold and multivalent uses, it takes for granted the activity and effects of the elements and considers them as implicit, impersonal background for subjects that are always human or, at least, anthropocentric.
At the same time, I will argue that if weathering has a semantic history that is anthropocentric, its multivalent grammar also points to a process of weathering that not only precedes oppositions of subject and object, culture and nature, or language and science but that can also be understood as the ground from which such oppositions emerge. And while language may not be able to do without the subject, reading the ambivalences of linguistic weathering with those of scientific weathering opens the possibility of re-working time-honoured 4 Ibid., p. 19. Walter Kaufmann's translates the bracket as 'which is all that is left of the honest little old ego', which suggests to my mind more accurately that the ego is still present in this 'it' . See Friedrich Wilhelm Nietzsche, Beyond Good dualistic oppositions so that they come to matter otherwise, that is, in a less anthropocentric manner. The idea here is not that science can avoid the linguistic predicate-subject separation, which tends to substantialize the subject even before it is opposed to an object, but rather that physics, for instance, in seeking to model unpredictable phenomena on the basis of natural laws, conjures up other kinds of 'subjects', which may help keep the grammatical subject from defaulting into an emphatic human subject. 5
CO-CONSTITUTIVE WEATHERING
In their 2014 article 'Weathering: Climate Change and the "Thick Time" of Transcorporeality', Astrida Neimanis and Rachel Loewen Walker propose to create 'weathering' as a concept 'to counter the fallacy of a bifurcated understanding of "nature" and "culture" -or of weather and humans' . 6 They draw on feminist new materialist and posthumanist approaches that highlight the fundamental entanglement, mutual imbrication, and inseparability of 'human and nonhuman natures' . 7 In particular, they invoke Stacy Alaimo's notion of 'transcorporeality', which stresses 'the extent to which the substance of the human is ultimately inseparable from "the environment"', 8 and propose an 'understanding of ourselves as weather bodies': 9 We seek to cultivate a sensibility that attunes us […] toward ourselves and the world as weather bodies, mutually caught up in the whirlwind of a weather-world, in the thickness of climate-time. In short, as weathering. 10 5 In other words, the general strategy could be described as countering a dualism of substance, which would oppose the human to the nonhuman, by a dualism of method -or more precisely a complementary of methods, which takes the risk of universalizing both anthropomorphism (or vitalism) and mechanism in order to attend in either case to the emergence of differences that have not been pre-supposed. Neimanis and Loewen Walker seek to radicalize their notion of weathering further by drawing on Karen Barad's theory of 'intra-action', according to which separable entities (onto)logically do not precede their relations but co-emerge through them. While Alaimo's transcorporeality allows for relations of contiguity, continuity, or immersion, they maintain that intra-action clarifies the claim that 'humans and nonhuman climate and weather phenomena are co-constitutive. We are mutually emergent, coextensive. Together, we weather the world.' 11 Moving from 'Transcorporeal Weather' to 'Transcorporeal Temporalities', Neimanis and Loewen Walker enlist the notion of intraaction to radicalize the collapse and co-constitution of distinctions even further and extend them to space, time, and matter. The principal target remains the belief that human bodies can be separated from their environment. They had already made the intriguing and subtle suggestion that it is not enough to speak of immersion: 'the weather and the climate are not phenomena "in" which we live […] but are rather of us, in us, through us.' 12 Referring to Claire Colebrook's observation that 'our attempts to externalize climate deny the fact that we are already entangled in its forces and flows', they now move to a critique of the 'exteriorization' and 'spatialization of time' . In particular, they object to narratives of sustainability, progress, or apocalypse that 'rely on a linear earth time where past, present, and future make up a time-line of human progression' . 13 It is in order to counter such an exteriorization and spatialization of time that Neimanis and Loewen Walker turn again to Barad's notion of intra-action to consider 'the co-constitutive functionings of matter and meaning that collapse any notion of distinct space and time into an "iterative becoming of spacetimemattering"' . 14 Shifting attention towards non-spatialized temporalities thus ends up involving the collapse of all distinctions -of and within space and time as well as matter and meaning into what one might well call space- 11 Ibid.,p. 564. 12 Ibid.,p. 559. 13 Ibid.,pp. 569 and 567. 14 Ibid.,p. 569 timematteringsemiosis 15 -so as to be able to think their intra-active co-emergent co-constitution. The move from 'Transcorporeal Weather' to 'Transcorporeal Temporalities' does not imply that Neimanis and Loewen Walker abandon weathering. On the contrary, it is meant to lead them deeper into weathering as 'the intra-active process of a mutual becoming' through which 'humans and climate change come to matter' . Thus, they make the striking claim that 'matter is weathering in its making of temporality' . 16 Such a concept of weathering 'means to think of bodies as part and parcel of the making of time […]. Our very bodies, thoughts, actions, and behaviors make the present, past, and future' . 17 Understood as a 'making of temporality', weathering leads them to the provocative 'claim that we are time' -or at least to the notion of 'a time that we weather together' . 18 Time, here, has taken the place of the world in the rallying cry 'Together we weather the world' . Such formulae are intriguing and highly suggestive, but upon closer inspection, the sense of weathering emerging from them is quite ambivalent, difficult to ascertain, and hard to retain. Indeed, they may serve rather as examples for the difficulty of fully grasping the 'profound conceptual shift' that the notion of intra-action represents in Barad's own words. Whereas the more common idea of 'interaction' presumes the 'prior existence of independent entities or relata', the notion of 'intra-action' insists that 'relata do not preexist relations' and claims instead that they 'emerge through specific intra-actions' . 19 Such a definition of intra-action is both compelling and easily repeated. However, articulating or even just thinking relations without preexisting relata -or, for that matter, the very notions of 'preexisting', 'preceding', or processes of co-emergence without a linear (temporal or logical) order -remains hard without getting seduced by the duplications, separations, and reifications of language.
A sentence such as 'Together, we weather the world' no doubt describes a relational process, but insofar as it has a clear subjectpredicate-object structure, it invariably conveys the sense that these identifiable, distinct elements precede their relation in the sentence. As Neimanis and Loewen Walker had previously defined 'weathering' as 'mutual worlding', the sentence should be read as 'Together, we mutually world the world', which becomes even more redundant and tautological if one remembers that 'we are the world' insofar as 'humans and nonhuman climate and weather phenomena are co-constitutive' and 'we are mutually emergent, coextensive' . In other words, by identifying the relata preceding the relation, one arrives at a triplicating but also manifestly circular 'The world worlds the world', while what emerges from the original formulation is not only the separation of 'we' from the 'world', but also a subject-object relation with a defiant predicate resonating well with how weathering is commonly used as a transitive verb when one says, for instance, 'the crew weathered the storm' .
My point here is not to criticize some particular, perhaps unfortunate formulations, but rather to note that when invoking intra-action to emphasize the ontological indeterminacy of anything preceding relations, one should not forget that this is only the premise for the claim that intra-action is meant to account for the co-constitution of separable entities. Such co-constituted entities seem to acquire strong, separable identities -stronger than what is suggested by contiguity, continuity, or immersion, which intra-action supposedly radicalizes. Indeed, the process of intra-active co-constitution is often referred to as a 'coming to matter', 20 but what is remarkable is that the coconstitution happens here in and through common language. If 'Together, we weather the world' has any referent and describes anything, it is the imaginary of a 'we', of a 'world', and of their mutual constitution that the sentence performatively produces. While the notion of intra-20 Neimanis and Loewen Walker, 'Weathering', adopt this pervasive language of mattering in another formulation that works to equate weathering with intra-active processes: 'it is through weathering -the intra-active process of a mutual becoming -that humans and climate change come to matter' (p. 560). However, the premise and claim of intraaction that relata come to matter through the relation rather than preexisting it, is immediately contradicted by the subsequent sentence offering an alternative definition by way of conclusion: 'Weathering, then, is a logic, a way of being/becoming, or a mode of affecting and differentiating that brings humans into relation with more-thanhuman weather' (p. 560).
action may well succeed in dispelling the belief in separate entities being ontologically pre-given, it is far less clear to what extent newmaterialist talk of mattering through intra-active, agential separation does not yield to the seduction of language of foiling subjects into everything and doubling or even tripling processes into a doing-doing or world-worlding world. Again, there is much to be said for taking the risk of anthropomorphizing in order to help balance the alternative risk of anthropocentrism, but it requires critical, methodological self-reflection if the twin risks are to be avoided. The problem with the way in which Neimanis and Loewen Walker create the concept of weathering is not so much that it involves a performative contradiction, which as such may well be unavoidable if language is always performative and enacts divisions into separable entities that appear to have pre-existed their (linguistic) relation. 21 Such a temporal entanglement, whereby language produces what it presupposes, could well be considered part of their critique of 'linear time' and correspond to the 'cultivation of the sensibility of thick time' they propose and describe. However, what would warrant critical reflection is the deliberate gesture of 'conceptcreation', 22 which implies sovereign, anthropocentric subjects that can create a concept such as 'weathering' ex nihilo (even as it explicitly draws on theoretical sources) and disregard the multiple significations that centuries of language use have sedimented into that signifier.
Disclaimers here seem insufficient and function rather as disavowals, denials, or negations in Sigmund Freud's sense of Verneinung, 23 as when the authors distance themselves from both anthropomorphism and human exceptionalism, 24 but otherwise quite consistently seek to reduce distances and deny the relevance of scale, thereby suggesting a 21 For an attempt to move beyond this logic of presupposition arguably characterizing the experience of language from Aristotle to Derrida, see Damiano Sacco's chapter 'The Weathering of the Trace: Agamben's Presupposition of Derrida' in this volume. 22 Neimanis and Loewen Walker,'Weathering',p. 560. 23 Highlighting the difficulty of retain the ambiguity of this term in translation, Jean Laplanche and Jean-Bertrand Pontalis, The Language of Psycho-Analysis (London: Hogarth, 1973) conclude their entry on negation, Verneinung, (dé)négation by highlighting three closely related assertions in Freud's analysis: 'taking cognizance of what is repressed' and engaging in 'a kind of intellectual acceptance of the repressed', and 'thinking frees itself from the restrictions of repression' (p. 263). 24 Neimanis and Loewen Walker,'Weathering', strict symmetry and even equivalence in relations of mutuality and in figures such as 'weather bodies' . In particular, they explicitly question common distinctions between climate and weather based on different time-scales and aim 'to reduce the distance between the enormity of climate change and the immediacy of our own flesh', or again they propose to 'bridge the distance of abstraction [which they criticize in climate discourse] by bringing climate change home' . 25 Similarly, Barad's disclaimer that she does not intend to 'make general statements […] about all entanglements, nor to encourage analogical extrapolation from [her] examples to others' 26 sits uneasily with the far-reaching consequences she and others draw from the notion of intra-action as 'mutual constitution of entangled agencies' . After all, Barad's neologism is only spelled out and specified in a precise, technical sense for the example of quantum diffraction experiments and their interpretation in terms of Niels Bohr's particle-wave complementarity. Barad's point that quantum entanglement is not limited to microscopic scales and can therefore apply also on macroscopic scales in a literal rather than analogical manner, is well taken, but in practice, when entanglement is invoked to deny the relevance of scale, literality seems to be a question of language rather than physics. 27 Seeking to negotiate between conflicting risks perhaps always runs the worse risk of disavowal, that is, of glossing over internal tensions, while reproducing what is to be avoided. Immunizing itself against critique, disavowal fosters a voluntarist decisionism concerning, for instance, the question of when separability is to be rejected in favour of 25 Ibid.,pp. 562 inseparable entanglements and when it is to be embraced as the mode in which anything 'comes to matter' . In order to negotiate conflicting risks, it may be better to focus on the opportunities that are in apparent conflict and envisage the possibility of fully endorsing heterogeneous accounts, even if they are mutually incompatible and cannot be pictured together. Such a possibility is envisaged by Bohr's principle of complementarity, which plays a decisive role in Barad's inspiring advancement of the diffractive methodology that Donna J. Haraway had proposed as alternative to 'reflection' . While I find the development of an ontology of indeterminate matter through a method of complementary diffraction compelling, I would highlight that this approach crucially relies on disparate scales and insist on retaining the method across different scales and fields rather than extrapolate the ontology linearly or even simply unchanged. In particular, it may be productive to think of the relation between the weather and human bodies, nature and culture, physics and language as entangled and perhaps even in some sense as intra-active, but not in the same sense in which electrons are entangled and materialize through the measurement process: unlike entangled electrons, each side of the relation has already come to matter -if anything ever does. The weathering article by Neimanis and Loewen Walker makes it admirably transparent that the appeal of new-materialist theories and notions such as intra-action often lies in opening up possibilities by enabling a 'new imaginary' -a way to 'reimagine our bodies', 'reimagine climate change', and 'reimagining our literal inextricability from that towards which we are called to respond' -rather than in providing a less human-centred, somehow more direct account of matter. No doubt, the 'reimagination of ourselves as weather bodies […] is already a politics' . 28 However, it is more difficult to see how a politics of reimagination and intensive feeling can be reconciled as such with the declared premises and aims of new-materialist and posthumanist approaches, such as the explicit critique of human exceptionalism and of the bifurcation of nature and culture. At least in its general outline, it ultimately seems rather to be a fully anthropocentric, humanist, even idealist politics. Perhaps in response to such reservations, the 2018 28 Neimanis and Loewen Walker,'Weathering',p. 572. article 'Weathering', which Neimanis co-authored with Jennifer Mae Hamilton, appears more anchored and explicit in associating weathering with 'specific feminist, antiracist and decolonial intersectional attentiveness' . 29 For instance, it exemplifies its expanded, 'naturalcultural' understanding of the weather by making contact with Christina Sharpe's notion of weather as 'totality of our environments' and 'total climate' characterized as 'antiblack', and helpfully situates weathering 'between the neo-liberal heroics of resilience and the victim politics of vulnerability' . 30 In these contexts, the image of intra-active, coconstitutive weathering is particularly provocative, and while probing its specific political potential falls beyond the purview of this chapter, I will concentrate on exploring its logic both on the level of language and on the level of physical reality to which Barad's argument lays claim.
WEATHERING THE DICTIONARY
Creating weathering as a concept of co-constitutive worlding is provocatively counter-intuitive insofar as both 'co-constitution' and 'worlding' suggest constructive processes. Even if one takes these processes also as undermining the autonomy of the subject, the most common use of the noun 'weathering' is rather more destructive: The Oxford English Dictionary (OED) defines the noun as the 'action of the atmospheric agencies or elements on substances exposed to its influence; the discoloration, disintegration, etc. resulting from this action' . 31 However, especially the verb 'to weather' is actually remarkably ambivalent. Now usually associated with disintegration, deterioration, and decay -or at least the risk thereof -its first meaning was '1. to subject to the beneficial action of the wind and sun; to air' . 32 The subentries in the OED indicate that 'weathering' was initially understood as a cultural technique encountered in such diverse contexts as hawking (the 'sport or practice of chasing birds or small animals by means of trained hawks'), house-and farm-work (airing linen, drying harvested crop), or handicraft (exposing clay for brick-or tile-making). 33 While the second meaning is formulated quite neutrally in terms of a change through exposure, the several sub-entries indicate that a complete reversal to the weather's detrimental effects has taken place: The final sub-entry (2d), which could suggest a return to a neutral sense of change, only serves to confirm the full reversal in relation to the beneficial weathering of hawking and other cultural techniques. Indeed, 'wearing well' merely seems to mean 'wearing less badly', and the only sample phrase is: 'For outside work, boiled oil is used, because it weathers better than raw oil.' Subsequent entries, which often invoke a nautical context, entail another kind of reversal without entirely restoring the initial, beneficial meaning. 34 Weathering here refers to the ability to resist or escape detrimental, even disastrous effects: 'to get safely round' (3b), 'to withstand and come safely through (a storm)' (4a), 'to sustain without disaster' (4b), or 'to pass through and survive (severe weather)' (4c). The previous meaning of weathering as a deteriorating change is implied here as a threat or risk, but the verb is now used in an opposite sense, namely as successful opposition to such a change.
The final, rather specialized meanings listed in the OED return to specific cultural techniques. One of them is directly beneficial, namely '6. To set (the sails of a windmill) at the proper angle to obtain the maximum effect of the wind-force'; while the other one consists again rather in avoiding detrimental effects, namely the technique in architecture 'to slope or bevel (a surface) so as to throw off the rain' (7).
Of course, many words are polysemic, but the OED definitions suggest that the verb 'to weather' may be regarded, more radically, as enantiosemic, that is, as having mutually opposite meanings. Such words, which are their own antonyms, are more common than one might expect, and include in English, for instance, 'to cleave' (to adhere or separate), 'to sanction' (to approve or penalize/boycott), or 'to rent' (to purchase use of something or sell it). 35 The verb 'to weather' is even more peculiar insofar as it can be said to be doubly enantiosemic and confound oppositions both of activity and quality: Usually signifying a deteriorating change, it can also mean, on the one hand, successfully opposing such a change and, on the other hand, undergoing a beneficial change.
While context often clarifies the intended meaning, the coincidence of opposites in a single word remains highly unstable. Sometimes, enantionyms are true homonyms, that is, words that have different etymological origins and just happened to have become homographs and homophones in the course of linguistic evolution. 36 But the possibility of opposite meanings coming together in a word through condensation, contraction, reduction, or equivocation can also be indicative or evocative of a generative kernel from which such oppositions have emerged in the first place and from which diverse meanings continue to emerge. The puzzling, vertiginous implications of such productivity appears reflected in the ongoing proliferation of terms proposed to name this category of words: from 'Janus word', which seems to be the only one to have found its way into the OED, to 'self-antonym', 'auto-antonym' or 'autantonym', 'con-tronym' or 'contranym', as well as 'enantionym', 'enantiadrome', and 'antagonym' . 37 Going further back, there is the notion of 'primal words [Urworte]', Carl Abel's thesis of their 'antithetical meaning [Gegensinn]', and the analogy that Sigmund Freud establishes with the logic and language of dreams. 38 Highlighting that a 'no' does not seem to exist for dreams, which 'feel themselves at liberty, […] to represent any element by its wishful contrary', Freud notes: [T]he most ancient languages behave exactly like dreams in this respect. In the first instance they have only a single word to describe the two contraries at the extreme ends of a series of qualities or activities […]; they only form distinct terms for the two contraries by a secondary process of making small modifications in the common word [Urwort]. 39 Without necessarily endorsing Abel's and Freud's claims concerning the historical evolution of languages and acknowledging that many enantiosemic words have separate etymologies that only happened to converge into single words, I maintain that the double enantiosemy of 'to weather' is more than a contingent, historical accident. Instead, my suggestion is that it is indicative of an undecidability that lies at the heart of the fraught relationship humans entertain with their environment and allows for manifold articulations. While such an understanding of weathering resonates with the concept Neimanis and Loewen Walker develop by drawing on contemporary theories of transcorporeality and intra-action, I consider it as a case less of 'concept-creation' than of learning from the experience and peculiar logic sedimented and implied in the word's enantiosemic uses. If the appeal of weathering lies in its generativity and beneficial potentials, This scheme can help clarify the grammar of weathering by being more consistent than the OED, which likewise organizes different meanings on two levels and foregrounds the transitive/intransitive distinction, but uses this distinction on both levels and entangles it with other categories such as 'passive' or even 'figurative', 'nautical', and 'architecture' . For example, the OED entries one and four have the heading 'transitive' and five is entitled 'intransitive', while three is entitled 'nautical' and has subheadings '3.a. transitive', '3.d. intransitive', and '3.b. figurative' . 41 Clearly, 'figurative' or 'nautical' exclude neither each other nor the categories of 'transitive' or 'intransitive' . Indeed, in this case the figurative use 'to get safely round' is not just transitive but corresponds well to Merriam-Webster's second transitive meaning of 'to bear up against and come safely through' .
The simplicity of the MWD scheme is deceptive and would require considerable commentary to unfold all its implications. No doubt, the historical use of weathering is richer and more nuanced, but I would like to give some indications of how the MWD's compact grammatical scheme suggests another story, one that is less historical than logical, reductive, but also constructive, ultimately leading to a generative kernel of indistinction and the possibility of as yet uncommon uses.
The compact three-fold definition formalizes well the double reversal from beneficial to detrimental exposure via successful resistance on which I have already remarked. However, what is perhaps most significant is that the MWD definitions conjure up an anthropocentric division of nature and culture. The grammatical subject of weathering indeed always seems to be human or an object of human interest, such as ships that have weathered storms. It is true that one can also speak of rocks that weather even if they have no specific human interest. But although there is a sense in which the activity of the weather or, more generally, 'the action of the elements' are always implied, the possibility of the elements becoming the subject of weathering is not envisaged. In other words, weathering is an anthropocentric predicate insofar as it excludes the weather or the elements as core argument in the subject function.
One might say that weathering denies the agency of the elements by excluding them from the subject position, but my claim concerning the anthropocentric bias of weathering needs to be articulated more carefully. After all, one could say that the elements are excluded as subjects only because their action is taken for granted and already implied in the predicate 'to weather' . Yet, as already mentioned, a characteristically anthropocentric mode of relating to nature is precisely to take nature for granted as a (back)ground for the figure of the human and their actions. 42 This also means that emphasizing the activity or even agency of nature is insufficient to counter anthropocentrism. There is indeed good reason to maintain that the elements are the primary source of all weathering activities, while humans are at best reactive if not entirely passive. But such a simple reversal of activity and passivity merely re-affirms the anthropocentric nature-culture divide as long as it involves predicates that enforce that divide. In other words, weathering is anthropocentric not because it denies agency of the weather as such -it does not preclude the existence of other predicates taking the weather as active subject -but insofar as it is premised upon a divide such that the weather is contained in the predicate (and possibly other arguments the predicate controls), but excluded as subject.
This analysis of weathering's anthropocentric bias indicates a possible strategy for countering it, namely to insist that there is no pregiven divide that would preclude the weather as subject of different forms of weathering; to expand accordingly the use of weathering at the risk of anthropomorphism; and thereby to allow for an exploration of how nature and culture may be understood as more entangled and/or became separated through weathering.
The OED actually allows for such an expanded use of weathering with some qualifications. As already mentioned, it includes the transitive use 2a 'to wear away, disintegrate, or discolour' and gives a couple of examples in which atmospheric phenomena or processes such as clouds, smoke, or percolation do the weathering. 43 At the same time it notes that the verb is then 'chiefly in passive', meaning that the agent 42 Cf. Neimanis and Loewen Walker, 'Weathering', which repeatedly opposes the tendency to externalize weather or climate: 'When we hold onto the belief that we can separate our human bodies from climate (close our doors, resist the winds), we maintain a worldview of relating to the earth, rather than worlding with it. As Colebrook has argued, our attempts to externalize climate deny the fact that we are already entangled in its forces and flows' (p. 567, emphasis in the original); 'The consequence of time's exteriorization is that, we, as bodies, are conceived as only ever in time, subject to forces that carry on beyond and outside of us. The weather/environment serves only as background, thus making for a particular mode of relating to the earth, as though human beings are somehow separate from the natural elements' (p. 568). 43 The sample phrases are: 'The rain-cloud hangs low..overhead; the smoke hovers around; and they weather the fines sculptured surface' and 'It [sc. percolation] acts also very powerfully in weathering the rocks through which the water passes' ('weather, v.', in OED Online; ellipsis and square bracket in the original).
is in a peripheral function or omitted entirely: the rock is weathered by persistent rain or it is weathered tout court. 44 While one sees here how the agency of the weather is at once presumed and pushed into the background, this is only the beginning of what becomes visible and available to analysis when this less common use of weathering is included. In particular, I would like to ask to what extent this use of weathering is distinct from the other uses or can be related to them, and ultimately what the different weathering verbs have in common other than their name.
AMBIVALENT WEATHERING OF A STORM I propose to augment the three-fold scheme of the Merriam-Webster dictionary by explicitly adding the missing transitive use of weathering, which takes the weather or the elements as subject, and place it alongside the second transitive meaning of weathering (see Table 1 below). Sample sentences of these two uses have the same basic subjectpredicate-object structure and differ only insofar as the weather is, in one case, the object and, in the other, the subject. Corresponding formally to a simple interchange of subject and object, they can be combined into a single, seductively simple sentence, such as: 'I weather the storm as the storm weathers me.' 45 This sentence could well be taken as paradigmatic for a transitive sense of weathering that relates nature and culture without being anthropocentric. Indeed, the sentence suggests symmetry, relationality, mutuality, perhaps even a form of co-constitution. It is as if the transitive subject were here not atmospheric action or the weather but once again human beings, that is, as if the whole second definition were ultimately just the malignant counter-part to the first definition of 'subjecting to the beneficial action of the wind and sun' . 45 Note that dictionaries tend to define verbs without specifying the subject, as if the verbs could be universally used for any subject, while they do often specify the kind of object that comes with certain uses: 'to air (linen, etc.)', 'to expose (land, clay for brick-or tile-making)', 'To set (the sails of a windmill)', 'to slope or bevel (a surface)', etc. In order to mark the anthropocentrism veiled by the universalism of the subject, I find it necessary to specify not only the object but also the implied subject, using in this case 'I' and 'the storm' .
However, there is something instructively wrong here -not in the introduction of posthumanist weathering nor in the sentence combining it with the more common, anthropocentric transitive use, but in being led astray by the polysemy of both the predicate 'weathering' and the conjunction 'as', which here should be read temporally or causally rather than as establishing an equivalence. I may manage to weather the storm while it weathers me, but I do not weather it 'just as' it weathers me.
The suggestion of symmetrical co-constitutional relations is actually plausible for similar sentences with predicates such as fighting or attracting: A arguably cannot fight, attract, or repel B if B does not also fight, attract, or repel A -at least not if one understands such verbs in a mechanical sense cohering with Isaac Newton's third law 'actio=reactio': every action not only elicits a reaction but is precisely equal (and opposite) to it, which means, among other things, that neither comes first, and it is arbitrary which one is called action and which reaction. A simple example would be the mutual gravitational attraction between earth and moon, or the (electromagnetic) repulsion preventing a building from falling to the centre of the earth and keeping it on the ground.
However, applying Newton's third law can be tricky. The law resonates strangely with Michel Foucault's dictum '[w]here there is power, there is resistance' and with his relational understanding of force and power that views resistance as a condition for power relations rather than a passive reaction to a power that that would precede it. 46 Yet, a balance of attractive and repulsive forces between two bodies by no means implies that a building could not collapse under its own weight. Two different kinds of forces, such as gravity and electromagnetism (ensuring rigidity), are at work here, and there is no law of equality or reciprocity for their relationship. Instead, they are quite independent from one another, each action eliciting a reaction in the other body, and there is equilibrium only if and when they balance each other out.
WEATHERING AMBIVALENCES
I may weather the storm while it weathers me, but there is no guarantee that I will succeed: although the two instances of weathering in this sentence are equally transitive, they do not only exchange subject and object but also have no necessary relationship between them. [To this extent we are far from being coextensive, let alone co-constitutive.] The deceptive symmetry can be grasped even more compactly through the ambivalent formulation of 'the weathering of a storm' . Grammatically, the genitive can be subjective or objective, that is, it can be the agent or the patient to yield a storm that weathers things or a storm that is being weathered. There may be symmetry in grammar but there is none in the action. The difference here is not so much in scale -a storm is generally much larger than anything that may weather it -but in kind: a storm threatens everything within its scope with death and destruction, whereas that which weathers the storm only saves itself and need have no effect on the storm.
Transitive weathering thus splits into two meanings that are quite independent from one another even if they are equally relational and mutually opposed. To make this explicit, one could write as defining phrase '2α I weather the storm while 2β the storm weathers me' . The grammatical reversibility of weathering -the possibility of interchanging subject and object in its transitive use -therefore does not imply relations of symmetry, mutuality, and co-constitution, as it does with verbs for fighting or mutual attraction and repulsion. Instead it is the result of weathering's specific enantiosemy that makes it appear as its own reciprocal complement. What I mean by this is that transitive weathering behaves in many ways like predicates that change into a contrary, reciprocal predicate when their subject is interchanged with another argument and that form couples such as 'attack-defend', 'inflict-suffer', or 'show-watch', except that it takes the same word for both meanings. In other words, the grammatical reversibility of weathering implies the couple ' 2α weathering-2β weathering', where 2α weathering involves self-preservation, sustained identity, and survival, and 2β weathering denotes destruction, disintegration, and death. Reducing even further, an equation of life and death could be said to lie at the core of weathering's enantiosemy.
FANTASY AND WEATHERING WEATHERING
Insisting on transitive weathering as enantiosemic and split into two different meanings is ambivalent insofar as it can both re-enforce ontological divisions and suggest different strategies of countering them. The division is reproduced if the contrary meanings are understood to be uniquely determined by the subject, that is, if a human(-centred) subject implies self-preserving 2α weathering and an atmospheric subject implies destructive 2β weathering. However, the division is undermined by insisting not merely on the atmosphere's agency, but on the possibility of human and non-human agents each being capable of being the subject of both 2α weathering and 2β weathering.
The enantiosemy of weathering -joint by common use and experience -makes it rather difficult to imagine and convey the possibility that I could weather the storm just as it weathers me, that is, that I 2β weather it, wear it out, and make it decay; or that the storm could weather me just as I weather it, that is, that it 2α weathers me, resists being worn out, and safely comes through all my attempts to annihilate it.
At the same time, the enantiosemy of weathering may be suggestive in indicating that such exercises in logical permutations and pedantic differentiations may be unnecessary and that there actually is a continuity between contrary meanings, that is, that some common, perhaps non-differentiated, ground exists from which they emerge. While such a continuity goes counter a logic of non-contradiction and is therefore hard to think, it can be dreamt and, in some languages, also said.
What I am proposing here is that the enantiosemy of weathering, its multiple reversals, and grammar can be read with Freud's remarks on antithetical primal words in his Interpretation of Dreams and his analysis of the vicissitudes of drives, which he tightly links to 'grammatical transformations' . 47 At one point, he derives masochism from sadism, which he considers as a 'pair of opposites', through a change from active to passive aim, which he calls 'reversal into its opposite', and through an interchange of subject and object, which he describes in two steps: first, the object of sadism is given up through a 'turning round upon the subject's own self ' and secondly, another person is sought to take over the role of the subject. Noting that only the final stage corresponds to what is commonly called masochism, Freud also highlights the necessity of assuming the existence of the intermediate stage, which he explicitly describes in grammatical terms as a change from the 'active voice […] not into the passive, but into the reflexive, middle voice' . 48 Jean Laplanche scolds Freud here for clouding his remarkable grammatical analysis by confusing the reflexive voice (e.g., to hit oneself) and the middle voice, which is somehow between or beyond the active and the passive. Unlike Ancient Greek, English and most other modern language have no verb form for the middle voice, but Romance languages, for instance, often approximate it through a particular use of the reflexive. Laplanche invokes the French example 'se cogner [to knock oneself]' and distinguishes between an accidental knocking oneself against a chair in the dark (corresponding to the middle voice) and an intentional knocking of one's head against the walls (which is more properly reflexive).
While Laplanche privileges the properly reflexive form for its clear distinction of subject and object over the middle form, where the terms 'remain in something of a state of coalescence', 49 in the context of my chapter, the intermediate stage between opposites is most promising precisely when it coalesces rather than distinguishes subject and object, activity and passivity. An intermediate stage characterized by a 'reflexive, middle voice' plays a pivotal role also in Freud's discussion of the pair of opposites 'scopophilia [voyeurism]-exhibitionism', where he links this stage to a much earlier, autoerotic stage, which he understands as the 'source of both the situations represented in the resulting pair of opposites' . 50 Questions of original autoeroticism and of a primary, reflexive masochism have been much debated as Freud remains notoriously 48 Freud,'Instincts and Their Vicissitudes',in The Standard Edition of the Complete Psychological Works of Sigmund Freud, ed. and trans. by James Strachey, 24 vols (London: Hogarth, 1953-74), xiv: On the History of the Psycho-Analytic Movement, Papers on Metapsychology, and Other Works (1914-1916) (1957. 49 Laplanche,Life and Death,p. 143n7. 50 Freud,'Instincts and Their Vicissitudes',p. 130; emphasis in the original.
-but also overtly -ambiguous, contradictory, or undecided about them. 51 In a beautiful essay on original fantasies, fantasies of origin, and origins of fantasy, Jean Laplanche and Jean-Betrand Pontalis have insisted on this question for an understanding of the origin of sexuality in the properly psychoanalytic sense as a deviation from biological needs 'into the field of fantasy' and/or through a 'breaking in of fantasy ' . 52 Rather than speculating further on the onto-or phylogenetic origin of an emergent splitting off of a specifically psychoanalytical -and human -order from the order of nature, I will return to the MWD's three-fold anthropocentric definition of weathering and my pairing of the central transitive definition with a weathering that -at the risk of anthropomorphizing -takes the elements as subject. Perhaps even more speculatively, I would like to propose that this scheme can tell another story, which, rather than by emergent splittings, proceeds by a progressive reduction that may well end up turning weathering into a primal weather verb that does not even have a subject.
PROGRESSIVE VALENCY REDUCTION
The story I propose progresses by reducing what linguists call a verb's 'valency' . This notion is taken from chemistry, where it indicates the number of bonds an atom can establish with other atoms such as hydrogen, and refers to the number of arguments controlled by a predicate. Transitive verbs are typically divalent -taking a subject and an object -or trivalent, when they also take an indirect object, while intransitive verbs are typically monovalent. Most languages have valency-lowering and valency-raising mechanisms, such as the passive or reflexive, which lower the valency, or the causative, which raises it: 'to make someone do something' . 53 Like the historically oriented OED, the MWD begins with weathering as a cultural technique. Grammatically, this transitive use of weathering is divalent: in weathering a hawk or weathering linen there are two arguments: subject and object. Yet, as the definition 'to subject to the action of the elements' suggests, the verb's valency is effectively increased to three. Two agents indeed seem to be present, not only the grammatical, usually human, subject of the sentence, but also the weather, the air, the sun, or other atmospheric elements. Although grammatically divalent, the historically first use of weathering thus involves a triangulation: to weather hawks or linen means that one has the weather weather them.
On this view, the second transitive use of weathering involves a reduction of the number of predicate arguments, leading to a semantically as well as grammatically binary subject-object relation: to weather a storm or crisis. However, upon closer inspection, it is hard to see how this transitive use could be obtained from the first one through valency-reduction, which points rather to its opposite 2β: I weather a hawk = I have the elements weather a hawk → the elements weather a hawk. Proceeding to the third meaning in the MWD involves a further reduction of valency. In rocks that weather, there is a subject but no object and weathering as an intransitive verb is monovalent. Verbs that can be both transitive and intransitive are sometimes called 'ambitransitive' or 'labile' verbs. 54 But here, too, it is hard to see how such a use could be derived from the original transitive use 2α, whereas it comes rather easily from 2β: The storm weathers me = I am weathered by the storm ≈ I am weathered → I weather. There is something quite remarkable in the last step from the passive (which is already monovalent) to the intransitive insofar as grammatically it involves a reversal from the passive to the active. Yet, it is linguistically not that unusual and is the mode of valency reduction of so-called 'patientive ambitransitives' . A standard example in English is the verb 'to break': I break the cup = The cup is broken by me ≈ The cup is broken → The cup breaks.
Adding the use 2β thus facilitates an understanding of the threefold MWD scheme in terms of a progressive reduction of valency. In a way, this is just the result of 2β making explicit the 'action of the elements', which 1 and 3 take for granted but place in the background. One might indeed object that if one considered the first use (i.e., weathering hawks) as trivalent, one should now also say that the intransitive verb is effectively divalent insofar as the weather remains implied as agent.
In other words, one might say that the use 2β was not entirely absent in the MWD but effectively contained in 3. If this observation can further the strategy of countering the anthropocentric nature-culture divide by insisting on the agency of the elements, I would now like to suggest that there is also something to be said for taking the grammatical reduction of valency from 2β to 3 more seriously.
While transitivity is premised upon a separation of subject and object, which easily aligns with nature and culture -regardless of whether one says 'I weather the elements' or 'the elements weather me' -intransitivity can do without a system-environment distinction and makes no reference to the environment. Is there not indeed a sense in which everyone and everything weathers and does so 'by itself ', without any particular external influence? This would mean understanding intransitive weathering not as an implicit or 'agentless' 28 WEATHERING AMBIVALENCES passive that disavows the 'action of the elements', taking it for granted and placing it in the background, but rather as a kind of reflexive or middle voice, perhaps as an activity that turns against itself when it finds no object, or, vice versa and more radically, as an 'anticausative' verb from which a subject-object distinction co-emerges without being implied. 55 It could even mean understanding intransitive weathering as a form of reflexive masochism or, with Freud's re-articulation of primary masochism in Beyond the Pleasure Principle, as a manifestation of the death drive.
Weathering in this sense is a correlate of ageing or the passage of time, as Romance languages using the same word for 'weather' and 'time' suggest. 56 Even though some things age well and growing up and maturing are often welcomed, and even though some storms and crises may be weathered, deterioration, decay, and death appears as ineluctable fate of all temporal existence. In other words, 'the action of the elements' undergone or endured in intransitive weathering may well be just a way of speaking of a law of nature. A famous law quickly comes to mind: the second law of thermodynamics, the law of increasing entropy, which is often invoked to define the arrow of time.
Perhaps there is a good reason, then, that using weathering for 'the action of the elements' is uncommon and that this action is taken for granted or even dismissed. Not only would it be a pleonasm to say that 'the weather weathers', but following Nietzsche one could insist that the process of weathering is all there is and that a subject is added only out of grammatical habit producing a 'doing doing' . 'The weather' and 'the elements' would then only be abstract fictions without real existence. All there is then is the process of weathering 55 Cf. Dixon and Aikhenvald on the distinction between the 'prototypical passive', where the agent is named, the 'agentless passive', where the agent is implied -'the glass is broken (implied: by someone)' -and the 'anticausative', where an agent is neither stated nor implied: 'the glass broke' (p. 7). (In this case, the object is explicitly assumed as grammatical subject, and my suggestion of co-emergence gestures already further towards an avalent 'it weathers' .) Concerning the term 'middle', the authors warn that it is used with a 'frightening variety of meanings' -including the anticausative and reflexive -lacks in 'typological clarity', and often amounts to a '(general) intransitivizer' (p. 11-12). 56 On the association of weather and time, or weathering and change, see Niccolò Crisafi and Manuele Gragnolati's chapter 'Weathering the Afterlife: The Meteorological Psychology of Dante's Commedia' in this volume.
that takes place in everything -in the elements themselves -without the intervention of some subject or agent. In other words, weathering would be an avalent weather verb after all: 'it weathers', and even speaking of an 'it' may be already saying too much.
But what about the more positive meanings of weathering: the possibility of weathering storms and crises, or even the beneficial actions of the weather? Do they have an intransitive, even avalent correlate or substrate from which they can be imagined to emerge, or are they in inherently bound to human subjects?
No doubt the positive evaluation of weathering is to a large extent anthropocentric. The outright beneficial character of the elements in the first use indeed seems due to cultural mediation -be it because we channel the actions of the elements, because the weathering is good for us but not the hawk (not to speak of sun-dried tomatoes or raisins), or because weathering hawks means temporarily lifting their confinement and putting them back in their element. Most strikingly, the OED counts the meaning 'to expose […] to the pulverizing action of the elements' under the rubric 'to subject to the beneficial action of the wind and sun', which sounds violent and could even appear sadistic if the utility for 'brick-or tile-making' were not mentioned in my ellipsis.
However, there is also a sense in which beneficial weathering can be understood with respect to an entity's identity, self, or unity as a system rather than human utility. Garments that are being weathered lose their odour and other contaminations, making them less susceptible to moths, for instance. Weathering here purifies; it takes off what is not essential, making the object more durable. Clothes that are dried in the open air do not rot, nor will grapes and tomatoes, for instance. Perhaps this weathering takes out some life -or the potential for unwanted life, like mould -but even disregarding human utility, it seems distinguishable from entropic processes of disintegration and dissipation that involve homogenization and a loss of structure and order. Indeed, it lays bare what persists, such as an underlying structure otherwise hidden under a superficial, perhaps only ornamental layer (as it is in plastered buildings or artworks).
In other words, weathering can accentuate a thing's identity in its constitutive difference from the environment by enacting the differentiation of that which in it weathers and that which does not. Or, more suggestively, weathering constitutes identity by enacting a differentiation between two kinds of weathering, between 'what 3 weathers' and 'what 2α weathers the elements' .
Again, one could insist that intransitive 3 weathering is really a passive form that implies the environment as agent and therefore presupposes a self that is distinguished from the environment. But one might also take the valency reduction more seriously and extend instead the intransitive meanings of weathering. In this way, if I argued that the ordinary intransitive use of weathering forms with 2β weathering a patientive ambitransitive pair, 2α weathering could well be considered part of an agentive ambitransitive pair, where the intransitive simply omits the transitive object. Examples in English include eating but also winning. Not only does 'to come safely through' -the MWD definition for 2α weathering -resonate with winning, but it is also itself intransitive, and perhaps one can hear in intransitive weathering not just resignation but also defiance: 'I weather!' . 57 With such a doubling of intransitive weathering, a sense of identity becomes conceivable that relies not on a pre-given distinction from the environment -the subject-object opposition of transitive verbs -but that is instead constituted through the internal differentiation of 'what αa weathers' and 'what βp weathers' . The additional superscript is necessary because one can also imagine another permutation of having the two transitive meanings of weathering enter the two varieties of ambitransitivity, arriving at both a βa weathering and a αp weathering of the elements (subjective genitive). 58 For a full account of self-constituted identity that does not rely on ontological divisions, the intransitive weatherings must not, of course, predetermine their 57 While defiant weathering may provide a (politically) unattractive model insofar as it approaches 'neoliberal resilience' (Neimanis and Hamilton,'Weathering',p. 83) No doubt, we have long left behind even less common uses of weathering, and there is also no reason to assume that every verb should be doubly ambitransitive. However, the combinatorics of valency reduction on the one hand raises critical questions worth exploring further. What would it mean if attempts at going beyond fixed subject-object binaries led to a proliferation of monovalent or eventually avalent weatherings? Does it suggest that the problem of binaries repeats itself in a different guise? Or could one take weathering's enantiosemythe use of the same word 'weathering' for all the different meanings that can be kept separate only with considerable analytical effort -as a hint that weathering defies language, logic, and reason, that one can only attempt to get to a fuller account by bringing together all these different, contrary meanings even though they cannot be pictured together, and that it forms an ultimately ungraspable kernel from which all the distinctions emerge that are retroactively used to project out complementary aspects? On the other hand, I would maintain that the linguistic combinatorics can be productively related to different material phenomena and some thorny questions in the history and philosophy of physics. I have already suggested that thermodynamics and its second law of irreversible entropy production and dissipation strongly resonates with ubiquitous intransitive weathering, which now should be specified as βp weathering. However, the second law continues to raise profound questions and elicit much debate. Insofar as it defines an arrow of time -an irreversible tendency towards larger entropy, disorder, equilibrium, and homogeneity -it confirms the intimate connection between weather and time suggested by Romance languages. As such, it is arguably necessary for any kind of phenomenal experience.
Yet while often regarded as the most fundamental law, the second law is alternatively, sometimes even simultaneously, considered as re- ducible to more fundamental laws at lower scales and/or in profound conceptual conflict with these laws. 59 The basic reason for the conflict is that physics tends to theorize fundamental laws as reversible and conservative rather than dissipative, that is, despite its early twentiethcentury revolutions, it continues to follow the paradigm of Newtonian mechanics and its laws of inertia and conservation, which it extends from energy and momentum to other quantities and information. The basic conflict between classical mechanics and thermodynamics can be aligned with the difference between α-weathering connoting persistence and β-weathering connoting decay, deterioration, and loss. More precisely, the conflict here corresponds to the difference between αa weathering (a persisting obtained from the active 'I α weather [the storm]') and βp weathering (a decaying obtained from the passive 'I am being β weathered [by the storm]'). Furthermore, there is the counter-intuitive claim -forcefully advanced especially by Ilya Prigogine, the 1977 Chemistry Nobel Prize winner for his work on dissipative structures -that it is the second 59 While the literature on this topic is vast, a helpful point of entry may be Craig Callender, 'Taking Thermodynamics Too Seriously', Studies in History and Philosophy of Science Part B, 32.4 (2001), pp. 539-53. law that accounts for the emergence of order and structure. 60 Prigogine's claim is remarkably hard to grasp in its provocatively general ramifications. It is often emphasized, for good reasons, that a key point for making the self-organized order and growth of living organisms consistent with the second law is to consider them as open systems that exchange energy with their environment and, more specifically, discharge more entropy into their surroundings than they ingest. 61 To this extent, one could say that the key insight is that organisms behave like little storms that βa weather in the form of exhaling, sweating, defecating, and urinating. However, Prigogine goes further to suggest that self-organization is not merely compatible with the second law but follows from it, that is, that a system's identity and organization emerge and persist thanks to dissipation. It is as if α-weathering of mechanical persistence were subsumed under thermodynamic β-weathering and pushed towards an ambiguous vitalism. Such a move can already be found in Spinoza's notion of conatus as a 'striving to persevere in one's being', which is modelled upon mechanical inertia but tends to retain sense of desire and teleology of upward progression. 62 The basic issue remains how anything can emerge to grow in the first place. Although the βa weathering verbs of excretion are all intransitive, they -and the very notion of an open system -all presuppose a distinction between system and environment. I would argue that it is therefore misleading to say that a system's identity and organization can emerge through dissipation as such, and maintain instead, as already indicated, that a differentiation between different modalities of intransitive weathering is required, which can now be specified as a differentiation of persisting αa weathering and dissipative βa weathering. Among the most familiar and deceptively simple examples from physics with which to test out these ideas is the formation of crystals in a super-cooled liquid. The process depends as much on the formation of persistent bonds in the emergent solid as on the flow of energy into the surrounding liquid, where it dissipates to increase the overall entropy. Furthermore, the liquid keeps threatening to dissolve the nascent crystal again, and it would quickly do so if its relative volume were not so large that it can easily absorb and dissipate the crystallization energy without being significantly affected by it. One could well speak here of a mutual weathering and even of the co-emergence and coconstitution of 'things' and their environment. Indeed, crystallization provides a model with which to think through the intriguing but hard to grasp notion of intra-action as process through which separable entities emerge and materialize in relations without preexisting relata. The model of crystallization helps understand, for instance, that mutuality, co-extensiveness, and co-constitution need not imply symmetry or similarity between the emergent entities, nor do they question the solidity and durable persistence of what emerges. It is helpful in clarifying the roles of indeterminacy, virtuality, and contingency, but also predictability. A closer analysis could thereby help sharpen the larger claims and expectations attached to such notions as intra-action, coemergence, and co-constitution, even if -or rather precisely asit may disappoint insofar as the general trajectory is quite predictable and progresses linearly from homogeneous liquid to immersed crystals, leaving only -though, to be sure, significantly -their actual location and shape open to a range of variation.
Much the same could be said about the more dynamic phenomenon to which Prigogine repeatedly refers as paradigm for dissipative structures, namely the spontaneous emergence of so-called Bénard convection cells in a homogeneous layer of liquid that is heated from below. This model relates in several ways to the multiple, ambivalent meanings of weathering. Above all, it provides the basic mechanism for the formation of wind patterns in the earth's atmosphere under the effect of the sun's influx of heat, which eventually gets radiated in a more isotropic and entropic form into outer space. But its closer analysis also gives an indication of what could be meant by αp weathering, which in my analysis of the grammatical possibilities of weathering remained perhaps most obscure: what does it mean to consider α weathering as a patientive ambitransitive, that is, to say that the storm αp weathers insofar as it is weathered by me? How could a storm possibly be affected by my weathering it?
The attempt to model the weather through numerical calculations of the equations governing atmospheric convection made meteorologist Edward Lorenz a pioneer of so-called chaos theory. Repeating his computer simulations, Lorenz discovered that the sequence of convection patterns is extremely sensitive to initial conditions, making the weather unpredictable for all practical purposes even when it is assumed to be governed by fully deterministic and reversible equations. 63 This defining characteristic of chaos as understood in mathematics and physics -the appearance of randomness and chance in fully deterministic systems due to sensitivity to initial conditions -is at the heart of the famous 'butterfly effect' . As Lorenz himself highlights in retrospect, this expression has a 'cloudy history' . 64 Often traced to his 1972 paper entitled 'Predictability: Does the Flap of a Butterfly's Wings in Brazil Set off a Tornado in Texas?', it refers to the possibility of small causes having large effects. 65 It suggests, in particular, that what an organism does to persist can have far-reaching effects -it can even set off a storm somewhere on the globe, or also prevent it.
As Lorenz notes, 'Perhaps the butterfly, with its seeming frailty and lack of power, is a natural symbol of the small that can produce the great.' 66 However, the agency that the butterfly effect attributes to everything, including the small and meek, is ultimately deeply ambivalent: its potentially enormous effects could seem to show the irrelevance of scale, make everything possible, and even feed fantasies of omnipotence, but they are utterly unpredictable and knowable only retroactively. Furthermore, even if an agent knew what it was doing, any intention, plan, or strategy would be foiled by myriads of equally powerful and unpredictable agents. Whatever happens ultimately depends on the constellation and behaviour of all agents, which are so thoroughly entangled as to make the very notion of agency meaningless.
If anything, it is the weather that gains agency through the entanglement of zillions of effective butterflies. As everything weathers the weather, the weather is weathered and therefore could be said to αp weather in an entirely unpredictable manner, behaving, for all practical purposes, as if it had a mind of its own. Furthermore, if αa weathering is above all about self-preservation, survival, and persisting self-identity -the storm figuring only as a grammatical object but not as the object towards which one's action is directed -the point of chaos theory could be said to assert much the same of αp weathering.
Chaos as understood by mathematics and physics is not pure randomness. Instead, it is defined in terms of a causally fully determined dynamic and involves a peculiar and ambivalent kind of unpredictability. If the figure of the butterfly ultimately established itself for chaos theory, it is arguably because it symbolizes not only the 'small that can produce the great', but also another defining aspect of chaos, namely the presence of so-called 'strange attractors', which constrain and govern the apparently random behaviour of chaotic systems. An attractor represents a small subset of possible configurations or dynamic patterns towards which a dynamic system is 'attracted' no matter where it begins. In the simplest case, the attractor is a point, such as the bottom of a well where a ball will end up whichever way it is thrown in. The attractors of chaotic systems are 'strange' insofar as they have a very complicated, so-called fractal geometry (Figure 1 shows an example). This strangeness accounts for the appearance of randomness as a system will be attracted to a different part of its attractor when the initial conditions are slightly different. Yet the attractor itself can be understood as a system's signature, that is, as a robust, intrinsic telos towards which the system will evolve and to which it will return when disturbed.
Among the earliest and most famous attractors is the one that Lorenz found while modelling some aspects of atmospheric convection. In its by now customary representation, it happens to resemble a butterfly (see Figure 2), which provides another genealogy and meaning of the butterfly effect. 67 In this case, the butterfly would be the symbol less of small external causes having large effects in systems that are very sensitive to initial conditions and therefore effectively random than of the robust identity of the system itself, which is drawn to sets of patterns in a way that is no doubt highly irregular, but still characteristic and recognizable like the dazzling colours of a butterfly's wings.
Combining the two genealogies, the butterfly can symbolize a profound ambivalence of chaotic systems like the global weather system. On the one hand, it highlights their frailty and instability if even the flap of a single, frail butterfly can cause or prevent the occurrence of a storm. But the utter unpredictability resulting from this sensitivity, which leads to an inextricable entanglement with all elements, is only half the reason of why the weather behaves as if it had a mind of its own. Just as importantly, the weather behaves on the other hand in a manner that is not random but indeed of its own, always moving towards and along the intricately folded lines of its strange attractor, which characterizes the weather system and endows it with its own robust autonomous identity and idiosyncratic dynamic.
According to Lorenz, the strange attractor of the global weather system is 'simply the climate', 68 and it imposes an indeed remarkably stable identity on the weather. While explaining a difficulty in numerical weather simulations, he asks his readers to 'imagine an enormous creature from outer space that swoops down close to the earth, reaches out with a giant paddle, and stirs the atmosphere for a short while before disappearing' . 69 Such a drastic intervention will have violent effects, of course, but Lorenz maintains that they will quickly dissipate; the perturbation of the weather will then be 'hardly detectable and the weather will be back to normal, although the particular sequence of weather patterns will undoubtedly not be the one that would have developed without the disturbance.' 70 Clearly, Lorenz was not thinking of climate change here, but his argument does not deny its possibility and remains instructive in understanding how one can insist both on the weather's resilient autonomy and on its instability, unpredictability, and sensitivity: Even after an enormously powerful intervention from outer space, the weather system quickly returns to 'normal' (that is, to its strange attractor), but at the same time the particular sequence of weather patterns (where and when a tornado forms, for instance) remains susceptible to even the most minute interventions.
While much more could be said about chaos theory, its premises, and consequences, I will conclude by proposing that it can provide useful models with which to think some of the ambivalences of weathering that the enantiosemic uses of the word have revealed. In particular, by engaging with profound ambivalences of un/predictability, im/potency, and dis/entanglement it may help grasp the peculiar kind of subject evoked when one uses weather verbs and says 'it storms', 'it rains', 'it is hot', or 'it is cold' . What remains perhaps necessarily ambivalent is the question of whether the weather as subject anthropomorphizes the weather or deflates the notion of a subject. As noted at the outset, it may be preferable to err on the side of anthropomorphism and exclude an anthropocentric dualism in which the human (subject) constitutes itself in opposition to the (nonhuman) elements. Yet blurring all distinctions runs the risk of re-asserting through disavowal an emphatic autonomous subject with a vengeance, making the subject omnipotent by considering everything inseparably entangled, denying the relevance of scale, and allowing for decisionist agential cuts that determine what comes to matter and what is left to draw on. What I am suggesting instead is to trouble human exceptionalism through the peculiar subject of weathering that transpires when weathering is methodologically situated in both language and physics and grasped in its generative enantiosemy and ambivalence. | 16,900.8 | 2020-01-01T00:00:00.000 | [
"Physics",
"Philosophy",
"Environmental Science"
] |
Aerosol Optical Depth Retrieval for Sentinel-2 Based on Convolutional Neural Network Method
: Atmospheric aerosol significantly affects the climate environment and public health, and Aerosol Optical Depth (AOD) is a fundamental optical characteristic parameter of aerosols, so it is important to develop methods for obtaining AOD. In this work, a novel AOD retrieval algorithm based on a Convolutional Neural Network (CNN) method that could provide continuous and detailed aerosol distribution is proposed. The algorithm utilizes data from Sentinel-2 and Aerosol Robotic Network (AERONET) spanning from 2016 to 2022. The CNN AOD data are consistent with the AERONET measurements, with an R 2 of 0.95 and RMSE of 0.049 on the test dataset. CNN demonstrates superior performance in retrieving AOD compared with other algorithms. CNN retrieves AOD well on high reflectance surfaces, such as urban and bare soil, with RMSEs of 0.051 and 0.042, respectively. CNN efficiently retrieves AOD in different seasons, but it performs better in summer and winter than in spring and autumn. In addition, to study the relationship between image size and model retrieval performance, image datasets of 32 × 32, 64 × 64 and 128 × 128 pixels were created to train and test the CNN model. The results show that the 128-size CNN performs better because large images contain rich aerosol information.
Introduction
Atmospheric aerosols are solid, liquid, or solid-liquid mixed particles suspended in the atmosphere with an aerodynamic diameter of less than 100 µm. These particles are mainly distributed in the stratosphere and troposphere at the lower portion of the atmosphere. Aerosols exhibit a complex and variable composition resulting from a combination of natural emissions and human activities, including anthropogenic aerosols, such as sulfates and biomass-burning aerosols, as well as natural aerosols, such as dust and sea salt, mineral dust, and primary biological aerosol particles [1].
Aerosol, which plays a crucial role in the Earth's atmospheric system, has a significant influence on radiation balance, climate change, atmospheric environmental quality, and human health [2][3][4][5][6]. Aerosol particles could change Earth's radiation budget through scattering and absorbing incoming solar radiation [7]. Aerosols also have an indirect effect on climate by altering cloud microstructure [8,9]. In addition to the above-mentioned aerosol cloud microphysical effects, aerosols can also change the atmospheric thermodynamic characteristics through aerosol-radiation interaction, thereby altering the atmospheric circulation, thus affecting the occurrence of clouds and precipitation [10]. For example, black carbon aerosols can heat the atmosphere, evaporate cloud droplets to reduce cloud cover [11], create stable atmospheric stratification [12], weaken convection, and reduce rain showers. High concentrations of aerosol particles seriously affect human health [13][14][15]. A number of studies have shown that the incidence of cardiopulmonary disease is closely related to the content of particulate matter in polluted air [16][17][18], especially the inhalable
Aerosol Robotic Network (AERONET) Data
AERONET is a global aerosol optical property monitoring network established by the National Aeronautics and Space Administration. This network uses the CIMEL automatic sun-photometer as the basic observing instrument [40], which can observe aerosol characteristics at eight wavelengths from visible to near-infrared every 15 min, and the network consists of more than 1500 sites around the world [41]. These sites have been providing researchers with long-term, stable, and easily accessible aerosol data for decades. AERONET aerosol data are divided into three levels: Level 1.0 (unscreened), Level 1.5 (cloud-cleared and quality controls), and Level 2.0 (automatically cloud-cleared and quality-assured). The number of ground-based AODs available at level 2.0 is limited. Accordingly, level 1.5 data are selected for unified sampling to have sufficient samples. East Asia is the ideal research area for training and testing an AOD retrieval model because of the complex and variable composition of aerosols and abundant distribution of AERONET sites. V3 Level 1.5 data from 22 sites in East Asia from 2016 to 2022 were obtained in this study based on the climate, the land cover types around the sites, and the effectiveness and availability of observation data. Figure 1 shows the distribution of 22 sites used in this work, and Table 1 provides the site name, location, abbreviation, corresponding coordinates, and matched number with satellite images of each site.
Sentinel-2 Data
Sentinel-2 is an Earth observation mission in the European Space Agency's Copernicus program, and it consists of two polar-orbiting satellites: Sentinel-2A and Sentinel-2B. Every satellite carries a multi-spectral instrument to acquire 13 spectral bands from visible to shortwave infrared along a 290 km-wide orbital belt. The revisit period is 10 days for each satellite and 5 days for the combined system [42]. Sentinel-2 provides land surface images with spatial resolutions of 10, 20, and 60 m/pixel covering the global land, which is convenient for scholars to apply in remote sensing analysis fields, such as forest monitoring, land cover change monitoring, and natural disaster management. Sentinel-2 is superior for characterizing high spatial resolution and relatively high temporal resolution simultaneously, making it possible to obtain a continuous and detailed aerosol distribution map. The European Space Agency provides free Sentinel-2 Level-1C products, which
Sentinel-2 Data
Sentinel-2 is an Earth observation mission in the European Space Agency's Copernicus program, and it consists of two polar-orbiting satellites: Sentinel-2A and Sentinel-2B. Every satellite carries a multi-spectral instrument to acquire 13 spectral bands from visible to shortwave infrared along a 290 km-wide orbital belt. The revisit period is 10 days for each satellite and 5 days for the combined system [42]. Sentinel-2 provides land surface images with spatial resolutions of 10, 20, and 60 m/pixel covering the global land, which is convenient for scholars to apply in remote sensing analysis fields, such as forest monitoring, land cover change monitoring, and natural disaster management. Sentinel-2 is superior for characterizing high spatial resolution and relatively high temporal resolution simultaneously, making it possible to obtain a continuous and detailed aerosol distribution map. The European Space Agency provides free Sentinel-2 Level-1C products, which are the atmospheric apparent reflectance products that have been finely orthorectified and sub-pixel geometric corrected. In this study, Sentinel-2 L1C images from Bands 2, 4, 8, and 12 from 2016 to 2022 were selected ( Table 2).
MODIS and Himawari-8 Data
MODIS has provided scholars with different types of global aerosol products for decades, among which MODIS AOD standard products over land include AOD retrieved by the MAIAC algorithm, DT algorithm, DB algorithm, and combined Dark Target and Deep Blue (DTB) [34]. At present, the MODIS MAIAC aerosol products have been developed for the C6.0 version, and the DT, DB, and DTB aerosol products have been developed for the C6.1 version [28]. In 2015, the Japan Meteorological Agency began operating the geostationary satellite Himawari-8 and providing remote sensing data with a high temporal resolution (10 min) [43]. The MODIS data used in this study are the daily aerosol Table 3). The Himawari-8/AHI level 3 hourly Yonsei Aerosol Retrieval (YAER) product with a spatial resolution of 0.05 • /pixel was used in this work [44]. These aerosol products were selected to evaluate the reasonability of the spatial distribution of the CNN AOD.
Data Preprocessing
The dataset was constructed using multiple data preprocessing steps, including band interpolation, resampling, pixel screening, and spatiotemporal matching ( Figure 2). Considering that CNN AOD must be compared with MODIS aerosol products, MODIS AOD is the AOD at 550 nm, whereas AERONET AOD does not have such a wavelength. Consequently, band interpolation is required. The AERONET AODs at 500 and 675 nm are interpolated into AOD at 550 nm using the Ångström exponent.
where τ 500 denotes the AOD at 500 nm, τ 675 represents the AOD at 675 nm, α is the Ångström exponent at 500-675 nm, and τ 550 indicates the AOD at 550 nm obtained by interpolation.
Considering that Sentinel-2 L1C products are resampled at different ground sampling distances (10, 20, and 60 m/pixel) based on the native resolution of different spectral bands [45], the L1C products must be first resampled to unify the spatial resolution before aerosol retrieval. In this study, all satellite images were resampled to 10 m/pixel. Satellite images are pixel-screened to remove samples containing light, clouds, cloud shadows, and snow from the dataset to improve the quality of samples and minimize the influence of the aforementioned factors. The AERONET AOD data after band interpolation are selected to match with the screened satellite images. Assuming that the aerosol concentration does not change in a short period of time, the average sun-photometer AOD value within ±30 min of the satellite overpass is selected to match with the satellite image. Assuming that the aerosol properties do not change within an area of 1.28 km × 1.28 km, the Sentinel-2 image within the range of 128 × 128 pixels is cropped as a satellite image sample with the AERONET site in the center. The satellite image and ground-based AOD data in the same spatiotemporal are paired together to provide samples for the model.
When training deep learning models, data augmentation methods, such as rotation, scaling, and translation, are frequently used to produce more samples to improve the generalization ability and robustness of the model [46]. The remote sensing images are processed by a horizontal translation of 0.1 to expand the dataset, enabling the model to better learn features from these images and avoid model overfitting. In this study, 2575 pairs of samples were successfully matched from the AERONET AOD data and Sentinel-2 images. Before training, all image samples are resized to 128 × 128 pixels. A total of 2575 pairs of samples are randomly assigned to the training dataset (2060 pairs) and the test dataset (515 pairs) in an 8:2 ratio.
CNN Model
CNN, which is a deep learning algorithm, consists of a series of convolution and pooling layers that extract relevant features from the input image, followed by one or more fully connected layers that use these features to make a prediction. It has the characteristics of sparse connections, parameter sharing, and translation invariance, greatly minimizing the number of network parameters that need to be optimized, improving the training speed, and facilitating the extraction of partial features in images. CNN is able to capture spatial features and patterns in images using a hierarchical architecture of layers that perform convolution operations and extract features at different levels of abstraction [47]. So, this algorithm has remarkable advantages for image analysis and recognition. The CNN model proposed in this work consists of 10 convolution layers, 4 max-pooling layers, 1 Global max-pooling layer, and 1 fully connected layer, as shown in Figure 3. Every convolutional block is followed by a max-pooling layer to minimize the number of parameters and compress image features to reduce computation and memory consumption. The last max-pooling layer is followed by a Global max-pooling layer, which can flatten the multidimensional image to obtain a feature vector with a size of 512. In addition, dropout regularization is used to randomly turn off some neurons with a certain probability during each iteration of training of the neural network [48]. Accordingly, each neuron does not necessarily participate in network training every time, and the updating of network weights no longer depends on the joint action of implicit nodes with fixed
CNN Model
CNN, which is a deep learning algorithm, consists of a series of convolution and pooling layers that extract relevant features from the input image, followed by one or more fully connected layers that use these features to make a prediction. It has the characteristics of sparse connections, parameter sharing, and translation invariance, greatly minimizing the number of network parameters that need to be optimized, improving the training speed, and facilitating the extraction of partial features in images. CNN is able to capture spatial features and patterns in images using a hierarchical architecture of layers that perform convolution operations and extract features at different levels of abstraction [47]. So, this algorithm has remarkable advantages for image analysis and recognition. The CNN model proposed in this work consists of 10 convolution layers, 4 max-pooling layers, 1 Global maxpooling layer, and 1 fully connected layer, as shown in Figure 3. Every convolutional block is followed by a max-pooling layer to minimize the number of parameters and compress image features to reduce computation and memory consumption. The last max-pooling layer is followed by a Global max-pooling layer, which can flatten the multidimensional image to obtain a feature vector with a size of 512. In addition, dropout regularization is used to randomly turn off some neurons with a certain probability during each iteration of training of the neural network [48]. Accordingly, each neuron does not necessarily participate in network training every time, and the updating of network weights no longer depends on the joint action of implicit nodes with fixed relationships. This approach prevents some features from being effective only under other specific features, reduces the dependence of model on some partial features, and enhances the generalization ability of the model. All convolutional layers use small-size 3 × 3 convolutional filters with size 1 stride and select the "same" padding to maintain the spatial resolution of the feature map. All max-pooling operations are performed over a window of 2 × 2 pixels with size 2 stride and the "same" padding to minimize the spatial resolution of the feature map by half. the loss. Accordingly, the predicted value output by the model will be closer to the true value. Mean Absolute Error (MAE) was used as the loss function of the model in this study. The nonlinear activation function ReLU is selected as the activation function to minimize the interdependence between parameters. During the back propagation, the optimizer can guide the model parameters to update in the correct direction. Consequently, the updated parameters maintain the loss value close to the global minimum. The Adam optimization algorithm is used for network training, and the learning rate is set to 0.001 to minimize the MAE between retrieved AOD and true AOD. After the CNN is trained, the model with the smallest MAE between CNN AOD and AERONET AOD is selected as the final CNN model.
Evaluation Methods
Spearman R, Pearson R, R 2 , Root Mean Square Error (RMSE), MAE, Within the MSI Expected Error (Within EE), Above EE, and Below EE were used as evaluation indicators in this study. Pearson R is sensitive to outliers and suitable for evaluating linear relationships. On the other hand, Spearman R is not affected by outliers and is suitable for evaluating nonlinear relationships. The model performance can be effectively evaluated using these two correlation coefficients as evaluation indexes. The loss function, activation function, and optimizer should be configured before training. After the training data of every iteration is inputted into the model, the model will output the predicted value through forward propagation, and the loss function will calculate the difference between the predicted and the true values. After the loss value is obtained, the model will update each parameter through back propagation to minimize the loss. Accordingly, the predicted value output by the model will be closer to the true value. Mean Absolute Error (MAE) was used as the loss function of the model in this study. The nonlinear activation function ReLU is selected as the activation function to minimize the interdependence between parameters. During the back propagation, the optimizer can guide the model parameters to update in the correct direction. Consequently, the updated parameters maintain the loss value close to the global minimum. The Adam optimization algorithm is used for network training, and the learning rate is set to 0.001 to minimize the MAE between retrieved AOD and true AOD. After the CNN is trained, the model with the smallest MAE between CNN AOD and AERONET AOD is selected as the final CNN model.
Evaluation Methods
Spearman R, Pearson R, R 2 , Root Mean Square Error (RMSE), MAE, Within the MSI Expected Error (Within EE), Above EE, and Below EE were used as evaluation indicators in this study. Pearson R is sensitive to outliers and suitable for evaluating linear relationships. On the other hand, Spearman R is not affected by outliers and is suitable for evaluating nonlinear relationships. The model performance can be effectively evaluated using these two correlation coefficients as evaluation indexes.
Atmosphere 2023, 14, 1400 Random Forest (RF) and VGG16 algorithms are selected for comparison and verification with the CNN model to verify the rationality of the CNN model. RF is an ensemble learning algorithm [49] that can evaluate the importance of every feature in classification problems, judge the interaction between different features, and capture the complex relationship between input feature vectors and output values. VGG is a classic CNN structure that is widely used in computer vision tasks, such as image classification, object detection, and semantic segmentation. VGG16 is a network structure in VGG [50].
Overall Validation of the CNN Model
AERONET AODs were used to compare and analyze the AOD retrieved by the CNN model on the training and test datasets. In the training dataset, the retrieved AOD is in good agreement with AERONET AOD, with an R 2 of 0.99, RMSE of 0.021, and Within EE of 99%, as shown in Figure 4. Only 1% of the data is higher than EE, indicating a very slight overestimation of the model in the training dataset. The result of CNN performance on the test dataset is similar to that on the training dataset. In the test dataset, R 2 is 0.95, RMSE is 0.049, and Within EE is 95%, indicating that CNN can provide accurate AOD concentration and distribution at a relatively fine temporal resolution (5 days). Approximately 3% of AODs are higher than EE, and 2% of AODs are lower than EE, indicating a slight deviation in the CNN AOD. The AOD retrieval performance of the CNN model on the test dataset was subdivided by season and land cover type. CNN performs better in summer and winter than in spring and autumn (Table 5). From the perspective of the various validation indicators, In the satellite images, the top-of-atmosphere signals are composed of coupled contributions from the surface and atmosphere, hence, the signal-to-noise-ratio of the high-AOD aerosols is larger than that of the low-AOD aerosols [35]. In the test dataset, 52% of the AERONET AODs are in the range of 0 to 0.16, and 48% are in the range of 0.16 to 1.6. The test dataset was divided into two parts to examine the accuracy of the CNN model in Table 4 illustrates that MAE and RMSE in the low AOD are lower than those in the high AOD. On the other hand, the Within EE in the low AOD is similar to that in the high AOD. This finding indicates that CNN can better retrieve AOD in areas with good air quality. In the low AOD, 6% of AODs are higher than EE and no data is lower than EE. In the high AOD, 1% of AODs are higher than EE and 4% of AODs are lower than EE. The result shows that the CNN model slightly overestimates the low values and underestimates the high values. The AOD retrieval performance of the CNN model on the test dataset was subdivided by season and land cover type. CNN performs better in summer and winter than in spring and autumn (Table 5). From the perspective of the various validation indicators, the performance of CNN in winter is the best, with an R 2 , MAE, RMSE, and Within EE of 0.97, 0.024, 0.037, and 97%, respectively. The excellent CNN performance in winter may be related to the large amount of winter data. The performance of CNN in summer is the second best, with an R 2 of 0.96 and RMSE of 0.045. On the other hand, the performance of CNN in autumn is similar to that in summer, with an R 2 of 0.94 and RMSE of 0.046. The CNN performance in spring is poorer compared with those in the other three seasons, with an R 2 of 0.93 and RMSE of 0.062. CNN efficiently performs AOD retrieval on the different land cover types (Table 6). CNN exhibits excellent retrieval accuracy in the vegetation surface, with an R 2 of 0.92 and RMSE of 0.044. When the surface reflectance is high, the satellite sensors can obtain less information about aerosols, making it difficult to retrieve aerosols on high-reflectance surfaces [51]. However, CNN still shows great retrieval performance on land cover types with high surface reflectance, such as urban and bare soil, with an R 2 of 0.96 and 0.90 and RMSE of 0.051 and 0.042, respectively.
Retrieval Performance of the CNN Model at Different Scales
The AOD retrieval performance of the CNN model on the test dataset was subdivided into two scale types: regional scale and ground-based AERONET site scale, to analyze CNN performance in different regions and sites. The retrieval performance of CNN in North China, Taiwan, South Korea, and Japan is shown in Figure 5. There is no analysis of Tibet and Hong Kong here because of the small amount of data in these regions. CNN performed best in South Korea, with an RMSE of 0.037, R 2 of 0.92, and Within EE of 98%. Moreover, 2% of AODs are lower than EE, indicating a slight underestimation of CNN in South Korea. In North China, CNN could accurately retrieve AOD with an R 2 of 0.97, RMSE of 0.047, and Within EE of 96%. Furthermore, 2% of AODs are higher than EE, whereas 2% are lower than EE, indicating a slight deviation of CNN AOD in North China. The more data in the region, the more regional features the model could capture. The largest amount of data can be found in North China. However, the retrieval performance of CNN in North China is not as good as that in South Korea, which may be because the aerosol concentration in North China is complex and changeable, and it is difficult for CNN to learn the aerosol characteristics in this region. CNN did not perform well in Taiwan, with its RMSE being the highest of the four regions. CNN performed the worst in Japan, with an R 2 of 0.78, RMSE of 0.045, and Within EE of 91%, which may be related to the smaller amount of data obtained in Japan. The annual average AOD in Japan is the lowest, and the CNN model slightly overestimates the low AODs with an Above EE of 6% (Table 4), so the CNN tends to overestimate the AOD in Japan, which has an Above EE of 8% and is significantly higher than those in North China and South Korea.
Atmosphere 2023, 14, x FOR PEER REVIEW 11 of 17 amount of data obtained at TW03 is small, with only 55 pairs of samples. Accordingly, the accuracy of CNN AOD at TW03 is low. CNN exhibits a slight AOD deviation at most sites. Nonetheless, the CNN AOD at HK01, HK02, KR4, and JP3 reaches 100% Within EE, and no AOD is higher or lower than EE at all. CNN retrieved AOD better at sites with more data, indicating that CNN can obtain more features at sites with more data. Expanding the dataset may help the model to better extract image features related to aerosols, and a large dataset could be made for model training in the future. The CNN performance at each site is shown in Figure 6. Given the limited amount of data at TB01 and KR1, only one TB01 data and two KR1 data were divided into the test dataset. Consequently, the retrieval performance at these two sites is not shown. The AOD retrieval accuracy at NC02 is the highest, with an R 2 , RMSE, and Within EE of 0.98, 0.027, and 99%, respectively. This result may be due to the largest amount of data being at NC02, with up to 311 pairs of samples, from which CNN could capture more accurate image features. Moreover, the land cover type at NC02 is simple and mainly urban. In addition, CNN performed well at TW01, with its RMSE being the lowest among all sites (0.016). The land cover type at TW01 is simple, with the vast majority of the area covered by vegetation, and CNN can obtain more aerosol information in an area with low surface reflectance. The AOD retrieval accuracy at JP4 is the lowest, with an R 2 of 0.73, RMSE of 0.095, and Within EE of 67%. CNN overestimated the AOD at JP4 with an Above EE of 33%. The low accuracy of CNN AOD at JP4 is due to two reasons. First, high AOD more frequently occurs in China and South Korea than in Japan [52], and the annual average AOD in Japan is the lowest, so CNN tends to overestimate the AOD in Japan. Second, the data at JP4 are few (only 52 pairs of samples). The deep learning model requires a large number of samples, and CNN would have difficulty exploring the characteristics of the site using only a few samples. In addition, CNN did not perform well at TW03, with its RMSE being the highest of all sites (0.122). TW03 is characterized by a relatively complex land cover type, primarily dominated by urban areas with some vegetation coverage. Moreover, the amount of data obtained at TW03 is small, with only 55 pairs of samples. Accordingly, the accuracy of CNN AOD at TW03 is low. CNN exhibits a slight AOD deviation at most sites. Nonetheless, the CNN AOD at HK01, HK02, KR4, and JP3 reaches 100% Within EE, and no AOD is higher or lower than EE at all. CNN retrieved AOD better at sites with more data, indicating that CNN can obtain more features at sites with more data. Expanding the dataset may help the model to better extract image features related to aerosols, and a large dataset could be made for model training in the future.
Comparison with Other Models
The MODIS and Himawari-8 aerosol products were selected to evaluate the rationality of the spatial distribution of CNN AOD. Figure 7a shows a Sentinel-2 satellite image over Chiba and Ibaraki in Japan on 12 February 2019, with no snow and clouds in most areas, making it suitable for AOD retrieval. Figure 7c-g shows MODIS MAIAC, DT, DB, DTB, and Himawari-8 aerosol products over Japan, with spatial resolutions of 1 km/pixel, 10 km/pixel, 10 km/pixel, 10 km/pixel, and 0.05 • /pixel, respectively. Figure 7b shows the spatial distribution of AOD retrieved by CNN. CNN AODs are calculated with a step size of 64 pixels and an input image size of 128 × 128 to facilitate display and comparison, resulting in a spatial resolution of 640 m/pixel. Figure 7a depicts some cloud and snow cover in the lower left corner and middle upper areas of the image. Incorrect AODs are retrieved by CNN due to the interference of clouds and snow. Except for the cloud and snow coverage area, the spatial distribution of CNN AOD and MODIS AOD is similar, and the overall AOD concentration of the image coverage area is low. Some coastal areas show relatively high AODs due to dense human habitation and intensive industrial agglomeration. On the other hand, inland and southern mountain areas show relatively low AODs. The Himawari-8 aerosol product shows an overall higher AOD, which is not consistent with MODIS AOD and CNN AOD. The MODIS AOD product is more reliable than the Himawari-8 AOD product [34]. Specifically, the CNN AOD that is more similar to MODIS AOD can reasonably reflect the AOD concentration. The distribution maps of the CNN AOD and Himawari-8 AOD cover the entire land. The data coverage rate of the MODIS aerosol products is low, and many no-value regions exist. The CNN AOD and Himawari-8 AOD can provide continuous AOD spatial distribution. Furthermore, the high spatial resolution map of the CNN AOD shows the spatial details that cannot be reflected in the rough spatial resolution of 1 km/pixel, 10 km/pixel, and 0.05 • /pixel. The AOD retrieved by CNN is closer to the ground-based AOD, with an RMSE of 0.049. In contrast, the RMSE values of MODIS MAIAC, DT, DB, and Himawari-8 AOD are 0.15, 0.22, 0.17, and 0.14, respectively, which are less accurate than CNN.
Comparison with Other Models
The MODIS and Himawari-8 aerosol products were selected to evaluate the rationality of the spatial distribution of CNN AOD. Figure 7a shows a Sentinel-2 satellite image Himawari-8 AOD can provide continuous AOD spatial distribution. Furthermore, the high spatial resolution map of the CNN AOD shows the spatial details that cannot be reflected in the rough spatial resolution of 1 km/pixel, 10 km/pixel, and 0.05 °/pixel. The AOD retrieved by CNN is closer to the ground-based AOD, with an RMSE of 0.049. In contrast, the RMSE values of MODIS MAIAC, DT, DB, and Himawari-8 AOD are 0.15, 0.22, 0.17, and 0.14, respectively, which are less accurate than CNN. The CNN model was compared with RF and VGG16 to further evaluate the AOD retrieval performance of CNN. The RF and VGG16 algorithms were trained and tested using the dataset produced in this study. The output layer of VGG16 was changed to a The CNN model was compared with RF and VGG16 to further evaluate the AOD retrieval performance of CNN. The RF and VGG16 algorithms were trained and tested using the dataset produced in this study. The output layer of VGG16 was changed to a fully connected layer with size 1 because it is a classification model that eventually divides into 1000 categories. The AOD retrieval performance of each model is shown in Table 7. The results of MAIAC, DT, DB, and YAER are from a study by Choi et al. [53], in which the MODIS and Himawari-8 aerosol products in East Asia are validated. CNN shows outstanding advantages in AOD retrieval, and all verification indicators are the best values, with an RMSE of 0.049, and significantly lower than those of other models. The R 2 of CNN is 0.95, which is higher than those of other models. RF performs well, with an R 2 of 0.86, which is higher than the 0.70 of VGG16. The RMSE of RF is 0.085, which is lower than the 0.123 of VGG16. The results show that CNN demonstrates significant advancements in addressing AOD retrieval problems compared with most previous algorithms.
Retrieval Performance as a Function of Satellite Image Size
Image datasets of 32 × 32, 64 × 64, and 128 × 128 pixels were constructed to explore the relationship between image size and AOD retrieval performance. The model was retrained with datasets of different sizes, and the hyperparameters of CNN were adjusted. The smaller the size, the less information the satellite image contains. Accordingly, the image feature vector of the fully connected layer in the CNN was reduced to 80 when the model was trained with datasets of 64 and 32 sizes. The AOD retrieval performance on the dataset with different sizes is shown in Table 8. The 128-size model works best, with an R 2 of 0.95, which is 0.03 and 0.02 higher than those of the 32-size model and 64-size model, respectively. The RMSE of the 128-size model is 0.049, which is 0.016 and 0.011 lower than those of the 32-size model and 64-size model, respectively. The 32-size model performed worst, with the lowest R 2 , highest MAE, and highest RMSE among the three models, and its overestimation and underestimation are the most severe. The better results of the model trained with a larger image dataset indicate that the 128-size image contains richer aerosol information.
Conclusions
In this study, a novel AOD-retrieval model based on CNN was developed, which can directly retrieve AOD from Sentinel-2 satellite images, improving the stability and spatiotemporal adaptability of aerosol retrieval. The CNN model consists of 10 convolutional layers, 4 max-pooling layers, 1 Global max-pooling layer, and 1 fully connected layer. The image features related to aerosol load in Sentinel-2 images of four bands were extracted using the excellent image feature extraction function of CNN. The correlation between the AOD and the remote sensing images was modeled to achieve AOD retrieval. Taking East Asia as the research area, the data of 22 AERONET sites from 2016 to 2022 were obtained, and the satellite image and ground-based AOD data in the same spatiotemporal were paired together to make 2575 pairs of samples. The samples were enhanced by a horizontal shift of 0.1. The model was trained and tested by selecting data in a ratio of 8:2 from samples. The results are as follows: (1) The proposed model can accurately retrieve AOD, with an R 2 of 0.95, RMSE of 0.049, and Within EE of 95% on the test dataset. The AOD-retrieval accuracy of CNN is higher compared with those of the DT, DB, DTB, MAIAC, YAER, RF, and VGG16 algorithms. In addition, CNN could provide continuous and detailed aerosol distribution to fill the observation gap in existing ground-based monitoring networks. (2) CNN efficiently performs AOD retrieval on different land cover types: vegetation surface, urban, and bare soil. When the surface reflectance is high, the satellite sensors can obtain less information about aerosols, making it difficult to retrieve aerosols on a high-reflectance surface. However, CNN still shows great AOD retrieval potential on surfaces with high surface reflectance, such as urban and bare soil, with an R 2 of 0.96 and 0.90 and RMSE of 0.051 and 0.042, respectively. (3) CNN performs better in summer and winter than in spring and autumn. The performance of CNN in winter is the best, with an R 2 of 0.97 and RMSE of 0.037. The performance of CNN in summer is the second best. The CNN performance in spring is poorer compared with those in the other three seasons, with an R 2 of 0.93 and RMSE of 0.062. (4) To investigate the relationship between image size and model retrieval performance, datasets of 32, 64, and 128 sizes were created to train and test the CNN. The 128-size CNN performed better because of the rich AOD information in the 128-size image. | 7,905.8 | 2023-09-05T00:00:00.000 | [
"Environmental Science",
"Computer Science"
] |
Dose calculation accuracy of lung planning with a commercial IMRT treatment planning system
The dose calculation accuracy of a commercial pencil beam IMRT planning system is evaluated by comparison with Monte Carlo calculations and measurements in an anthropomorphic phantom. The target volume is in the right lung and mediastinum and thus significant tissue inhomogeneities are present. The Monte Carlo code is an adaptation of the msnp code and the measurements were made with TLD and film. Both the Monte Carlo code and the measurements show very good agreement with the treatment planning system except in regions where the dose is high and the electron density is low. In these regions the commercial system shows doses up to 10% higher than Monte Carlo and film. The average calculated dose for the CTV is 5% higher with the commercial system as compared to Monte Carlo. PACS number(s): 87.53.‐j, 87.66.‐a
I. INTRODUCTION
The purpose of this study is to evaluate the accuracy of doses calculated by an IMRT planning system in the presence of inhomogeneities and under clinically relevant conditions. We have made comparisons in the thorax of an anthropomorphic phantom between the dose distribution computed by an IMRT treatment planning system ͑Corvus v4.6, NOMOS Corp., Cranberry Township, PA͒, measured doses and Monte Carlo ͑an adaptation of MCNP͒ calculated doses. The measurements were made with film and thermoluminescent dosimeters ͑TLD͒. A comparison of this study with previously published work of a similar nature is discussed in Sec. VII.
Most clinical physicists carry out patient specific IMRT validation measurements in slab phantoms which are homogeneous. It is known that accurate dose computation in the presence of large tissue inhomogeneities is challenging and that many algorithms exhibit shortcomings in this domain. 1 Furthermore, accurate dose calculation may be more crucial for inversely planned IMRT than for conventional treatment planning because plan optimization is based on the calculated dose matrix. 2 For these reasons it is important to assess the dose calculation accuracy of IMRT treatment planning systems under conditions of realistic geometry and composition.
II. METHODS AND MATERIALS
We have used the anthropomorphic phantom with the trade name RANDO ͑The Phantom Laboratory, Salem, NY͒. This phantom was fabricated using a tissue equivalent resin ( ϭ0.985 g/cm 3 ) molded around a natural skeleton. The phantom has air cavities ͑pharynges, larynx, trachea, stem bronchi, etc.͒ which were made from impressions taken from a cadaver. RANDO has realistic lungs with a density of 0.32 g/cm 3 . RANDO is sectioned axially into 2.5 cm thick slabs ͑see Figs. 1 and 2͒. Film can be placed between slabs for dosimetry studies. Within each slab there are plugs arranged in a grid with a spacing of 1.5 cm between adjacent plugs ͑see Fig. 1͒. These plugs may be removed and replaced by TLDs.
Three fiducial markers ͑2 mm diameter lead BBs͒ were placed on the phantom prior to a CT scan, one each on the right lateral, left lateral, and anterior surfaces. Paper spacers were placed between the slabs, where it was anticipated that film would be positioned, in order to preserve geometric spacing between slabs. The relationship between CT number and relative electron density was calibrated with a Gammex rmi 467 CT electron density phantom ͑Gammex rmi, Middleton, WI͒. The slice thickness and table feed were 4 mm. The scan extended from a point superior to the apex of the lungs to inferior of the diaphragm. Each CT slice contains 512 by 512 pixels. Each pixel is about 1 mm in size.
The IMRT treatment planning system is Corvus ͑v4.6͒ produced by the NOMOS corporation. The dose calculation is based on a finite size pencil beam algorithm with a beamlet size of 1 cm by 1 cm. The tissue inhomogeneity correction is based on a path length correction for each pencil beam. The linac used for delivery is a Varian Clinac 2300 C/D with 52 leaves ͑26 pairs͒ ͑Varian Medical Systems, Inc., Palo Alto, CA͒. Leaf width is 1.0 cm projected to isocenter. The beam energy is 6 MV.
The target was contoured by a physician in Corvus. It is intended to be representative of a typical lung tumor target volume ͑see Figs. 1 and 2͒. The target occupies a portion of the mediastinum and right lung. The volume of the clinical target volume ͑CTV͒ is 148 cm 3 . The portion of the target in the lung has electron density values characteristic of lung tissue. A real tumor may have higher electron density values. This is a caveat that must be considered when interpreting our results. The organs at risk that were contoured are the left lung ͑2300 cm 3 ͒, right lung ͑2500 cm 3 ͒, and the spinal cord. Overlapping structures are not permitted in Corvus and therefore the contoured right lung does not include the portion of the target that is in the right lung. Margins for the expansion of the CTV to the PTV are 11 mm in both the anterior and posterior direction, 14 mm in both the right and left direction, and 18 mm in both the superior and inferior direction. 3 FIG. 1. ͑Color͒ Shows an axial cross section CT image of the anthropomorphic phantom used in this study. The grid of plugs that can be removed and replaced with TLDs is evident. The plugs are approximately 1.5 cm apart. The region colored red is the CTV. The dose distribution for the Corvus plan is shown with full correction for inhomogeneities. The Monte Carlo software is an adaptation of the MCNP 4B code. 4 Validation measurements for this adaptation were performed in both homogeneous and heterogeneous phantoms and are described in detail by He. 5 Dose was calculated in 3 mmϫ3 mmϫ3 mm voxels with statistical uncertainties less than 1%.
The goal dose for the target is 180 cGy. We were restricted to axial beams because the MCNP calculations are only possible for axial beams due to current software limitations. For these reasons the plan is not the best possible plan, and perhaps not even a good plan, but it is not unrealistic. Our focus here is not on the quality of the plan, but on the accuracy of the dose calculation. A total of five beams were used: an anterior beam ͑gantry angle 0°͒, two anterior oblique beams ͑25°and 341°͒, and two posterior oblique beams ͑127°and 225°͒. All gantry angles are based on the International Electrotechnical Commission ͑IEC͒ scale. The optimization was performed with inhomogeneity corrections turned on. An axial view of the dose distribution is shown in Fig. 1 and a coronal view in Fig. 2. The Corvus voxel size for dose calculation is 1 mmϫ1 mmϫ1 mm.
Ready pack 10Љϫ12Љ Kodak EDR2 film ͑Kodak, Rochester, NY͒ was placed in between slabs 12 and 13, 13 and 14, and lastly 14 and 15 ͑see Fig. 2͒. Three widely spaced pinholes were made in each film at known plug positions for later image registration. The phantom was set up in the treatment room using the fiducial markers that were placed prior to the CT scan. Before delivering the irradiation, anterior and lateral orthogonal films were exposed. These films were compared to DRRs generated by Corvus to ensure proper positioning. Small adjustments in the position of the phantom were made as a result of this procedure. A film calibration curve was constructed using an MLC step wedge previously calibrated with an ion chamber. Eleven data points were used to define the H&D curve with doses between 0 and 350 cGy. All films were taken from the same batch and processed simultaneously. The films were scanned with a 16 bit CCD film scanner ͑Vidar VXR-16DP, Vidar Systems Corp, Herndon, VA͒. The scanning parameters were 357 m resolution and 3 by 3 median filtering. This provides resolution of 1 mm or better. The isodose lines from the film were superimposed on the Corvus treatment planning images ͑see Sec. IV͒.
van Dyk has proposed criteria for the accuracy of dose calculations for a composite anthropomorphic phantom. 6 The criterion used in a specific region depends on whether the dose is high or low there and on whether the dose gradient is high or low. The criteria are based on one standard deviation. This means that 68% of comparisons should lie within the stated tolerance. In a high dose region where the dose gradient is low, the dose accuracy at corresponding locations should be 4%. In high dose gradient regions ͑Ͼ30%/cm͒ the distance to agreement criterion is used. The distance to agreement is the distance between a point with a specific dose value and the nearest point in the comparison dose distribution that has the same dose value. In such regions the distance to agreement should be no more than 3 mm. In low dose regions ͑Ͻ7% of the normalization dose͒ where the dose gradient is also low, the dose agreement should be within 3%.
It is difficult to make experimental dose measurements in an anthropomorphic phantom with the accuracy necessary to discern these distinctions. The error in the registration of the film to the treatment planning dose distribution would need to be Ͻ3 mm to clearly demonstrate a distance to agreement of 3 mm. This is difficult to accomplish. In high dose, low gradient regions, dose measurements would need to be accurate to considerably better than 4% to discern discrepancies on that level. This is again a difficult proposition in that the accuracy of TLD measurements is on the order of 3%. 7 Dose distributions on Corvus can only be displayed for axial, sagittal, and coronal planes. The film plane was not precisely parallel to the CT axial plane. There was an anterior to posterior tilt of approximately 1°. Three pin pricks were made in each film on the left, right, and anterior for later registration with treatment planning images. It is estimated that the film registration accuracy is approximately 3 mm. Dosimetric accuracy of 2%-3% seems possible with EDR2 film. 8,9 The irradiation setup uncertainty, even with a rigid stationary phantom, may be on the order of 1-2 mm despite the use of localization films. On the slices that most closely correspond to the film location, adjacent 10% Corvus decrement isodose lines sometimes touch and even cross one another. As an example, in Fig. 5͑b͒ the Corvus 162 cGy line actually crosses the Corvus 144 cGy line in the right posterior lung where the gradient is very steep. In view of all of these considerations it is evident that the van Dyk criteria are difficult to use in practice because the tolerance level is on the same order, or perhaps even less, than the measurement uncertainties. These considerations do not apply, however, to the comparison between Corvus and MCNP. For this comparison the dose distributions can be very accurately registered.
III. COMPARISON OF TREATMENT PLAN STATISTICS
Three treatment plans have been generated, and these are compared in this section. Two of these are Corvus plans and one is a MCNP plan. In the ''corrected'' Corvus plan the dose optimization and the final calculated dose are based on dose computed with corrections for inhomogeneities ͑see Figs. 1 and 2͒. The normalization for the corrected plan has been chosen so that 5% of the CTV volume is below the goal dose of 180 cGy. The fluence map for this plan was taken and used to recalculate the dose in the absence of corrections. This is the Corvus ''uncorrected plan.'' The MU for each port was identical for each Corvus plan. The fluence delivery is precisely the same in each case, the only difference is whether corrections are made in the final dose calculation. This will show the effect of the corrections on the dose distribution.
For the MCNP plan, the intensity maps were obtained from the leaf sequencing files generated by Corvus. Corvus ͑v4.6͒ does not recompute the dose after leaf sequencing and this may lead to some small differences between the MCNP computed dose and the Corvus ͑preleaf sequencing͒ computed dose. The target and the normal structure contours were transferred via RTOG export from Corvus to the MCNP code.
The three plans are summarized and compared in Tables I and II and in Fig. 3. Table I uncorrected plan, 60% of the target volume is below the goal dose of 180 cGy, the minimum dose is about 110 cGy and the maximum dose is 209 cGy. The mean dose is approximately 15% lower for the uncorrected plan. The differences are due to the extra attenuation ͑in the uncorrected plan͒ associated with the assumption of unit relative electron density throughout the volume of the lungs. The MCNP doses in the target are generally lower than computed by the corrected Corvus plan; the mean dose is approximately 5% lower, although the maximum dose is about the same. The percentage of the volume below the goal dose is 20% for MCNP in comparison to only 5% for Corvus. Table II lists dose volume statistics for the normal structures for a 34 fraction treatment which delivers a total dose of 6120 cGy to the CTV. There is little difference between MCNP and Corvus ͑corrected͒. Figure 3 shows the cumulative dose volume histogram ͑DVH͒ for all anatomical structures for the Corvus corrected plan and for the MCNP plan. Also shown in the same figure is the Corvus uncorrected CTV DVH. For the target, the MCNP DVH is lower by a ͑possibly͒ significant amount. Corvus shows a somewhat higher volume of right lung at the highest doses, just as for the CTV.
IV. COMPARISON OF ISODOSE CURVES
The dose distribution on an axial slice between slabs 12 and 13 ͑see Fig. 2͒ calculated by Corvus is superimposed on the MCNP dose distribution in Fig. 4͑a͒. The agreement is excellent except in the medial aspect of the right lung, where the dose is high and the electron density is low. The MCNP dose is as much as 10% lower than the Corvus dose. The dose distribution measured on film between slabs 12 and 13 is displayed superimposed on the Corvus dose distribution in Fig. 4͑b͒. The circles in the anterior mediastinum and in the left lung are from the pin holes in the film. The measured dose distribution is therefore unreliable in the vicinity of these holes. The dose distribution on the film extends above the anterior surface of the phantom because the film projected above this surface. There are discrepancies in the same region as in Fig. 4͑a͒. In this region the dose to the film is lower than the Corvus doses just as for the MCNP doses. There are locations where the corresponding isodose lines are more than 3 mm apart.
The interface between slabs 13 and 14 ͑see Fig. 2͒ is in the middle of the high dose region. The Corvus dose distribution is superimposed on the MCNP dose distribution in Fig. 5͑a͒. The low dose lines correspond with one another fairly closely. In the right lung and mediastinum and generally throughout the region of low electron density which is occupied by the target, there are some significant discrepancies. In this region the MCNP dose is lower than the Corvus dose by 10% or more. The same features are seen in Fig. 5͑b͒, which shows the film dose measurements superimposed on the Corvus dose distribution. Figure 6 shows a profile in the left/right direction through the crosshair shown in Fig. 5͑a͒.
The interface between slabs 14 and 15 is near the inferior border of the high dose region ͑see Fig. 2͒. The MCNP dose distribution is superimposed on the Corvus dose distribution in Fig. 7͑a͒ FIG. 4. ͑Color͒ ͑a͒ Superposition of Corvus and MCNP dose calculation at the location of the interface between slabs 12 and 13 of RANDO. The thick lines are MCNP and the thin lines are Corvus. The agreement is excellent, so much so that in many locations it is not possible to see the superposition because isodose curves lie on top of one another. The exception is in the medial aspect of the right lung. MCNP lines appear to be pushed in compared to Corvus. This indicates that Corvus doses tend to be higher in this region. ͑b͒ Isodose lines from film ͑thick lines, sometimes dashed͒ superimposed on the dose distribution calculated by Corvus for slabs 12 and 13. The 54 cGy film isodose curve extends above the anterior surface of the phantom because the film protruded above the surface. The circles are the locations of the pin pricks in the film. Registration accuracy between the film and the Corvus image is expected to be on the order of 3 mm. The agreement between the film and Corvus is good except in the posterior mediastinum and the medial aspect of the right lung. The film dose is somewhat higher in the posterior mediastinum and lower in the right lung ͑as is MCNP͒. and the film dose is superimposed on the Corvus dose in Fig. 7͑b͒. The agreement is seen to be fairly good. The doses tend to be lower than the Corvus doses.
On the whole, the film and the MCNP doses tend to be lower than the Corvus dose. This is consistent with the results presented in Sec. III which suggest that the mean Corvus dose delivered to the CTV may be 5% high. The regions in which the discrepancies are greatest are the locations where the dose is highest and the electron density is lowest. This is in the medial aspect of the right lung ͓see Figs. 5͑a͒ and 5͑b͔͒. FIG. 5. ͑Color͒ ͑a͒ Corvus isodose lines superimposed on the MCNP isodose lines for slabs 13 and 14. This is a plane near the center of the dose distribution. The thin lines represent the Corvus dose distribution. This plane shows some of the largest discrepancies between Corvus and MCNP which occur in the high dose low electron density region ͑medial aspect of the right lung͒. The Corvus doses are as much as 10% higher than MCNP. A left/right dose profile through the cross hair is shown in Fig. 6. ͑b͒ Corvus isodose lines superimposed on isodose lines measured with film. The thin isodose lines were computed by Corvus; the thick lines represent film measurements. On this plane we see the largest differences between Corvus and the measured doses.
VI. TLD MEASUREMENTS
Twenty four TLD powder capsules ͑LiF TLD-100, 30 mg͒ from the same batch were used inside RANDO slabs 13 and 14 ͑see Fig. 2͒. The powder capsules are 16 mm long and 4 mm in diameter ͑slightly smaller than the 5 mm diameter of the holes͒. The capsules were placed in the inferior portion of the cavities. Rolled paper spacers were placed in the superior portion of each cavity to hold the TLD capsules in place and to fill the air cavity. The positions of the TLDs were chosen to be in the expected high dose region and in regions of relatively low dose gradient. If the dose gradient across a capsule is large, then interpretation of the reading is problematic. Ten TLD capsules were used as standards. The standards were given a dose of D 0 ϭ180 cGy. The average reading for the ten standards was Q 0 ϭ15.33 C and the standard deviation is 0 ϭ0.53 C ͑3.5%͒. We assume that this is characteristic of this batch of TLDs. The relatively large standard deviation limits the accuracy to which the Corvus and the Monte Carlo dose calculations can be tested. It is estimated that the uncertainty ͑one standard deviation͒ in the dose D measured by a TLD is: where nϭ10 is the number of standards irradiated and Q is the reading of the individual TLD. The value of D is very uniform among the TLDs at approximately 6.5 cGy. Thus differences between TLD measurements and predicted values of less than 2ϫ D ϭ13 cGy are not significant. The Corvus and MCNP predicted dose for each TLD capsule is based on the dose calculated by the treatment planning system at the center of the TLD capsule. A coordinate system transformation relating the coordinates used by the MCNP code and that used by Corvus was established based on the position of three fiducial markers. This transformation was used to ensure that the dose was evaluated at the same location for MCNP and Corvus.
The average measured TLD dose is 191.7 cGy, the average dose predicted by Corvus is 192.0 cGy, and the average predicted dose by MCNP is 186.9 cGy. The ratio of the Corvus average to the measured average is therefore approximately 1.002. A histogram of the dose ratios is shown in of 0.045. The average ratio for the MCNP predicted dose to the dose measured by the TLDs is 0.977 with a standard deviation of 0.045. It does not seem possible to distinguish the TLD measured dose from either the Corvus or MCNP predictions within the accuracy of the TLD measurements ͑3%-5%͒.
VII. DISCUSSION
Wang, Yorke, and Chui have compared the MSKCC pencil beam algorithm against an adaptation of the EGS4 Monte Carlo code for 6 MV IMRT plans for lung and head and neck treatments. 10 They have used patient CT data and have examined lung plans for five different patients. The pencil beam algorithm uses the equivalent path length method for inhomogeneity corrections. The location of our tumor volume is similar to that reported by Wang et al. in their study. All of their tumor volumes were in the right upper to middle medial lung and partially involved the mediastinum. These authors do not quote volumes for the CTV. Our CTV numerical volume is between FIG. 7. ͑Color͒ ͑a͒ The Corvus isodose distribution is superimposed on MCNP on the plane between slabs 14 and 15. The thick lines are MCNP. ͑b͒ The isodose curves from film measurements superimposed on the Corvus isodose curves. The thick lines are from measurement and the thin lines are from Corvus. their quoted GTV and PTV volumes. The lung plans employed 4 -6 beams, but we are not told the beam directions. The results of their comparison are very similar to our results. These authors find significant discrepancies in the medial and posterior portions of the right lung. The discrepancies are predominantly in the high dose, low electron density regions. The average ratio of the Monte Carlo mean dose ͑PTV͒ to the pencil beam mean dose for the five patients is 0.965.
Laub, Bakai, and Nüsslin have carried out a study in which they have compared the IMRT dose distribution in an Alderson anthropomorphic phantom thorax computed by a pencil beam algorithm with Monte Carlo calculated doses and measured doses ͑TLD and film͒. 11 The IMRT planning system is the KonRad IMRT system. Intensity modulated beams are delivered with the use of compensators which does present some complications. They have used five non-coplanar 6 MV beams. The Monte Carlo code is an adaptation of EGS4. They found good agreement between Monte Carlo, pencil beam and measured doses. These authors do not find that the pencil beam algorithm overestimates the dose in the target volume. They offer three possible explanations for this: ͑i͒ a phantom was used instead of a patient data set, ͑ii͒ non-coplanar beams suppress the differences, and ͑iii͒ intensity modulation. Reasons ͑i͒ and ͑iii͒ cannot be correct, as a phantom and intensity modulation have been used here. We cannot address reason ͑ii͒, as we have used coplanar beams.
Pawlicki and Ma describe a comparison between EGS4 based Monte Carlo calculations and Corvus for an upper thoracic target. Eight coplanar 4 MV beams are used for plan calculation. 12 The Monte Carlo calculations predict a 9% lower mean dose to the target than Corvus. Once again the pencil beam calculated doses are higher in the target than the EGS4 calculated doses because of electron transport out of the target into low density surrounding lung tissue. The Corvus pencil beam kernel is not laterally scaled to account for changes in lateral electron transport due to inhomogeneities. 13 In Sec. II of this paper a caveat is mentioned regarding the true electron density of tumor tissue. In our study, the electron density of the PTV is based on the CT numbers inside the physician drawn ͑phony͒ tumor volume. The electron densities inside the lung portion of the tumor volume are therefore characteristic of normal lung tissue. Real lung tumors may have electron densities which are higher than surrounding normal lung tissue. As the Wang et al. study is based on patient CT data, it can be presumed that their PTV electron densities accurately characterize lung tumor tissue. The fact that the results of Wang et al. are similar to ours suggests that either the electron densities that we have used are similar to patient tumor densities or that the results are insensitive to this. FIG. 8. ͑Color͒ A histogram of the ratios of Corvus predicted dose values to TLD measured doses and to MCNP predicted doses. The ratio of MCNP predicted doses to TLD measured values is also shown. Given the spread in these distributions ͑3%-5%͒ it is not possible to distinguish them.
VIII. CONCLUSION
We have made comparisons between calculated dose distributions and the measured dose distribution in the thorax of an anthropomorphic phantom for a lung treatment plan. The calculated dose distributions have been computed using the Corvus treatment planning software and by the MCNP Monte Carlo code. The MCNP computed and the measured dose values are in good agreement with Corvus values except in regions where the electron density is low and the dose is high. Corvus computes dose values which are up to 10% higher than MCNP in these regions. The average Corvus calculated dose for the CTV is 5% higher than for the MCNP computation. The clinical significance of these results is not clear to us.
These results are in good agreement with those of other workers who have compared a variety of pencil beam algorithms to EGS4 Monte Carlo calculations and measurement. The explanation suggested for this phenomenon is that pencil beam algorithms underestimate the degree of lateral electron transport out of low density regions and therefore overestimate the dose. 10 | 6,097 | 2003-09-01T00:00:00.000 | [
"Physics",
"Medicine"
] |
STATISTICAL CORRELATION BETWEEN SOCIOECONOMIC INDICATORS AND PROTECTED NATURAL AREAS AROUND THE WORLD
Protected natural areas are important for nature conservation. Nevertheless, some critiques point out that establishing and managing these areas may inhibit socioeconomic development and generate poverty. However, this strategy has presented varying results since some studies show positive eff ects on human development. The use of development indicators such as the Gross Domestic Product, the Human Development Index, the Gini Index, and the World Happiness Index can favor the understanding of how PNA management and goals may infl uence human development. This study aimed to correlate the existence of protected natural areas registered in the World Database on Protected Areas with the variables Gross Domestic Product, Human Development Index, Gini Index, and World Happiness Index of each country involved in the research. The correlations were tested for 145 countries through a non-parametric Spearman test. The correlation between the Gross Domestic Product and protected natural area percentage was positive yet nonsignifi cant. The correlation between the Human Development Index and protected natural area percentage was positive and signifi cant. As for the correlation between the Gini Index and protected natural area percentage, it was negative although non-signifi cant. Regarding the correlation between the World Happiness Index and protected natural area percentage, it was also negative but non-signifi cant. These results are possibly due to situational diff erences among the countries assessed. It is suggested that future studies such as this one be carried out by economically similar countries or regions to better elucidate the links between protected natural areas and socioeconomic development.
INTRODUCTION
A protected natural area (PNA) is a geographical space with clearly defi ned boundaries that is acknowledged, managed, and dedicated, through legal means or other eff ective means, to reach longterm goals of nature conservation while also ensuring the maintenance of ecosystem services and associated cultural values (Dudley, 2008). PNAs can vary in size, management, typology, and property depending on the country.
PNA creation and implementation are important in situ nature conservation strategies, responsible for the maintenance of biodiversity and natural resources (Adams et al., 2004). Despite their importance, some critics argue that PNA establishment and management can inhibit local or regional development and, besides failing to meet conservation goals, drive local populations into poverty, because environmental restrictions can preclude the maintenance of traditional practices needed for the subsistence of the peoples that used to occupy the protected land or its surroundings (Diegues, 2001;Colchester, 2010).
On the other hand, PNA establishment has presented varying results. Some studies have shown that PNA establishment on diff erent continents, countries, and regions can have positive socioeconomic consequences by generating income and stimulating regional and local development (Andam et al., 2010;Ferraro et al., 2011;Canavire-Bacarreza and Hanauer, 2013). PNAs generate income by supporting and promoting sustainable tourism programs, cooperatives of eco-friendly products, and encouraging educational activities and scientifi c research (Araújo, 2007). On average, they contribute to reducing poverty as much as they contribute to reducing deforestation rates (Leverington et al., 2010;Ferraro et al., 2011). Andam et al. (2010) found that, although communities near PNAs in Costa Rica and Thailand are substantially poorer than the countries' national averages, such observation does not support the hypothesis that inequality owes to the existence of PNAs. Several studies have shown that the net impact of environmental protection reduces poverty. In Brazil and other countries, PNAs allow local, regional, and national development, contributing signifi cantly to both the local subsistence of human populations and environmental conservation (Zambrano et al., 2010;Medeiros and Young, 2011;Naughton-Treves et al., 2011;Stevenson et al., 2013).
Other studies have also contributed to eradicate the myth that nature conservation is a hurdle for socioeconomic growth. For instance, Oldekop et al. (2016) showed that PNAs with positive socioeconomic results were more likely to have positive conservation results when they adopted co-management regimes, training the locals, reducing economic disparities, and maintaining cultural and subsistence benefi ts. Pawlewicz et al. (2017) noted that the Nature 2000 network (N2K) does not signifi cantly infl uence investment attractiveness in Polish territorial units, which could help to mitigate social confl icts that emerge following the establishment of N2K sites. Zeeshan et al. (2017) observed that the socioeconomic status of human settlers around the Keoladeo National Park in India improved after reconciling the community's interests with the governance strategy and through changes in resource utilization in the park, considered an internationally relevant humid zone recognized both as a Ramsar site and a World Heritage site by UNESCO. Chechina et al. (2018) noted through a community-level analysis that communities that depend upon a forest reserve in the Philippines have higher socioeconomic status due to greater access to natural resources.
An alternative to mitigate social confl icts generated by PNA establishment is payment for environmental services (PES). Silveira-Junior et al. (2020) found that PES can eff ectively contribute to confl ict resolution. However, PES results diff ered among the study cases due to certain attributes such as local environmental conditions, local economic activities, and stakeholder involvement capacity. The authors concluded that joint PES strategies are needed to improve effi ciency throughout the PNA management process, from diagnosis to decisionmaking stages.
Coupling environmental conservation and economic development is possible. For instance, Strassburg et al. (2014) found that the current productivity levels of cultivated pastures in Brazil fall between 32 and 34% of their full potential. According to these authors, raising these fi gures between 49 and 52% of their potential should be enough to meet meat consumption demands until the year 2040, without the need to convert natural ecosystems into new pastures or to expand the agricultural frontier. Also, in a comprehensive report on the implications of nature protection on several economic sectors (including agriculture, silviculture, fi shing, and the conservation sector itself), more than 100 economists and scientists found that the global economy would benefi t from a 30% increase in PNA extents in terrestrial and marine areas (Campaign for Nature, 2020). Analyzing the environmental Kuznets curve, Tritsch and Arvor (2016) verifi ed that socioeconomic growth does not seem to be a driver of deforestation in the Brazilian Amazon past the curve's infl ection point. The authors also highlighted that the areas subjected to active deforestation in 2010 present lower socioeconomic indicators than stabilized areas. The environmental Kuznets curve posits a relationship between carbon emissions and income in an inverted U-shape. At the initial stages of economic growth, income and emissions have a direct relationship. But as income grows and reaches a certain level, emissions present a downward trajectory. This is because higher-income populations are willing to couple economic growth with lower environmental degradation (Ávila and Diniz, 2015).
Some studies have aimed to relate socioeconomic development with PNAs. Hebron (2018) conducted a bivariate regression analysis using 2014 World Bank data and found no signifi cant correlations between number of PNAs and democracy levels, between social inequality and reduction in the number of PNAs, and between population increases and reduction in the number of PNAs. They observed, however, a signifi cant correlation between Gross Domestic Product (GDP) and number of PNAs. In this sense, Oprsal et al. (2018) conducted cross-sectional regression and panel regression analyses on variables that could infl uence PNA cover, using data collected between 1990 and 2012. They found that higher economic, social, or institutional development is related to higher proportions of PNA cover, revealing positive trends and important steps towards nature conservation.
The results obtained by Upton et al. (2008) suggested that high-income countries have a higher number of small-sized PNAs than low-income countries but with a higher extent of total protected land than the latter. However, the authors found few signifi cant links between poverty indices and PNA extent at the national level. On the other hand, Salvio et al. (2016), who correlated the distribution of conservation units with the Municipal Human Development Index (MHDI) and the Gini Index (IGini) of diff erent municipalities of Minas Gerais state (Brazil), observed that IGini varied signifi cantly between municipalities with conservation units and those without them. In other words, what they found is that municipalities with conservation units have higher income concentration, higher social inequality, and thus, higher IGini. They also observed that, in general terms, MHDI neither infl uences nor is infl uenced by the existence of conservation units, regardless of their nature, either full protection or sustainable use.
Using socioeconomic development indices, such as the Human Development Index (HDI) and the Gini Index (IGini) can improve our understanding of how PNA goals and management infl uence PNA surroundings (Salvio, 2017). The HDI was fi rst released in 1990 and broadened the perspective of human development towards incorporating the social dimension (PNUD, 2010). Currently, three principles underpin the HDI: education level, measured through a combination of adult literacy scores (with a 2/3 weight) and school attendance rate combined for primary, secondary, and higher education (with a 1/3 weight); and standard of living measured by real GDP per capita (PNUD, 2010). The IGini is a statistical measure of inequality, developed by the Italian statistician Conrado Gini in 1912, broadly used to indicate the degree of income concentration in a given region. It is based on the Lorenz curve (Hoff mann, 1998) and represented by a number between 0 and 1, whereby 0 corresponds to perfect income equality and 1 corresponds to perfect income inequality (Hoff mann, 1998). The Gross Domestic Product (GDP) is a synthetic indicator of the economy of a country, state, or city, and consists of the sum of all its goods and services produced in a specifi c time period, usually one year (IBGE, 2021). Finally, the World Happiness Index (WHI) is generated by the research data gathered for the World Happiness Report, a research conducted in more than 150 countries and based on citizen satisfaction about their own lives. The data are updated annually (World Happiness Report, 2020).
The most comprehensive source of information on the number of terrestrial and marine PNAs and the area covered by them on a global level is the monthly updated World Database on Protected Areas (WDPA). The WDPA is one of the main global data sources on PNAs and is broadly used by scientists, companies, governments, international secretariats, and others to inform policy planning, decision-making, and management. The WDPA is a joint venture between the United Nations Environmental Programme (UNEP) and the International Union for Conservation of Nature (IUCN). Compilation and management of the WDPA are done by the UN World Conservation Monitoring Centre -WCMC, with the collaboration of governments, non-governmental organizations, academia, and the industry.
The total number of PNAs recorded in the WDPA version released in March 2020 was 248,330, which encompassed 227,090 polygons and 21,240 points covering 245 countries and territories. Most of these areas are located within continents and collectively protect over 20 million km ² , which is 15.1% of the global terrestrial land surface. The marine protected areas (MPAs), despite being lower in number, cover 28,665,325 km ² , representing 7.91% of the global ocean surface. National waters encompass 39% of the Earth's ocean surface and, currently, 18.40% of these waters are designated as PNAs. On the other hand, only 1.20% of the extent of international waters, which make up the remaining 61% of the global ocean surface, are designated as PNAs. This lower fi gure among international waters owes to the complex jurisdictional context that aff ects PNA creation in these waters (WDPA, 2020). The magnitude of these numbers motivates several authors to consider PNAs as the main tool for biological diversity protection. PNA creation and management are important parts of any national environmental policy (Juff e-Bignoli et al., 2014).
Therefore, here we aimed to conduct statistical correlation analyses between PNA existence (as recorded in the WDPA) and the GDP, the HDI, the IGini, and the WHI of their countries of location in a global study.
Quantitative survey of PNA existence and socioeconomic indicators of each country
We performed a quantitative survey of PNAs located in 245 countries through the WDPA database (https://www.protectedplanet.net), version released in March 2020. We obtained 2019 GDP data on 200 countries from the offi cial website of the International Monetary Fund (IMF). The HDI and IGini data were surveyed from the Human Development Report published by the Human Development Report Offi ce of the United Nations Development Programme (UNDP), with data collected for 2019 in 176 and 145 countries, respectively. The WHI data were surveyed from the 2019 World Happiness Report (Helliwell et al. 2019) for 156 countries. These were the most upto-date data during the surveys.
We organized the data in a spreadsheet and removed the lines with absent data. This resulted in a list of 145 countries, which composed the study sample.
Sample size assessment
We performed a sample power analysis through the software GPower v. 3.1.9.4 adopting a signifi cance threshold of 0.05, which informed that a minimum of 134 observations was needed for a reliable correlation test. The samples evaluated herein were based on at least 145 observations, a number higher than the minimum threshold of 134 (Figure 1).
Normality test and correlation analyses
We tested data normality through Kolmogorov-Smirnov, Lilliefors, and Shapiro-Wilk tests, which informed us that the variables used in this study do not present normal distribution. We then opted for a non-parametric Spearman correlation test, which measures the degree of correlation between two variables (Vieira, 2010) and presents a non-parametric alternative for the Pearson correlation coeffi cient. The Spearman test must be used when the analyzed variables do not display two-dimensional normal distribution.
Both the Pearson and the Spearman correlation tests measure the degree of linear correlation between two quantitative variables. But in the Spearman correlation, specifi cally, the values are transformed into ranks and the correlation represents the linear correlation coeffi cient of the ranks (Hammer et al., 2001). The results vary between -1 and 1 and, while a value of 0 indicates no linear relationship at all, a value of 1 indicates a perfect linear relationship, and a value of -1 represents a perfect, but inverse, linear relationship, indicating that when one of the variables increases in value, the other decreases. The closest to 1 or -1, the strongest is the linear association between the variables (Hair et al., 2005). The correlation analyses were conducted in software SPSS version 24.
These results only revealed a positive and signifi cant correlation between PNA and WHI. No other signifi cant correlation was found between the other indices and the presence or absence of PNAs.
DISCUSSION
We found a non-signifi cant correlation between PNA percentage and GDP. This result disagrees with Hebron (2018), who observed a statistically signifi cant correlation between GDP and PNA. Our fi ndings also disagree with those by Andam et al. (2010), which indicated that the net impact of environmental protection reduces poverty. Conversely, our results resemble those of Upton et al. (2008), who found few signifi cant relationships between poverty indicators and PNA extent at the national level.
The countries with the highest HDI values also have greater percentages of protected land extent, which suggests that these countries acknowledge the relevance of natural resources and environmental services conservation on guaranteeing the standard of living of their populations. This result is congruent with the fi ndings of Oprsal et al. (2018), who concluded that higher economic, social, or institutional development is linked to higher PNA proportions.
We found a non-signifi cant correlation between PNA percentage and IGini. This result disagrees with the fi ndings of Salvio et al. (2016), which found a positive and signifi cant correlation between the existence of conservation sites and IGini, suggesting that conservation areas in the state of Minas Gerais (Brazil) are linked to greater social disparities, with greater income concentration and social inequality. On the other hand, this result reinforces the observations by Andam et al. (2010) who evidenced that, although communities near PNAs in Costa Rica and Thailand are substantially poorer than the national averages, this analysis does not support the hypothesis that inequality owes to the existence of PNAs.
Our results also suggest that the existence of PNAs does not infl uence neither is it infl uenced by the WHI. This fi nding corroborates the results of Neve and Sachs (2020), who observed that the UN Sustainable Development Goals 14 and 15 present an insignifi cant correlation with human well-being. Goals 14 and 15 are dedicated to the conservation and sustainable use of, respectively, marine and terrestrial ecosystems. Both feature among the list of 17 sustainable development goals that originated from the United Nations Conference on Sustainable Development that took place in Rio de Janeiro in 2012 aiming to meet urgent global challenges concerning environmental, political, and economic aspects.
According to the World Wide Fund for Nature -WWF (2008), the relationship between PNA and poverty is complex and multifaceted because, while in some cases PNAs have contributed to deepening populations further into poverty, in others, PNAs have had a positive role in poverty reduction. This may be due to the non-signifi cant correlations predominantly found in our analyses, which possibly owe to situational diff erences among the countries involved in our research. Upton et al. (2008) indeed confi rmed that the links between poverty and conservation are dynamic and, often, site-specifi c; at larger scales, the links between PNA and poverty remained undetermined.
Our results reinforce the WWF (2008) recommendation that site-specifi c studies be carried out, linking clear PNA management recommendations, local populations, and poverty reduction. Such studies could advance the current understanding of the relationships between PNA management and poverty, as well as evaluate the possibility to replicate a successful management action from one place to another. Besides, good governance is indispensable for attempts to link poverty reduction and PNAs to be successful, reinforcing the perception of PNAs as positive and welcome elements, rather than guarded and threatened sites.
CONCLUSIONS
Among the studied countries, those with higher HDI levels have also higher percentages of their extent under protection. At the same time, the correlation analyses between the existence of PNAs and GDP, PNAs and IGini, and PNAs and WHI did not produce signifi cant results, which possibly owes to situational diff erences among the diff erent countries. Regardless, our results allowed us to conclude that PNAs do not necessarily produce poverty. We would encourage future similar studies to segment economically similar countries or regions for a more refi ned understanding of the links between protected area management and socioeconomic development.
AUTHOR CONTRIBUTIONS
Ricardo Elói de Araújo: Data collection and organization, bibliographical research, analysis and discussion of results, and article writing.
Helder Antônio da Silva: Designed the methodology, statistics analysis and discussion of results.
Geraldo Majela Moraes Salvio: Conceived the ideas, designed the methodology, analysis and discussion of results. | 4,368 | 2022-01-01T00:00:00.000 | [
"Economics"
] |
Institutional Ownership and Firm Performance: Evidence from an Emerging Economy
: Using the Ordinary Least Square (OLS) estimation technique based on a sample of 180 listed firms from 2008 to 2018, this study investigates the impact of institutional ownership on firm performance in the Bangladeshi setting. Consistent with the “active monitoring” view, the results indicate that both domestic and foreign institutional investors have a positive effect on firm performance measured by Tobin’s Q and Return on Asset (ROA). In addition, this study explores whether the other corporate governance attributes—board size and board independence—operate as mediators between institutional ownership and firm performance. Our findings indicate that both board size and board independence have a significant positive impact on the relationship between institutional ownership and firm performance.
Introduction
Institutions around the globe have become progressively more important in shaping business. Mergers and acquisitions in the integrated business world of today have brought about a drastic shift in the corporate ownership structure of firms, a shift that is more inclined towards the concentration of institutional and foreign ownership. It is believed that institutional shareholders are efficient implementers of company information for effective investment (Bartov et al. 2000;Douma et al. 2006;El-Gazzar 1998;Ferreira and Matos 2008). Prior studies have indicated that institutional ownership impacts abnormal stock returns and, hence, firm performance (Ali et al. 2004). Extant literature has investigated the impact of institutional ownership on firm performance in various contexts, which puts forward three differing views: "active monitoring", "passive monitoring", and "exploitation". Pressure-resistant institutions independently acquit and actively engage in monitoring firm managers, while pressure-sensitive investors carry themselves as passive monitors, have a business relationship with the investee firm, or exploit the rights of the minority shareholders for their own gain. These differing investors are deemed to have vastly different impacts on the firm's performance (Brickley et al. 1988).
Prior studies have shown mixed and inconclusive evidence on the relationship between institutional ownership and firm performance. Some studies demonstrated a positive relationship between institutional ownership and firm performance (Elyasiani and Jia 2010;Omran et al. 2008;Yan and Zhang 2009;Yeh 2019) and the same positive correlation was noted between foreign ownership and firm performance (Bentivogli and Mirenda 2017;Kao et al. 2018), hence, confirming the "active monitoring" view. On the contrary, a negative (Muttakin et al. 2012) or no association (Demsetz and Villalonga 2001) between institutional ownership and firm performance has also been concluded in some past studies, leading towards a more "passive monitoring" view of investors. In the context of Bangladesh, Mollah et al. (2012) concluded there was no impact of board ownership on firm performance; however, other corporate governance mechanisms like institutional shareholdings, board salary, dividend, etc., might boost firm performance. Rashid (2020) showed the relationship between institutional ownership and Tobin's Q was not statistically significant; however, he found a significant positive association between institutional ownership and accounting-based firm performance.
This study draws attention to an emerging market owing to the considerable differences in economic systems, tax systems, corporate ownership structures, and corporate governance that are noticed between East Asia and the United States (Claessens and Fan 2002;Hasan et al. 2020;Peng et al. 2008). The economic growth rate of SAARC countries has been commendable, and along with their ample natural and human resources, these markets show great potential. In addition, the linkage between corporate ownership structure, corporate governance, and firm performance in emerging markets have been of interest in past literature as well (Bose et al. 2017;Lou et al. 2020;Mukhopadhyay and Chakraborty 2017). We focus our study on Bangladesh for numerous reasons. First, in the last decade, due to sustained economic expansion and financial development in South Asia, Bangladesh has been declared an "emerging Asian tiger" (Alom 2018). The GDP growth rate was 7.86% in the fiscal year 2017-2018, one of the highest in the world (Centre for Research and Information (CRI) 2018), motivating us to select Bangladesh as a region of interest. Second, the majority of the firms in Bangladesh are primarily family-owned, which results in highly concentrated ownership that plays an important role in shaping corporate decisions. Third, the Bangladeshi economy has seen a shift from the conventional family-owned business structure and has witnessed a rise in institutional shareholdings. In 2020, the percentage of institutional shareholding of total Dhaka Stock Exchange market capitalization rose to 15.6%. We expect this restructuring of ownership composition to strengthen corporate monitoring, which might elevate firm performance. Fourth, the Securities and Exchange Commission Bangladesh in 2006 announced the Corporate Governance Notification (CGN), which required listed firms to have independent directors in the Anglo-American style on their boards, at a ratio of 1:10 (Rashid 2020). It will be interesting to see the role of institutional ownership in shaping firm performance, given this reform. Finally, Bangladeshi firms are afflicted by a lack of regulatory control, corporate accountability, and transparency. Another concern is market irregularities and malpractices, such as the unavailability of information to investors, insider control, and collaboration to manipulate stock prices in the market. Considering these issues, this study aims to test whether institutional shareholders are effective implementers of monitoring mechanisms within the investee firms in the context of Bangladesh.
Using the OLS estimation technique, this paper aims to demonstrate whether the presence of institutional investors within the listed companies in Bangladesh favorably influences firm performance. We obtained data from 180 listed companies on the Dhaka Stock Exchange over the period of 2008-2018. Our measures of institutional ownership comprise domestic and foreign ownership. Firm performance was measured by Tobin's Q and Return on Asset (ROA), which are widely used proxies for measuring firm performance. Our results support the "active monitoring" view, where Bangladeshi institutional investors are actively involved in monitoring and facilitating firms toward performance enhancement. We further analyze whether corporate governance attributes like board size and board independence are channels through which the relationship between institutional ownership and firm performance is mediated. The results are indicative that when they are included in the model along with institutional ownership, board size, and board independence, they show a significantly positive impact on Bangladeshi firm performance.
Our paper contributes to the existing literature owing to the plausible different impact institutional ownership might have on emerging markets, such as Bangladesh, given the different economic profile, weak implementation of the legal framework, and poor corporate governance compared to their developed counterparts. The closest to our study is the paper by Rashid (2020), which concluded that foreign and director ownership has a significant positive impact on a firm's performance considering both accounting and market-based performance, whereas institutional ownership positively impacts only accounting-based performance. However, this study was conducted for three years with a limited sample size, which might not lead to comprehensive results. This study includes a dataset obtained over a longer period of time (2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018), which eliminates estimation bias and assists in generalizing the outcomes to a much wider context. The relationship between institutional ownership and firm performance is a fundamental area of interest; hence, this study will aid managers, current and potential investors, and regulators in making key decisions and contributing to the value enhancement of firms. Since emerging financial markets possess similar market traits, our findings will enable not only Bangladeshi policymakers, but other emerging economies to set up an effective structure for corporate governance and make better and well-informed policy decisions.
The rest of the paper is organized as follows: Section 2 discusses relevant literature, and Section 3 describes the research methodology. Section 4 presents the empirical results, followed by Section 5, which displays channel analysis for the relationship between institutional ownership and firm performance. Section 6 concludes the paper and provides policy recommendations.
Theoretical Evidence
The separation of ownership and control leads to agency problems and moral risks if managers put short-term profits ahead of long-term profits by using the company's resources for their own personal gain (Hart and Quinn 1993). The corporate ownership structure acts as a crucial governing mechanism that pressurizes the managers to strive for company success. Shleifer and Vishny (1986) contend that, among several executive equity ownership mechanisms, institutional shareholding is deemed to be an ideal form by which to rigorously monitor the managers.
The existing literature on the effects of institutional ownership on corporate governance mechanisms has brought forward a few differing views: (i) Active Monitoring, (ii) Passive Monitoring, and (iii) Exploitation. According to the "active monitoring" view, institutional investors have the power to exert pressure on company managers to improve the company's corporate governance mechanisms in order to reduce information asymmetry and increase firm operational transparency (Lin and Fu 2017;Shleifer and Vishny 1986). With the aid of sophisticated managerial expertise, high-quality resources, and specialized skills, institutional investors can supervise managers, aiding them in improving the quality of managerial decisions to ensure increased corporate value (Elyasiani and Jia 2010;Firth et al. 2016). Through the large proportion of shareholding (and voting rights) of the institutional investors, the monitoring role also enables them to indicate any dissatisfaction with the concerned company's management (McCahery et al. 2016) which ensures that the firm's performance is never compromised.
On the other hand, the "passive monitoring" view states that institutional investors have no intention of monitoring the day-to-day management of the investee firm, as they might solely be interested in myopic financial goals or adjusting portfolio requirements (David and Kochhar 1996;Elyasiani and Jia 2010). The "passive monitoring" view suggests that institutional investors sometimes invest in a company based on some insider information advantages and might not have any intention of long-term investment or improvement of the company's corporate governance and firm performance (Elyasiani and Jia 2010). In this instance, institutional investors act as "traders", buying and selling shares within a short span of time. Thus, such institutional investors will have no significant impact on the investee firms' performance. Alternatively, the "exploitation" view states that institutional investors might conspire with the investee firm's managers to exploit the shareholder's wealth and obtain additional benefits from the company (Elyasiani and Jia 2010). According to the exploitation view, institutional investors ignore resolving organizational fraud or misappropriation of company wealth if they themselves also benefit from such misconduct (Cornett et al. 2007).
The distinguished monitoring role of institutional investors is further enhanced when the institutional shareholders come from a thriving foreign economy. Foreign institutional shareholders can minimize agency problems and enhance the firm's performance with the aid of their superior monitoring skills, global expertise, and multinational portfolios. Foreign institutional investors are able to offset the costs of board ownership through their substantial shareholdings, and they are able to oversee managerial behavior even though company boards within emerging economies are typically controlled by familyconcentrated boards. Corporate governance, management expertise, and strict monitoring are all domains where firms might benefit from the presence of foreign institutional investors (Huang and Zhu 2015). Due to their internationally diversified portfolios, independence from local management, absence of conflicts of interest, and professional monitoring abilities; foreign institutional investors play a stronger role than domestic institutional investors in internal corporate governance, and hence, improve the firm's value as well as operating performance (Lou et al. 2020).
Empirical Evidence
Although it is evident that institutional shareholding can possibly impact firm performance, studies, however, have shown inconclusive results. For instance, employing a large data set consisting of Chinese listed firms between 2004 and 2014, Lin and Fu (2017) documented the positive impact institutional investors posit on Chinese firm performance, supporting the "active monitoring" view. Using data from an emerging economy like India, Kansil and Singh (2018) opined that there is a significant positive relationship between institutional shareholding and firm performance. Institutional shareholders motivate companies to undertake good governance practices, and they have the duty to protect the interests of the principals of the company, which then translates into enhanced corporate performance (Connelly et al. 2010;Cornett et al. 2007;Hussain Tahir 2015;Tornyeva and Wereko 2012). Institutional shareholders also act as monitors, who tend to put a check on the expropriating behaviors of the self-serving owners of the firm and improve overall firm performance (Mokhtari and Makerani 2013;Su et al. 2013). The positive impact of the presence of institutional investors on firm performance is imminent, supporting the benefits the investee firm derives from the monitoring role of the institutional investors (Fitri and Surjandari 2022;Hai et al. 2018;Herdjiono and Sari 2017;Lin and Fu 2017;Lou et al. 2020).
Previous studies have suggested that foreign institutional shareholding is a source of good governance practices and improved firm performance (Bai et al. 2004;Douma et al. 2006). Companies with a high level of foreign institutional ownership are expected to experience higher growth and improved performance with the supply of external funding and foreign managerial expertise (Ferreira and Matos 2008). Because companies in developing countries have highly concentrated ownership, Santiago-Castro and Baek (2004) argue that foreign institutional investors are crucial, as their active monitoring role will mitigate insider exploitation. Studying 18 emerging economies, Lins (2003) suggested that domestic institutional shareholders of emerging markets have limited resources and encounter political constraints that obstruct them from effectively performing their monitoring role. However, foreign institutional shareholders can operate without any unjust political or governmental intervention. Thus, in an emerging economy, foreign shareholders are able to effectively perform the monitoring role, positively impacting firm performance. Similar results were documented by Khanna and Palepu (2000), who contended that for an emerging economy like India, firm performance is positively associated with foreign institutional ownership and negatively associated with domestic institutional ownership.
Contrary to the above findings that favor the presence of institutional shareholders within the firm, several researchers have put forward the adverse effect institutional investors have on firm performance (Musallam et al. 2018;Tsouknidis 2019). For instance, Musallam et al. (2018), using a dataset from the Indonesian Stock Exchange, and Tsouknidis (2019), using a dataset of U.S.-listed shipping companies, showed a significant negative correlation between institutional ownership and corporate performance, suggesting that institutional investors are unable to effectively monitor the managers, adversely affecting firm performance. Such institutional investors are more inclined towards earning shortterm benefits by aligning interests with incompetent managers rather than monitoring those managers for further improvement of the company's performance (Pound 1988). This supports the "Exploitation" view. Using data from India from the early 1990s, Khanna and Palepu (2000) contended that domestic institutional investors have detrimental effects on firm performance for companies within emerging economies, as domestic investors in an emerging economy may not have the business knowledge and expertise equivalent to their international counterparts. In terms of foreign institutional shareholders, Thanatawee (2014) showed that firms with higher foreign institutional shareholders tend to have lower corporate value.
Similar to several other East Asian economies, the corporate control framework of the majority of Bangladeshi firms is insider-oriented, or significant stakes in the company are family-owned. Insiders or family owners, therefore, can easily manipulate corporate valuation and may not act as appropriate corporate monitors. Due to such highly concentrated ownership within a non-efficient market, external governing mechanisms like the presence of institutional shareholders can positively impact firm performance. Rashid (2020) showed the relationship between institutional ownership and Tobin's Q is not statistically significant. but found a significant positive association between institutional ownership and ROA. Such ambiguous results give us the motivation to develop the following hypothesis to further delve into the analysis of the relationship between institutional ownership and firm performance: Hypothesis 1 (H1). Institutional ownership (domestic/foreign) positively impacts firm performance.
Sample Construction and Data Sources
We developed our sample size by extracting information from several sources. For the investigation of the proposed hypothesis, data was obtained from all the listed companies on the Dhaka Stock Exchange, from 2008 to 2018. The ownership structure of the sample data was manually collected from annual reports of publicly listed companies in Bangladesh. Excluding financial firms, our final sample included 180 firms. Financial variables for the study were all obtained from the DataStream Database.
Dependent Variable: Firm Performance
We employed two proxies in order to measure firm performance: Tobin's Q and Return on Asset (ROA) (Daryaei and Fattahi 2020;Hamza et al. 2020;Hsu and Wang 2014;Sakawa and Watanabel 2020). Tobin's Q is the market-based firm performance measured by adding the fair market value of common stocks with the book value of total liabilities and dividing it by total assets. ROA is an accounting-based measure of how well a company is doing. It is measured by dividing the company's net income by its total assets.
Independent Variable: Domestic Institutional Ownership and Foreign Institutional Ownership
Institutional ownership was our key independent variable. Institutional shareholding was categorized into domestic and foreign ownership. Domestic ownership (INST_DOM) included shares held by domestic institutions like insurance companies, banks, mutual funds, and pension funds, plus special accounts. Foreign institutional ownership (INST_FRGN) was calculated as the percentage of ownership held by foreign investors.
Control Variables
To ensure consistency and comparability with previous papers, Daryaei and Fattahi (2020); Drobetz et al. (2021) controlled for several variables that may potentially influence firm performance. The first control variable we apply is the firm size (Size) which is measured as the natural logarithm of total assets in millions of US $. It is expected that larger firms with higher bureaucracies may become inefficient at business development and the decision-making process. Such an incompetent business operation will adversely impact corporate value. Second, we control for liquidity (Liquidity). We expect that firms with higher liquidity will have better firm performance, as higher liquidity implies firms' solid ability to raise cash when needed, which can consequently be an underlying reason for higher firm performance. Our third control variable, leverage (Leverage), is calculated as the ratio of total debts to total assets. Leverage illustrates how a firm funds its assets and the firm's overall ability to meet its monetary obligations in due time. A higher leveraged firm denotes the firm's inadequate capital structure, making such firms riskier. Riskier firms tend to have an overall detrimental business operation, which is likely to adversely impact firm performance. Next, we control for asset growth (Asset growth) which shows the percentage change of company assets within a given period of time. Asset growth reflects the management efficiency of asset capitalization, and thus, is expected to positively impact firm performance. The fifth control variable is stock price volatility (Price volatility) which is measured by the average stock price fluctuations over a given period of time. Higher stock price volatility translates to higher risk for firms, which may also adversely influence firm performance. Lastly, we also employ property, plant, and equipment net (PPENT) as a control variable. All variables are defined in Table 1. The natural logarithm of the book value of a firm's asset Datastream Liquidity Cash divided by total assets Same as above Leverage The ratio of a firm's total debt to the book value of its assets Same as above Asset Growth Percentage growth in assets Same as above Price Volatility Measured as standard deviation of share price Same as above PPENT Property, plant and equipment divided by total assets Same as above
Estimation Model
To investigate the relationship between institutional ownership (domestic and foreign) and firm performance (Tobin's Q and ROA), we specify the following OLS regression model: where Performance it is measured by Tobin's Q and ROA, INST represents institutional ownership as domestic institutional ownership (INST_DOM) and foreign institutional ownership (INST_FRGN). Size denotes the firm size, measured as the logarithm of total assets; liquidity is measured as cash divided by total assets; leverage is measured as total debt divided by total assets; Asset growth is measured by percentage growth in total assets; price volatility is measured as the standard deviation of share price; and PPENT is measured as property, plant, and equipment divided by total assets, which are used to control for the effects of firm-level characteristics.
Summary Statistics and Correlation Matrix
The descriptive statistics for our sample are presented in Table 2. Two performance proxies, Tobin's Q and ROA, have mean (median) values of 1.59 (0.92) and 8.3 (6.69). INST_DOM has a mean (median) of 0.19 (0.17) and a standard deviation of 0.15. INST_FRGN has a mean (median) of 0.10 (0.02) and a standard deviation of 0.19. INST_DOM varies between 0% and 73%, whereas INST_FRGN varies between 0% and 90%. The mean (median) of firm size (Size) is 16.69 (16.59). The mean (median) of firm leverage (Leverage) is 0.92 (0.10), with a standard deviation of 4.15. Moreover, asset growth has a mean (median) value of 18.85 (13.10) with a standard deviation of 25.78, while the mean (median) of stock price volatility (Price volatility) is 0.13 (0.12). Table 2 presents summary statistics of Tobin's Q and ROA, institutional ownership proxies and other control variables used in this study. All variables are defined in Table 1 and winsorized at the 1% and 99% levels.
The pair-wise correlation matrix is shown in Table 3. The correlation between Tobin's Q and the ROA spread is 0.35, which is positive, as predicted. The stronger association implies that the majority of businesses with a high Tobin's Q also have a high ROA spread, implying that Tobin's Q and ROA spread are complementary proxies for company performance. Both domestic and foreign ownership has a positive correlation with Tobin's Q as well as ROA. All regression estimations are separately performed for institutional and foreign ownership to avoid multicollinearity. We also note that the correlation among the control variables is low, suggesting that multicollinearity is not a problem in the estimation. Additionally, we compute and examine each independent variable's variance inflation factor (VIFs). In all cases, the VIFs are below 2 (not reported here), which is far below the critical value of 10, suggesting multicollinearity is not an issue in the model.
Baseline Regression Results
Using OLS regression, we estimate Equation (1) to examine the influence of institutional ownership on market and accounting performance, as measured by Tobin's Q and ROA, respectively. The baseline regression results are presented in Table 4. In columns 1 and 3, we show the unincorporated impact of institutional ownership on firm performance without controlling for any firm-specific characteristics. Our first key variable of interest, INST_DOM, shows a positive relationship with Tobin's Q as well as ROA, both having a 5% significance level. In columns 2 and 4, we re-ran the regression results after controlling for the firm-specific characteristics. Our findings continue to exhibit a positive association between INST_DOM and both firm performance measures, with a 1% significance level. The coefficient of 3.241 indicates that an increase of one within-firm standard deviation (0.15) in institutional ownership is associated with a 0.49 increase in Tobin's Q, which is equivalent to 30.1% of the average Tobin's Q score and nearly 19.21% of the standard deviation of Tobin's Q. ROA's coefficient of 7.014 indicates that an increase of one within-firm standard deviation (0.15) in institutional ownership is associated with a 1.05 increase in ROA, which is equivalent to 12.67% of the average ROA score and nearly 14.74% of the standard deviation of ROA. Therefore, these findings are economically significant. Based on our findings, we can state that the monitoring role of the institutional owners significantly improves firm performance.
The result is consistent with the "active monitoring" and similar to the findings reported by Elyasiani and Jia (2010) (2019), who reported a negative relation between institutional ownership and firm performance in the context of the USA. The presence of institutional shareholders mitigates the value-reducing activities of opportunistic managers, as the institutional shareholders rigorously monitor entrenched managers. Hence, under rigorous monitoring by the institutional owners, managers are forced to act in the best interest of the shareholders and make decisions that are most beneficial for the firm. Our results are in line with the outcomes of Sakawa and Watanabel (2020), who, using a Japanese dataset, also showed that the monitoring role of institutional investors significantly improves firm performance. In summary, our results show that Bangladeshi institutional investors have become more involved in monitoring and aiding enterprises in improving their performance.
Turning to the control variables, we observe several significant relationships. The coefficient of firm size (Size) with both measures of performance is negative and significant at a 1% level, indicating that an increase in firm size decreases a firm's market performance (Tobin's Q), as well as accounting performance (ROA). The economic significance of firm size is significantly higher compared to the economic significance of other control variables. For example, an increase of one within-firm standard deviation in the firm size decreases Tobin's Q by 1.51, which is nearly 59.21% of the standard deviation and 94.96% of the mean of Tobin's Q. Likewise, an increase of one within-firm standard deviation (1.75) in firm size is associated with a 3.36 decrease in ROA, which is equivalent to 40.52% of the average ROA score and nearly 47.20% of the standard deviation of ROA. Thus, having more layers, more bureaucracy, and a more inflexible structure, the large firm can impair the board's capacity to make strategic decisions, resulting in a negative influence on the firm's performance (Gong et al. 2013). Leverage, another control variable, has a negative effect on Tobin's Q and is statistically significant at the 5% level. Leverage continues to exhibit a negative association with ROA. However, the coefficient is statistically insignificant. Therefore, firms with a high leverage ratio in their capital structure see their performance being negatively impacted (Anderson and Reeb 2003;Elyasiani and Jia 2010). Thus, financial managers should cautiously proceed while using financial leverage, which can impact the firm's value. We also find that stock price volatility (price volatility) has a negative, significant effect on both Tobin's Q and ROA at the 1% level, implying that both Tobin's Q and ROA fall as stock market volatility increases. Moreover, the correlation between board independence and firm performance (Tobin's Q and ROA) has great economic significance. For instance, an increase of one within-firm standard deviation in the stock price volatility (price volatility) decreases Tobin's Q by 0.959, which is nearly 60.31% of the mean of Tobin's Q, and an increase of one within-firm standard deviation in the stock price volatility decreases ROA by 2.54, which is nearly 30.64% of the mean of ROA. In Table 5, we examine the impact of foreign institutional ownership on firm performance. Columns 1 and 3 show the results without the application of the firm-level control variables. Our second key variable of interest, INST_FRGN, shows a significant positive association with both firm performance measures across all regression results at a minimum 5% significance level. After controlling for the firm-specific characteristics, we show that an increase of one within-firm standard deviation in the percentage of foreign ownership increases Tobin's Q by 0.55, which is 34.59% of the mean of Tobin's Q and nearly 21.56% of the standard deviation of Tobin's Q, indicating that an increase in INST_FRGN substantially increases Tobin's Q score of Bangladeshi firms. Additionally, at a 5% level of significance, INST_FRGN exhibits a positive correlation with ROA. An increase of one within-firm standard deviation in the percentage of foreign ownership increases ROA by 1.22, which is 14.68% of the mean of ROA and nearly 17.13% of the standard deviation of ROA, indicating that an increase in INST_FRGN substantially increases the ROA score. Similar to the findings of Ferreira and Matos (2008), Lin and Fu (2017), and (Bena et al. 2017) we also document a positive correlation between foreign ownership and firm performance. Through active monitoring, management skills, and financial resources, foreign institutional investors strengthen the firm's corporate governance mechanisms and, as a result, improve corporate performance. In a study based on an emerging market like Taiwan, Hsu and Wang (2014) showed that the positive association between institutional ownership and firm performance is more pronounced for foreign institutions, as foreign institutions can be more resourceful and have better executive skillsets. Our findings, therefore, suggest that in the context of Bangladesh, foreign institutional ownership tends to improve firm performance. The results are consistent with Imam and Malik (2007), who used a dataset from the Dhaka Stock Exchange to conclude that foreign institutional shareholders tend to improve the performance of Bangladeshi firms. Table 1. Superscripts *** and ** denote statistical significance at 1% and 5% level, respectively.
The Effect of Corporate Governance Mechanisms
In this section, we aim to examine whether corporate governance characteristics, such as the size of the board and the proportion of independent directors within the board, may influence the overall effectiveness of a company's ownership structure and its mechanisms. In the next section, we employ two corporate governance attributes-board size and board independence-to examine whether they have any mediating effect on the relationship between institutional ownership and firm performance.
Mediating Impact of Board Size on the Relationship between Institutional Ownership and Firm Performance
The larger number of directors allows the board to gather diverse intellect and skills and utilize such expertise to improve strategic decision-making (Gómez et al. 2017). In other words, the effectiveness of the board increases with an increased number of directors within the board. As a result, studies show that a larger board is associated with improved firm performance (Hamza et al. 2020;Sanda et al. 2010). Based on the discussion, we expect that board size will mediate the relationship between institutional ownership and firm performance in a positive direction.
The regression results are reported in Table 6. The results demonstrated in columns 1, and 2 show that board size has a positive effect on the relationship between institutional ownership and Tobin's Q at the 5% level of significance and on the relationship between institutional ownership and ROA at the 10% level of significance. Furthermore, in columns 3 and 4 of Table 6, board size has a favorable influence on the linkage between foreign institutional ownership and Tobin's Q at the 1% level of significance and on the relationship between foreign institutional ownership and ROA at the 10% level of significance. A larger board tends to provide greater and more heterogeneous expertise to develop superior business development strategies and contribute to optimum decision-making processes. In addition to that, the larger the board size, the more monitoring is assumed, assisting the board of directors in making solid corporate decisions (Pearce and Zahra 1992) by leveraging a range of diverse expertise from their diverse backgrounds (Kao et al. 2018) and thereby boosting the firm's performance (Choi and Wong 2007). So, our results suggest that board size plays an important role in solidifying the corporate governance mechanisms of Bangladeshi companies. Year effects Yes Yes Yes Yes Table 6 shows results for analyzing the mediating impact of board size on the relationship between institutional ownership and firm performance. T-statistics are in parentheses. Definitions of all variables are provided in Table 1. Superscripts ***, **, * denote statistical significance at 1%, 5%, and 10% level, respectively.
Mediating Impact of Independent Directors on the Relationship between Institutional Ownership and Firm Performance
The agency theory argues that the advisory and monitoring roles are more prominently performed by independent directors as they do not take part in the day-to-day operation of the business. Thus, board independence acts as an internal control method, and an independent director's decisions are more strategic and objective toward achieving the shareholder's interest (Brickley and Zimmerman 2010). In 2006, the Securities and Exchange Commission of Bangladesh created the Corporate Governance Notification, under which it was made mandatory for listed Bangladeshi firms to have at least one independent director on a board with fewer than ten members. Such a convergence of Bangladeshi corporate governance mechanisms with those of the Anglo-Saxon makes it rational to examine the influence of board independence on the relationship between institutional ownership and firm performance. Independent directors possess the capability to govern and challenge management decisions in order to protect the company's reputation and continue to improve firm performance. Zhu et al. (2016) show evidence that the effective monitoring role of independent directors is associated with higher firm valuation. In line with this discussion, we expect that, in the context of Bangladesh, the presence of independent directors within the board will have a positive reinforcement of the relationship between institutional ownership and firm performance.
The results are reported in Table 7. In columns 1 and 2, we show that board independence has a favorable influence on the link between institutional ownership and Tobin's Q and ROA, at a significance level of 5% and 1%, respectively. Furthermore, through columns 3 and 4 of Table 7, we demonstrate that board independence has a significant positive influence on the association between foreign ownership and Tobin's Q at the 5% level of significance and on the relationship between foreign institutional ownership and ROA at the 1% level of significance. Independent board members with extensive experience and competence, as well as enhanced monitoring capabilities and expanded networks (Fama and Jensen 1983), strengthen the relationship between ownership structure and business performance. Our results provide evidence that, in addition to the vigorous monitoring by the institutional shareholders, the monitoring role of the independent directors acts as a catalyst for the positive relationship between institutional ownership and firm performance. Table 7 shows results for analyzing the mediating impact of board independence on the relationship between institutional ownership and firm performance. T-statistics are in parentheses. Definitions of all variables are provided in Table 1. Superscripts ***, **, * denote statistical significance at 1%, 5%, and 10% level, respectively.
Conclusions
This paper investigates the impact of institutional ownership (domestic and foreign) on firm performance in the context of an emerging economy-Bangladesh-over the period of 2008-2018, using Tobin's Q and Return on Asset (ROA) proxies of firm performance measures. Our paper extends the existing empirical literature by exploring the role of both institutional and foreign shareholders on firm performance in the context of Bangladesh, using a dataset covering 180 non-financial firms from DSE. Research in the context of an emerging nation, such as Bangladesh, is scarce. Hence, this paper adds a view to the already vast existing literature, the data of which are mainly focused on developed nations.
After controlling for firm-specific characteristics (such as firm size, liquidity, leverage, asset growth, price volatility, and property, plant, and equipment net), we provide evidence that Bangladeshi institutional shareholders are facilitating firms towards boosting their performance. The result is coherent with the "active monitoring" view, which implies that institutional shareholders, with strong supervision of the managers, can in turn help managers make sound strategic decisions regarding the firm. This will mechanically lead to a reduction in information asymmetry and an intensification of the transparency with which the firm operates to ensure increased corporate value. This monitoring authority of institutional owners with specialized skills will align firm-specific decisions in favor of shareholders that are most beneficial for the firm. Following that, we also demonstrate a positive association between increased foreign institutional shareholding and firm performance. In an emerging economy, foreign institutional shareholders supply skilled management expertise and high-end resource endowment, and most importantly, they act as superior corporate monitors, which further enhance firm performance. Additionally, we explore two other corporate governance characteristics-board size and board independence-to determine whether they mediate the relationship between institutional ownership and firm performance. Our results indicate that board size has a significantly positive impact on firm performance in Bangladeshi companies. Next, we explored the impact of board independence on the relationship between institutional ownership and firm performance. Our results suggest that not only robust monitoring by the institutional shareholders, but also the strong monitoring role played by the independent directors are the foundations for the positive relationship between institutional ownership and firm performance. Independent board members, with their extensive experience and competency, coupled with their monitoring capabilities and a large network base, do reinforce the positive relationship between ownership structure and business performance.
Our study has several important implications for shareholders, firms, and regulators. Through the threat of exit due to incompetency, institutional shareholders can discipline the management, as well as mitigate the plausible lack of financing by increasing liquidity, which may ameliorate firm performance. As is evident in this paper, institutional (both domestic and foreign) shareholders, through their rigorous monitoring role, can trigger strong corporate governance, which will in turn increase firm performance in an emerging economy like Bangladesh. In Bangladesh, insider ownership currently comprises 30% of the overall ownership structure within Bangladeshi firms. Such a founder-family-controlled board lacks independence, and thus, inadequately represents minority shareholders and institutional investors. Based on the findings of our study, it is recommended that the government and standard setters must restructure the ownership setting and prioritize increasing institutional shareholdings within Bangladeshi firms for improved governance and performance. Consequently, regulators might want to look into the corporate ownership structure of firms in Bangladesh to aid in striking a balance in ownership structure. This is because institutional and foreign shareholders tend to be high-risk takers looking for a high-risk-return payoff. This high-risk-taking tendency might lead to indulgence in risky investments, which might have adverse implications for firm value. Hence, our study appeals to finding an equilibrium between internal regulatory mechanisms and ownership structure. Consequently, this research also contributes to the vast field of ownership structure in academic literature from an emerging market perspective.
We acknowledge that our regression results may have some limitations. We employed the OLS regression technique, for which the obtained results could be biased due to endogeneity issues. Future research could concurrently investigate the relationship between institutional ownership and firm performance on the one hand, and the cost of debt on the other. This will help get a better idea of the risk-return trade-off, and hence, aid managers in making informed decisions. In addition to that, instead of looking into the impact of institutional ownership on financial performance, studies can also delve into the possible impacts of institutional ownership on non-financial performances. Some studies carried out in the context of the USA included the effect of institutional ownership on carbon emission (Safiullah et al. 2022), firm transparency and information production (Boone and White 2015), sustainability reporting (Nulla 2015), and CSR activities (Kim et al. 2019). It would be noteworthy to see how these results might differ in the context of an emerging economy like Bangladesh.
Conflicts of Interest:
The authors declare no conflict of interest. | 8,846.6 | 2022-11-30T00:00:00.000 | [
"Economics",
"Business"
] |
RepeatFiller newly identifies megabases of aligning repetitive sequences and improves annotations of conserved non-exonic elements
Abstract Background Transposons and other repetitive sequences make up a large part of complex genomes. Repetitive sequences can be co-opted into a variety of functions and thus provide a source for evolutionary novelty. However, comprehensively detecting ancestral repeats that align between species is difficult because considering all repeat-overlapping seeds in alignment methods that rely on the seed-and-extend heuristic results in prohibitively high runtimes. Results Here, we show that ignoring repeat-overlapping alignment seeds when aligning entire genomes misses numerous alignments between repetitive elements. We present a tool, RepeatFiller, that improves genome alignments by incorporating previously undetected local alignments between repetitive sequences. By applying RepeatFiller to genome alignments between human and 20 other representative mammals, we uncover between 22 and 84 Mb of previously undetected alignments that mostly overlap transposable elements. We further show that the increased alignment coverage improves the annotation of conserved non-exonic elements, both by discovering numerous novel transposon-derived elements that evolve under constraint and by removing thousands of elements that are not under constraint in placental mammals. Conclusions RepeatFiller contributes to comprehensively aligning repetitive genomic regions, which facilitates studying transposon co-option and genome evolution. Source code: https://github.com/hillerlab/GenomeAlignmentTools
The nature of repetitive sequences such as transposons, however, leads to many paralogous alignments, which pose a challenge for comprehensively aligning orthologous repeats between vertebrate genomes. Most methods for aligning entire genomes use a seed-and-extend heuristic, originally implemented in BLAST [18], to find local alignments between the sequences of two genomes. The seeding step of this heuristic detects short words or patterns (called seeds) that match between the sequences of the two genomes. This can be computed very efficiently. Seed detection is then followed by a computationally more expensive alignment extension step that considers ungapped and gapped local alignments. Given that repetitive sequences provide numerous seed matches to paralogous repeat copies in a whole genome comparison, it is computationally infeasible to start a local alignment from seeds located in repetitive sequences. Therefore, seeds that overlap repetitive regions are not used to start a local alignment phase, either by masking repetitive regions before aligning genomes [19][20][21][22] or by dynamically adapting seeding parameters by the observed seed frequencies [23]. Consequently, alignments between highly-identical repeats are only found during the extension phase, initiated from seeds outside the repeat boundaries. This can be problematic if the regions flanking a repeat have been diverged to an extent that no seed in the vicinity of the repeat can be found.
Here, we investigated to which extent aligning repetitive sequences are missed in whole genome alignments. We show that ignoring repeat-overlapping seeds misses between 22 and 84 Mb of mostly repetitive elements that actually align between mammals and we provide a tool, called RepeatFiller, to incorporate such repeat-overlapping alignments into genome alignments. We further show that a subset of aligning sequences detected by RepeatFiller evolve under evolutionary constraint, which uncovers previously-unknown conserved non-exonic elements and thus improves the annotation of constrained elements.
RepeatFiller incorporates several megabases of aligning repetitive sequences to genome alignments
To investigate how many aligning repetitive elements have been missed in alignments between mammalian genomes, we adopted a previously-developed approach that was initially devised to detect novel local alignments between a pair of distantly-related species [24,25]. The original approach focused on unaligning regions that are flanked by aligning blocks in co-linear alignment chains [26], which are detected in the first all-vs-all genome alignment step. In a second step, this original approach used lastz [21] with highly-sensitive seeding and (un)gapped extension parameters to align the previouslyunaligning regions again. This second round of highly-sensitive local alignment can uncover novel alignments that are co-linear with already-detected alignment blocks. Here, we adopted this approach by introducing two key changes. First, we increased alignment parameter sensitivity only slightly, but unmasked the unaligning region. This implies that all seeds, including repeat-overlapping seeds, will be considered ( Figure 1). By restricting the size of the unaligning regions to smaller regions of at most 20 kb, we reason that novel local alignments detected with a similar sensitivity level likely constitute orthologous alignments. Second, while the previous approach computed all alignment chains again from scratch using previously-detected and novel local alignments, our new approach directly adds novel alignments to existing alignment chains, thus removing the need for a chain re-computing step. This approach is called RepeatFiller and is available at https://github.com/hillerlab/GenomeAlignmentTools.
To investigate how many aligning repetitive elements can be added by RepeatFiller, we built alignment chains between the human (hg38) genome assembly and the genomes of 20 other mammals that represent the major mammalian clades (Figure 2 Table 1), and thus adds little the runtime of the initial genome-wide all-vs-all pairwise alignment step that is typically around ~1000 CPU hours. Together, this shows that a considerable portion of aligning transposon sequences are missed when repeat-overlapping seeds are ignored and that RepeatFiller can detect such alignments with little extra computational runtime.
RepeatFiller application uncovers thousands of novel repeat-derived conserved non-exonic elements Next, we investigated whether some of the newly-aligning sequences show evidence of evolutionary constraint, which indicates purifying selection and a biological function. To this end, we used the pairwise alignments, generated either with or without RepeatFiller, to build two human-referenced multiple genome alignments of 21 mammals with Multiz [27]. Then, we used PhastCons [28] to identify constrained elements. We found that the majority (98%) of the 164 Mb in the human genome that are classified as constrained in the multiple alignment without RepeatFiller were also classified as constrained in the RepeatFiller-subjected alignment.
Dividing the conserved regions detected in the alignment without RepeatFiller into exonic and non-exonic regions, we found that 99.8% of the exonic and 97.4% of the non-exonic regions are also classified as constrained in the RepeatFiller-subjected alignment. Since conserved exonic regions are virtually identical, likely because they rarely overlap repeats, we focused our comparison on the conserved non-exonic elements (CNEs), which often overlap cis-regulatory elements [29][30][31]. This comparison first showed that 3.46 Mb of the human genome were newly classified as conserved non-exonic in the RepeatFiller-subjected alignment, representing 2.9% of all conserved non-exonic bases detected in this alignment. Requiring a minimum size of 30 bp, application RepeatFiller led to the identification of 30167 novel CNEs that are listed in Supplementary Table 2.
Two striking examples of newly-identified CNEs are shown in Figures 3 and 4. Figure 3 shows the genomic region overlapping MEIS3, a homeobox transcription factor gene that synergizes with Hox genes and is required for hindbrain development and survival of pancreatic beta-cells [32-34]. By revealing novel alignments to many non-human mammals, RepeatFiller identifies several novel repeat-overlapping CNEs in introns of MEIS3 ( Figure 3). Figure 4 shows the genomic region around AUTS2, a transcriptional regulator required for neurodevelopment that is associated with human neurological disorders such as autism [35,36]. Applying RepeatFiller revealed several novel CNEs upstream of AUTS2. For some of these CNEs, RepeatFiller incorporated a well-aligning sequence of 19 mammals, which then permitted the identification of evolutionary constraint. Overall, applying RepeatFiller led the identification of more than 30000 CNEs that were not detected before.
RepeatFiller improves annotations of Conserved Non-exonic Elements
Interestingly, the comparison of conserved non-exonic bases detected by PhastCons also revealed 3.08 Mb of the human genome that were classified as conserved non-exonic only in the multiple alignment without RepeatFiller, but not in the RepeatFiller-subjected alignment. These 3.08 Mb represent 2.6% of all conserved non-exonic bases detected in the alignment without RepeatFiller. The 29334 CNEs with a size ≥30 bp are listed in Supplementary Table 3. To investigate the reasons underlying these 'lost' CNEs, we first sought to confirm that the RepeatFiller-subjected alignment had an increased species coverage in these regions. Indeed, we found that RepeatFiller added on average 3.9 (median 3) aligning species to these lost CNEs. Inspecting many of these CNEs showed that the newly added sequences are similar to the already-aligned sequences; however, they exhibit more substitutions. These substitutions increase the overall sequence divergence across mammals, which likely explains why the same region was not classified as constrained anymore, despite having a higher coverage of aligning species. To confirm that the newly-added sequences increase the overall sequence divergence, we applied GERP++ [37] to both multiple alignments (Supplementary Figure 1A). For each alignment column, GERP++ estimates the number of substitutions that were rejected by purifying selection (RS = rejected substitutions) by subtracting the number of observed substitutions from the number of substitutions expected under neutrality. Since GERP++ computes the number of substitutions expected under neutrality from a phylogenetic tree that is pruned to the aligning species (Supplementary Figure 1B), we can directly compare RS between alignment columns that were only classified as constrained in either alignment to estimate whether the RepeatFiller-added sequences evolve slower than expected under neutrality. Specifically, for each alignment column, we computed the difference in RS before and after adding new alignments with RepeatFiller, as illustrated in Supplementary Figure 1B.
We found that the alignment columns, where constraint was only detected in the alignment without RepeatFiller, mostly exhibit slightly negative RS differences ( Figure 5C, grey background), which suggests that many positions in the RepeatFiller-added sequences do not evolve under strong constraint. Hence, the extent of constraint in the more limited set of aligning sequences was likely overestimated, providing an explanation of why these genomic regions were not classified anymore as constrained across placental mammals. It should be noted that these regions may still be under constraint in particular lineages. In contrast, most alignment columns, where constraint was only detected after applying RepeatFiller, exhibit a positive RS difference ( Figure 5C, orange background), which suggests that the newly-added sequences evolve under constraint. Overall, by uncovering previously-unknown alignments, RepeatFiller application led to an improved CNE annotation.
Discussion
While transposon-derived sequences can be co-opted into a multitude of biological roles and can evolve under evolutionary constraint, comprehensively detecting alignments between ancestral transposons and other repeats is not straightforward. The main reason is that considering all repeat-overlapping alignment seeds during the initial whole genome alignment step is computationally not feasible. However, it is feasible to consider all seeds when aligning local regions that are bounded by colinear aligning blocks. We provide a tool RepeatFiller that implements this idea and incorporates newly-detected repeatoverlapping alignments into pairwise alignment chains. We tested the tool on alignments between human and 20 representative mammals and showed that with little additional computational runtime RepeatFiller uncovers between 22 and 84 Mb of previouslyundetected alignments that mostly originate from transposable elements.
We further show that RepeatFiller application enables a refined and more complete CNE annotation by two means. First, applying RepeatFiller led the identification of thousands of CNEs whose aligning sequences were not detected before. This includes highlyconserved transposon-derived CNEs that are located near important developmental genes. Second, the sequences added by RepeatFiller may not evolve slower than expected under neutral evolution. In this case, providing a more complete set of aligning sequences led to the removal of thousands of putatively-spurious CNEs that overall do not evolve under strong constraint across placental mammals, though the possibility of lineage-specific constraint remains.
Taken together, RepeatFiller implements an efficient way to improve the completeness of aligning repetitive regions in whole genome alignments, which helps annotating conserved non-exonic elements and studying transposon co-option and genome evolution.
Generating pairwise genome alignments
We used the human hg38 genome assembly as the reference genome. To compute pairwise genome alignments, we used lastz version 1.04.00 [21] and the chain/net pipeline [26] with default parameters (chainMinScore 1000, chainLinearGap loose). We used the lastz alignment parameters K = 2400, L = 3000, Y = 9400, H = 2000 and the lastz default scoring matrix. All species names and their assemblies are listed in Supplementary Table 1.
RepeatFiller
The input of RepeatFiller is a file containing co-linear chains of local alignment blocks. This file must be in the UCSC chain format as defined here https://genome.ucsc.edu/goldenPath/help/chain.html. The output is a file that contains the same chains plus the newly-added local alignment blocks. By default, RepeatFiller only considers unaligned regions in both the reference and query genome that are at least 30 bp and at most 20000 bp long. We considered all chains with the score greater than 25000. For each unaligning region that fulfills the size thresholds, RepeatFiller uses lastz with the same parameters as above but with a slightly more sensitive ungapped alignment threshold (K=2000). All repeat-masking (lower case letters) was removed before providing the local sequences to lastz. Since lastz may find multiple additional local alignments in this second step, we used axtChain [26] to obtain a 'mini chain' of local alignments for this unaligning region. RepeatFiller then inserts the aligning blocks of a newly-detected mini chain at the respective position in the original chain if the score of the mini chain is at least 5000. All default parameters for the size of unaligning regions, minimum chain scores and local alignment parameters can be changed by the user via parameters. Finally, RepeatFiller recomputes the score of the entire chain if new alignments were added.
We compared the number of aligning bases in the chains before and after applying RepeatFiller. To this end, we used the coordinates of aligning chain blocks to determine how many bases of the human hg38 assembly align (via at least one chain) to the query species. We used the RepeatMasker repeat annotation for hg38, available at the UCSC Genome Browser [38], to determine how many of the newly-added alignments overlap repetitive elements.
Generating multiple alignments
Before building multiple alignment, we filtered out low scoring chains and nets requiring a minimum score of 100000. We used Multiz-tba [27] with default parameters to generate two reference-based multiple alignments using the pairwise alignment nets produced with and without RepeatFiller, respectively.
Conservation analysis
To identify constrained elements, one needs a tree with branch lengths representing the number of substitutions per neutral site. We used four-fold degenerated codon sites based on the human ENSEMBL gene annotation to estimate the neutral branch lengths with PhyloFit [28]. To identify conserved regions, we used PhastCons [28] with the following parameters: rho=0.31; expected-length=45; target-coverage=0.3. To obtain conserved non-exonic regions, we first obtained exonic regions from the human Ensembl and RefSeq annotation (UCSC tables ensGene and refGene). As done before [25], we merged all exonic regions and added 50 bp flanks to exclude splice site proximal regions that often harbor conserved splicing regulatory elements. To obtain Conserved Nonexonic Elements (CNEs), we subtracted these exonic bases and their flanks from all conserved regions.
To compare constraint in genomic regions classified as constraint in only one alignment, we used GERP++ [37] with default parameters (acceptable false positive rate = 0.05) to estimate constraint per genomic position. We denote genomic regions as 'gained' if they were classified as constrained by PhastCons only in the multiple alignment generated with RepeatFiller. We denote genomic regions as 'lost' if they were classified as constrained only in alignment generated without RepeatFiller (Supplementary Figure 1A). Gained and lost regions were identified using 'bedtools intersect' [39]. For each position in 'gained' and 'lost' non-exonic regions, we computed the RS score (number of rejected substitutions) with GERP++ [37] and calculated the difference between the RS score obtained for the alignment with and without RepeatFiller (Supplementary Figure 1B). These differences are plotted in Figure 5C. Positive differences indicate that the sequences added by RepeatFiller evolve slower than under neutrality, thus increasing the number of rejected substitutions. Differences close to zero indicate that the newly-added sequences evolve as expected under neutral evolution and negative differences indicate that they evolve faster than expected under neutral evolution.
Data Availability
The multiple genome alignments generated with and without applying RepeatFiller and the respective PhastCons conserved elements are available at https://bds.mpicbg.de/hillerlab/RepeatFiller/. The CNEs that differ between both alignments are available in Supplementary Tables 2 and 3. The RepeatFiller source code is available at https://github.com/hillerlab/GenomeAlignmentTools.
Competing interests
The authors have no competing interests. Phylogenetic tree of human and 20 non-human mammals whose genomes we aligned to the human genome. The amount of newly alignments detected by RepeatFiller is shown in megabases and in percent relative to the human genome. Bar charts provide a breakdown of newly-added aligning sequences into overlap with transposons, simple repeats and non-repetitive sequence. T A A A T T G A T T C A C C T T A T T T A T T T C C G A T A T T T A T indicating a higher alignment identity) shows that RepeatFiller adds several aligning sequences, some of which evolve under evolutionary constraint and thus are CNEs (red boxes) only detected in the RepeatFiller-subjected alignment. The RepeatMasker annotation shows that these newly-identified CNEs overlap transposons. The zoom-in shows the 21-mammal alignment of one of the newly-identified CNEs, which overlaps a DNA transposon. While this genomic region did not align to any mammal before applying RepeatFiller, our tool identified a well-aligning sequence for 17 non-human mammals (red font). A dot represents a base that is identical to the human base, insertions are marked by vertical orange lines, and unaligning regions are showed as double lines. UCSC genome browser screenshot shows a ~1.5 Mb genomic region around AUTS2, a transcriptional regulator required for neurodevelopment. CNEs only detected in the RepeatFiller-subjected multiple alignment are marked as red tick marks. The zoom-in shows the 21-mammal alignment of one of the newly-identified CNEs. While only the rhesus macaque sequence aligned to human before applying RepeatFiller, our tool identifies a well-aligning sequence for all 19 other mammals (red font). A dot represents a base that is identical to the human base. The RepeatMasker annotation (bottom) shows that this newly-identified CNE overlaps a DNA transposon. 1 1 1 A G G T T T C T T T T A T A T G T T T A T C G G T T A T T C G T G T A C T T T T T C A C A A G T T A T A G T T A C T A G A A A T T C C T T T A T G T T T A G T T . applying RepeatFiller. Dots in these alignments represent bases that are identical to the human base, insertions are marked by vertical orange lines, and unaligning regions are showed as double lines. The alignments show that the sequences of species added by RepeatFiller (red font) exhibit a number of substitutions. This explains why these regions were not classified as constrained anymore, despite adding more aligning sequences. Please note that in (B) only the sequence of the rhesus macaque was aligned before applying RepeatFiller. Sequences in both (A) and (B) overlap LINE transposons.
(C) Difference in evolutionary constraint in non-exonic alignment columns that are only classified as constrained in either alignment. For each alignment position, we used GERP++ to compute the estimated number of substitutions rejected by purifying selection (RS). The difference in RS between alignments with and without RepeatFiller is visualized as a violin plot overlaid with a white box plot. This shows that almost all non-exonic bases that were only detected as constrained in the alignment with RepeatFiller (orange background) have a positive RS difference, indicating that the newly-aligning sequences added by RepeatFiller largely evolve under evolutionary constraint. In contrast, nonexonic bases only detected as constrained in the alignment without RepeatFiller (grey background) often have slightly negative RS differences, indicating that many of the newly-added sequences do not evolve under constraint. The two distributions are significantly different (P<e -16 , two-sided Wilcoxon rank sum test). | 4,555.6 | 2019-07-09T00:00:00.000 | [
"Computer Science",
"Biology"
] |
Nickel-Titanium Alloys Welding of Thin Sheets Using GTAW: Comparative Study Between Similar and Dissimilar Welding with AISI 304 Stainless Steel
The objective of the present work was to evaluate welding of thin sheets (thickness < 1 mm) of similar and dissimilar NiTi alloys with AISI 304 stainless steel using the Gas Tungsten Arc Welding (GTAW) process and to study the mechanical and metallurgical properties of the joints with and without post-welding heat treatment (PWHT). The GTAW process was chosen because it is more economical than the usual welding processes for NiTi alloys, such as the Laser Beam Welding (LBW). The welded joints were characterized by the techniques of SEM, OM, electrical resistance in temperature (ERT), tensile test and Vickers microhardness. It was observed that the dissimilar joints presented a brittle behavior due to the formation of brittle elements along the weld metal by the excessive increase of the hardness in this region with peaks of hardness higher than 900HV. Similar joints presented superior mechanical behavior, with extensive plastic deformation before rupture and fracture surface with ductile appearance. The PWHT in the similar joints promoted a decrease in the rigidity of the material by minimizing the thermal stresses from the welding process, the PWHT did not influence the dissimilar joints mechanical behavior.
Introduction
Shape Memory Alloys (SMA) are metal alloys for which reversible phase transformations induced by appropriate temperature and/or stress changes are observed. A phase transformation of an SMA is unique, accompanied by high data recovery capacity and stress as a function of temperature changes. Due to its ability to recover its initial strain with a temperature variation, the SMA is intrinsically considered an intelligent material as thermo-mechanical actuators applications 1,2 .
Among the most used materials in this intelligent materials field, the nickel-titanium (NiTi) SMA are well known for its superelasticity properties, shape memory effect, biocompatibility and corrosion resistance when the atomic compositions of nickel and titanium are similar, i.e. close to equiatomic 1,3 . Joining the NiTi SMA is a technological challenge due to the formation of fragile intermetallic compounds, which leads to a marked loss of tenacity from the welded joints, as during solidification cracks associated to the dendritic microstructure of the weld metal are formed. Also, there is still precipitation of deleterious phases in the heat affected zone (HAZ) and the weld metal (WM), resulting in severe reduction of mechanical resistance 4 . These phenomena are responsible for limiting the applicability of the NiTi SMA in multiple areas of interest.
With the possibility of expanding the field of the NiTi SMA applications, literature reports the joining by welding of this material with other metal alloys to explore the properties of both alloys; a common pair is to weld NiTi with stainless steel 5,6 . Stainless steel has excellent biocompatibility properties, corrosion resistance and are easily welded, specially the austenitic classes, such as the AISI 304. However the dissimilar welding of the NiTi-AISI304 pair results in a brittle structure with undesirable mechanical properties due to the formation of intermetallic elements along the weld metal 5,7,8 . An alternative way to minimize the negative effects of this welding is the use of filler metals [9][10][11] . Other ways such as promotion of post welding heat treatment (PWHT) can be further explored.
The main joining process used in welding of NiTi SMA is the Laser Beam Welding (LBW), due to its excellent precision and focusing of welding heat, resulting in a shorter HAZ 4,5 . Despite its benefits, the LBW process requires high initial investment to install the necessary infrastructure, which limits its application.
Electric arc welding processes, such as the arc welding process with non-consumable tungsten electrode and gas shielding (GTAW), are commonly used in industrial applications and are known to provide high quality welded joints and mechanical strength for most steels, aluminum, copper and their alloys. It would be possible to expect the same behavior for the NiTi alloys. However, according to authors 12,13 the GTAW welding process negatively affects the mechanical properties of welded NiTi joints due to a large extent HAZ. The presence of contaminants also limited the functional properties of the welded joints. Few papers assessing the mechanical properties of joints welded by the GTAW process have been published 14 .
It is very important to note that most studies on similar and dissimilar welding involving NiTi SMA were performed with wires 7-12,15-17 and using the LBW process. There are only a few studies on thin sheets [18][19][20] and even fewer using simpler and less expensive processes, such as the gas tungsten arc welding (GTAW).
Thus, this work aims to report the possibility of welding similar (NiTi-NiTi) and dissimilar (NiTi-AISI304) joints by the Gas Tungsten Arc Welding (GTAW) process and to evaluate the influence of the realization of post-weld heat treatments in the microstructure and in the mechanical properties, such as hardness, tensile strength and strain variation; and further analyze the fracture surfaces of the joints after welding to understand the main failure mechanism.
Materials and Methods
For the development of this work, NiTi equiatomic and AISI 304 austenitic stainless steel thin plates (thickness < 1mm) were used. The welding was carried out manually, with no purge gas used to protect the root. It was used the Digiplus A7 from IMC -SOLDAGEM, which is an electronic multiprocess welding power source. The welding parameters were fixed, after previous study, as indicated in Table 1. The welding speed used was V = 1.2 mm/s (approximate time = 15 seconds for an 18 mm bead), diameter of the gas nozzle = 4 mm, electrode diameter = 1/16 inch.
For all welding conditions the filler rod used was the NiTi alloy itself which was taken directly from the thin plates ( Figure 1). The post-weld heat treatment (PWHT) on both similar welding (NiTi-NiTi) and dissimilar welding (NiTi-AISI304) consisted of heating the joints at 500ºC for 120 minutes than furnace cooling, with no controlled atmosphere. This heat treatment was suggested by the NiTi sheet supplier in order to release the thermoelastic martensitic transformation. Joints were evaluated by electrical resistance in temperature (ERT), Vickers microhardness test, optical microscopy (OM), tensile test and scanning electron microscopy (SEM).
In order to determine the phase transformation temperatures the evaluation of the electrical resistance variation temperature (ERT) an average rate of 4°C/min, for a temperature range from -60°C to 100°C was used. The test consists of heating the joints to the reference austenitic state (at 100°C), cooling it to -60°C and then heating it to the reference state again 21 , illustrated at Figure 2. The electrical resistance variation measurements were performed according to the scheme of Figure 3. This technique was chosen in detriment of the Differential Scanning Calorimetry (DSC) as it is a nondestructive technique for the determination of the phase transformation temperatures which were calculated from the tangent method 22 . The intersection of the tangent lines denotes the temperature at which the phase transformation occurs 2,23 . Joints microstructure was analyzed using metallurgical microscope Olympus, BX-51 model. A hydrochloric acid, nitric acid and hydrofluoric acid mixture was prepared for microstructure revelation. In the Vickers microhardness (HV) tests the welded joints were tested by applying a load of 50 grams-force (gf) for 15 seconds for the similar joints and a load of 300 grams-force (gf) for 15 seconds for the dissimilar joints, due to the greater hardness presented for these. The equipment used for the test was a Future Tech microindentator, model FM-700. The spacing between the impressions was roughly 0.2mm, passing through the weld entire length starting from one side of the base metal until it crosses the weld metal and reaches the other base metal side. For the tensile test, the standard ASTM F2516-14 was used ( Figure 4). The fracture surface was analyzed using a scanning electron microscope (SEM) of the TESCAN brand, model VEGA 3 SBH.
Results and Discussion
In Figure 5-a is presented the percentage variation curves of the electrical resistance in temperature (ERT) for the NiTi base metal, as received (AR), prior to the PWHT and after PWHT. In Figure 5-b is presented the NiTi similar joints, S11 and S12, taking the temperature of 100°C as the reference state for the calculation of electrical resistance variation. The start and end temperatures of transformation were determined from the intersection of the tangent lines.
It can be inferred that there is the direct martensitic transformation during the cooling phase, whereas the reverse martensitic transformation is observable in the heating phase. The ERT curves especially that of the base metal (NiTi sheet without weld), are typical of the two-stage martensitic transformation of NiTi SMA, austenite, cubic structure, to the rhombohedral R phase with great distortion at α angle, and finally to the monoclinic martensite, forming a peak in the cooling phase that begins at Rs temperature 2,24 . The S11 joint was submitted to a welding thermal cycle and this acted as a PWHT. The S12 joint was also submitted to a welding thermal cycle although in addition it was further heat treated by the real PWHT. These heating cycles promoted the aging of the microstructure and causing a second type of martensite to precipitate in the material which is characterized by the largest peak decrease in Figure 5-b.
According to Table 2 it is still possible to observe a slight increase in the martensitic transformation start temperature (Rs) when welding these alloys by the GTAW process. Joint S12 presented an increase of about 4°C relative to joint S11 and increase of 8°C relative to the base metal (reference). These increases probably occur due to the impoverishment of nickel in the NiTi matrix one since Ni 4 Ti 3 precipitation can occur, causing an increase in the phase transformation temperatures 12 .
Vickers microhardness profile of the evaluated joints is shown in Figure 6. The 0 mm mark was considered the center of the weld metal. The weld metal had dimensions of approximately 4mm. The positive values of distance are directed to the base metal AISI 304 (for the dissimilar joints), whereas the negatives to the base metal NiTi. It was noticed that it was not possible to determine the extent of the heat affected zone (HAZ).
This microhardness profile observed in Figure 6 for NiTi similar joints (S11 and S12) exhibited a more homogeneous microhardness profile, presenting small increase of hardness in the possible HAZ and towards the weld metal, mainly when contrasted to the behavior observed in the dissimilar joints (D11 and D12).
The S12 (430 ± 68 HV) welding condition presented higher hardness than the S11 (353 ± 45 HV) welding condition because it was subjected to PWHT, the hardness increases observed after the PWHT is linked to phase precipitation such as Ni 3 Ti 4 25 . It can be stated that the qualitative microhardness profile is identical to those reported by 12,26-28 . There was no partially diluted zone in the similar joints, so there was no such great increase in microhardness at the weld interface, as can be observed in Figure 7 for the S12 joint. Also, it is possible to observe the competitive grain growth resulting in a coarse microstructure of columnar grains. Columnar grain size is a function of the original base metal grain size, welding energy and joint geometry, thus impacting on the joint mechanical behavior 29 . Furthermore, it can be stated that the PWHT do not result directly on weld metal grain refine, with the possibility to degrade this properties due to phase formations.
For the dissimilar joints it was possible to note a considerably hardness increase in the weld metal (WM) region, decreasing towards both base metals, to the two sides of the joints, reaching expected values of HV hardness for both base metals, i.e. around 220 HV for NiTi and 170 HV for AISI 304.
Also, the existence of high hardness peaks, 800-1100 HV, for joints D11 and D12 were mainly concentrated towards the weld metal region. These are related to the intermetallic formation during welding, due to the large number of different alloying elements present in the steel matrix (Cr and Ni) in addition to Fe and C, and also nickel and titanium for the NiTi SMA 9,19 . The heat treatment used was not effective in the minimization and appearance of these peaks. Those elements lead to the presence of a partially diluted zone (PDZ), as can be seen on Figure 8.
The qualitative behavior of the hardness profiles observed in Figure 6 was also reported in the literature 8,17 . The authors reported that high hardness peaks (800-1400 HV) were observed in the melt zone extension of the NiTi-AISI304 dissimilar joint, associated with the precipitation of intermetallic as TiFe 2 , FeTi, TiCr 2 , TiC, TiNi 3 .
EDS analyzes in the PDZ region (D11 joint) is presented in Figure 9. The spectrum 1 is located at the AISI304 base metal side, the spectrum 2 is located at the PDZ and the spectrum 3 is located at the edge of the partial diluted zone, in the WM (Figure 9). Chemical composition reported by spectrum 1 indicated that only Mn (1.63 %wt), Ni (7.89 %wt), Cr (19.64 %wt) and Fe (70.84 %wt) are presented, mainly AISI 304 chemical composition. The spectrum 2 indicated that in the partial diluted zone nickel (16.13 %wt) and titanium (26.45 %wt) are also presented. Iron and chromium are presented in 44.11 %wt and 12.51%wt. Finally, the spectrum 3 indicated that in the weld metal, Ni (49.30 %wt) and Ti (37.28 %wt) are the majority, Fe (11.51 %wt) and Cr (1.92 %wt), the minority. This chemical composition variation is the mainly variable responsible for the PDZ formation and the microhardness behavior previously discussed.
What is observed in Figure 10 is that the dissimilar joints ruptured prematurely when reaching UTS of 100 MPa, with final deformation inferior than 0.5%. NiTi-AISI304 joints' UTS values are close to those found in studies in the literature, 6. Microhardness profile of the welded joints, the AISI 304 base metal is located at the right side for the dissimilar joints. Figure 7. S12 joint microstructure observed through OM, interface between BM and WM.
mechanical properties for the dissimilar NiTi-stainless steel joints [9][10][11] . The use of intermetallic layers is a viable approach for obtaining NiTi-stainless steel joints welded by the laser beam welding (LBW) process due to its precise control.
Other approaches, such as performing a post-weld heat treatment, may be further explored. Distinct behavior is presented by the similar joints that presented an extensive area of plastic deformation before failing, reaching UTS of ~650MPa and elongation of 16% to the condition without heat treatment and ~550MPa and elongation of 23% to the thermally treated joint, mechanical behavior comparable to the NiTi base metal used (UTS of 750MPa and elongation of 21%). It is also possible to observe that the heat treatment of the S12 joint promoted a decrease in the rigidity of the material by minimizing the thermal stresses from the welding process. This also promoted a final deformation increase. Therefore the PWHT could be a viable tool to manage the NiTi-NiTi rigidity and enhance its ductility. This mechanical behavior result for similar NiTi joints can be compared with those reported in the literature: Melo et al. 22 using factorial experiment were able to achieve NiTi-NiTi welded joints by laser with UTS of 682 MPa as (UTS = 134MPa) 11 and (UTS ~200MPa) 8 for welding using the LBW process. This behavior is directly related to the observation of regions of high hardness in the weld metal. During the plastic deformation the dislocations are accumulated in the regions of high tension and have their restricted mobility which culminates with the premature fracture. One reported approach for minimizing the negative effects of such difficulties in welding is the use of metallic interlayers between dissimilar metals. The use of these layers aims to mitigate these effects and to promote desirable ; e -S11 and f -S12. and elongation of 12.7%, Wang et al. 30 welding NiTi thin sheets obtained UTS of 450 MPa and elongation of 5.1% using the LBW process, Mirshekari et al. 8 achieved UTS of 835 MPa using the LBW and NiTi wires and Gugel et al 17 achieved a NiTi welded joint UTS of about 70% of its corresponding base metal. Thus, it is also possible to reach satisfactory mechanical behavior of NiTi-NiTi joints using the GTAW process.
Most recently, Oliveira et al. 14 presented a comparison of the different joining techniques for the NiTi shape memory alloys. For the NiTi similar welding mostly problems are related to formation of brittle intermetallics, compositional variations and precipitation. LBW is the most studied process for the NiTi welding due to minimal thermal effect on the joint. For the NiTi dissimilar welding the authors state that there is still limited research on this topic. Figure 11 (a-f) present the SEM images of the fracture regions for the dissimilar joints D11 (a-b) and D12 (c-d), and similar joints (e-f). Joints D11 and D12 have failed at the interface between the weld metal and the base metal. This characteristic is common for welded joints that fail due to brittle fracture, usually due to the presence of welding defects and/or the formation of intermetallic compounds precipitated in this region, as evidenced by the observed microhardness profile. It is possible to notice that the main mechanism of fracture was ductile, with the presence of dimples in the similar joints, S11 and S12. In addition, the presence of parabolic dimples (Figure 11-f) was observed when submitted to PWHT, whereas in the non-PWHT condition, dimples of smaller size were observed (Figure 11-e).
In the dissimilar joints, the micro aspect of these surfaces indicates that the mechanism of fracture was predominantly by cleavage. It is possible to observe planes of different heights along the fracture surfaces, characterizing the type of fragile fracture 8 . This fracture is fragile and of transgranular origin, since fracture cracks pass through the grains, presenting a faceted texture, as a result of changes in orientation of the cleavage planes from one grain to another. Most intermetallics lead to brittle fracture, with little crack tolerance. The complex structures of these compounds have large Burgers vectors, thus requiring less energy for fracture than for displacement during deformation.
Conclusion
In this work a comparative study was performed on the mechanical behavior of similar and dissimilar nickel-titanium joints welded by the micro GTAW process and evaluating the influence of post-weld heat treatment on these properties. It was possible to achieve good appearance weld beads with no oxidation or porosity. The Vickers microhardness test allowed the identification of high hardness peaks of the order of 1000HV in the weld metal for the dissimilar joints due to the formation of fragile intermetallic in this region. A slightly hardness increase was observed in the HAZ and WM of the similar joints. It was possible to notice that the dissimilar welded joints failed under UTS of 100 MPa exhibiting brittle fracture profile, by cleavage. Similar welded joints exhibited distinct mechanical behavior with extensive plastic deformation prior to failure. The post-weld heat treatment in the similar joints promoted a decrease in the rigidity of the material by promoting the relief of thermal stresses from the welding process. It was possible to achieve satisfactory mechanical behavior with similar NiTi welded joints using the GTAW process. | 4,520.6 | 2018-11-08T00:00:00.000 | [
"Materials Science"
] |
Modulation of polar mesospheric summer echoes (PMSEs) with high-frequency heating during low solar illumination
. Polar mesospheric summer echo (PMSE) formation is linked to charged dust/ice particles in the meso-sphere. We investigate the modulation of PMSEs with radio waves based on measurements with EISCAT VHF radar and EISCAT heating facility during low solar illumination. The measurements were made in August 2018 and 2020 around 20:02 UT. Heating was operated in cycles with intervals of 48 s on and 168 s off. More than half of the observed heating cycles show a PMSE modulation with a decrease in PMSE when the heater is on and an increase when it is switched off again. The PMSE often increases beyond its initial strength. Less than half of the observed modulations have such an overshoot. The overshoots are small or nonexistent at strong PMSE, and they are not observed when the iono-sphere is influenced
Introduction
Polar mesospheric summer echoes (PMSEs) are strong, coherent radar echoes observed from altitudes of 80 to 90 km at high and middle latitudes during the summer.It was first noted in the 1970s that these coherent radar echoes were unusually strong (Ecklund and Balsley, 1981;Czechowsky et al., 1979) and that they originate from the height of the extreme temperature minimum around the mesopause that occurs at high and middle latitudes in the summer months (Ecklund and Balsley, 1981).Later, the echoes were observed from various locations using radars with frequencies ranging from 50 MHz-1.3GHz (Cho and Röttger, 1997).The PMSE is observed from mid-May to the end of August in the Northern Hemisphere, with the main occurrence during local noon (Latteck et al., 2021).
The observed reflection of the radio waves results from strong variations in the electron density and, thus, the refractive index.The echoes are strong as the backscattered radio waves interfere constructively when the distance between the scattering centers is half the radar wavelength, called the Bragg condition.Scattering at the Bragg condition Published by Copernicus Publications on behalf of the European Geosciences Union.
is typically caused by neutral turbulence in the atmosphere.PMSEs arise from a combination of neutral turbulence and the presence of charged ice particles that form near the cold mesopause and influence the electron distribution; the presence of these ice particles expands the Bragg scales for which the echoes are observed (Rapp and Lübken, 2004).The spatial distribution of the ice particles at these altitudes is influenced by the complex neutral atmosphere dynamics caused by the upward-propagating gravity waves.It can also be seen in the structure of noctilucent clouds (NLCs) (Dalin et al., 2004).
The region of PMSE occurrence overlaps with that of NLCs, which is an optical manifestation of these ice particles.Temperature studies of the summer Arctic mesosphere suggest that both phenomena are temperature controlled and occur at temperatures of 150 K and lower around the mesopause (Lübken, 1999), where water ice particles can form.Since 2007, water ice particles have also been observed by satellites in so-called polar mesospheric clouds (PMCs); the optical properties of water ice explain the measured cloud extinctions with inclusions of smaller meteoric smoke particles (Hervig et al., 2012).The meteoric smoke particles are nanometer-sized dust particles that form from ablated meteoric material in the altitude range 70-110 km (Rosinski and Snow, 1961;Hunten et al., 1980;Megner et al., 2006).The satellite observations also support the existing hypothesis that the ice particles are formed by heterogeneous condensation, which has recently been supported by a study that applies a new theoretical condensation model (Tanaka et al., 2022).The surface charging of dust particles, be it meteorite smoke, ice particles, or a mixture of both, is a necessary process that influences the growth of ice particles and, at the same time, gives clues to their size and composition (Rapp and Thomas, 2006).The dust can, for example, become negatively charged from electron attachment in the PMSE altitude range.This is indicated by rocket measurements of so-called electron "bite-outs" (depletion in electron density), where PMSE is present (Rapp and Lübken, 2004, and references therein).
Previous studies have shown that the modulation of PM-SEs during artificial heating with high-frequency (HF) radio waves could be used to study the underlying plasma and dust particles (Biebricher et al., 2006;Mahmoudian et al., 2011Mahmoudian et al., , 2020)).During such heating experiments, the electron temperature is locally and temporarily enhanced (Rietveld et al., 1993); Chilson et al. (2000) first noticed that PM-SEs can be modulated during such heating.The PMSE often almost disappears when the heater is turned on and then returns when the heater is turned off again.It is assumed that the increased electron temperature during heating and the resulting increased diffusion reduces the fluctuations in the electron density and thus the PMSE power (Rapp and Lübken, 2000).Havnes (2004) found that with an adequate on/off time of the heater, a so-called overshoot characteristic curve could be generated, in which the PMSE power did not return to the original value after heating but exceeded it.Such overshoot curves have been observed in many simultaneous radar and heating studies of PMSE made with EISCAT.The overshoot curves have also been observed for some polar mesospheric winter echoes (PMWEs) (Kavanagh et al., 2006;Belova et al., 2008;Havnes et al., 2011).Most PMWEs do not appear to be associated with the presence of dust (Latteck et al., 2021).Still, those showing overshoots are more likely related to the presence of small dust particles, possibly meteoric smoke.
With this work, we want to investigate whether and how the PMSE modulation during heating can be used for systematic investigations of the charged dust component.We present observational studies of PMSE with the EISCAT VHF radar during four VHF/heating campaigns, which are all done in August during twilight or night conditions.This is the first systematic investigation of PMSE modulation under reduced sunlight conditions and toward the end of the PMSE season.
The remaining part of the paper is structured as follows.First, Sect. 2 introduces the PMSE modulation during heating and the overshoot effect.Section 3 describes the experiments we performed, including the radar and heating parameters, and gives an overview of the observational results.Then a discussion of the PMSE modulation is given in Sect.4, where we first discuss the cases of quiet ionospheric conditions and of an ionosphere that is moderately influenced by energetic particle precipitation; we then give an overview of the observed PMSE modulation.We make a comparison with a model calculation and discuss the overall outcome.A short conclusion is given in Sect.5, and additional information on observational data is provided in Appendix A and the Supplement.
PMSE and heating
The EISCAT heating facility transmits high-frequency radio waves of high power into the atmosphere (Rietveld et al., 1993).Electron oscillations associated with wave absorption translate into thermal motion, heating the electron component while the other plasma components keep their initial temperature.As mentioned above, it was found that this active heating influences the PMSE signal.During the experiments, the heating is switched on and off in pre-defined time intervals (48 s on and 168 s off).The PMSEs are simultaneously observed with the EISCAT radar.The time variation of the observed PMSE power is sketched in Fig. 1 to illustrate the observed phases of the PMSE heating cycle and the often seen overshoot curve: decline, heating phase, recovery/overshoot, and relaxation.
The amplitudes (R 0 , R 1 , R 2 , R 3 , and R 4 ) marked in Fig. 1 will be considered in our analysis of the observations below, where R 4 is then the start (R 0 ) of the next subsequent cycle.We follow previous studies (e.g.Havnes et al., 2015) and refer to the curves that describe the measured PMSE during one heating cycle (on and off time) as overshoot curves.
Decline
As the heater is switched on at R 0 , the power effectively falls off instantaneously (depending on the radar frequency used) (Havnes, 2004).The backscattered power drops as the heating enhances the electron temperature and, consequently, the electron diffusivity so that the large electron density gradients are reduced.Therefore, the backscatter is less efficient (Rapp and Lübken, 2000).
Heating
R 1 → R 2 : During the heater-on phase from R 1 and R 2 , there are some variations in the power amplitude.Because of the higher electron temperature, the charging electron flux on the dust particles increases during the heater-on period, and often an increase in the power can be seen.The charging timescales become shorter and compete more with the faster electron diffusion (Mahmoudian et al., 2011).
Recovery/overshoot
R 2 → R 3 : The power then increases when the heater is switched off (recovery), and in many cases, the power rises above the previous undisturbed level (overshoot).The electron temperature drops quickly to the initial value before the heater is on due to the highly collisional regime at these altitudes.The dust particles carry a higher charge than before and repel the electrons more strongly.The electrons follow the ion diffusivity, and as a result, the electron density gradients become larger.This causes the backscatter to be larger, creating an overshoot in power.
Relaxation
R 3 → R 4 : Now the power relaxes back to the previous undisturbed level, with a varying relaxation time depending on the conditions.With a long relaxation time, new and undisturbed plasma can enter the radar beam, or the dust present has time to discharge (Havnes, 2004).
Observations
We first describe the overall observation conditions, radar operations, and radar analysis, and then we present an overview of the data.
Overall observation conditions
The presented observations were carried out during the "Mesoclouds 2018" and "Mesoclouds 2020" campaigns in collaboration with UiT Tromsø and IRF Kiruna.The EIS-CAT VHF radar and the EISCAT heating facility are located in Ramfjord near Tromsø, Norway (69.59 • N, 19.23 • E).The observations were made on 11/12 and 15/16 August 2018 and 5/6 and 6/7 August 2020, during the night between 20:00 and 02:00 UT.These observations thus represent dusk and night conditions with reduced influence of sunlight on the observational volume compared to other observations, mainly done around noon in June and July.
The solar zenith angles during the observations are in the range of 88-97 • .PMSEs at 80-90 km altitude are still sunlit but to a lesser extent for most of the previous PMSE observations.To estimate the difference, we compare the solar illumination at the time of our 15 August (2018) observations to those of the summer solstice in the same year.We derive the solar UV flux by calculating the absorption of the solar UV flux by O 2 along its path through the atmosphere (described by Giono et al., 2018).We use solar Lyman-α line (121.56nm) flux from the SOLSTICE instrument on the SORCE satellite (https://lasp.colorado.edu/home/sorce/data/ssi-data/, last access: 27 February 2020) and O 2 densities from the NRLMSISE-00 atmosphere model (Hedin, 1991) for the location of the EISCAT VHF radar.We estimate that the solar photon flux in August at PMSE altitudes is reduced by at least 1 order of magnitude compared to noon conditions in June, as seen in Fig. 2.This translates to a reduced photoemission current by an order of magnitude.It thus influences the dust-charging conditions since the photoemission current is proportional to the photon flux (Mahmoudian et al., 2018).
Simultaneous optical measurements of NLCs were done using two cameras located at Kiruna and Nikkaluokta, Sweden (about 200 km south of Tromsø).There was, however, no NLC observation above the radar site, mainly because of weather conditions.During the night of 15/16 August, faint NLCs were observed from Kiruna close to the horizon, approximately above Andøya, i.e., more westward than the EISCAT site.Figure 3a (blue line) as measured by the Aura satellite and frost-point temperature profiles (green line) estimated using the Aura water vapor data (both the temperature and water vapor were measured with the Microwave Limb Sounder (MLS) instrument).The height ranges in which the temperature is lower than the frost-point temperature indicate the regions where ice particles can form.This gives a good indication of the conditions present in the atmosphere, showing that the temperatures are cold enough to facilitate ice particle formation at PMSE altitudes.However, there could be variations due to the spatial and temporal differences between the measurements that must be kept in mind.These measurement points were the closest in time and space to the PMSE observations; the horizontal distance to Tromsø is about 490 km in Fig. 3a and about 293 km in Fig. 3b.
Radar operation and data analysis
The radar observations were made in the zenith direction with the EISCAT VHF (224 MHz) antennas near Tromsø (69.59 • N, 19.23 • E).The radar code used was Manda, and reference to EISCAT documentation (Tjulin, 2017) and radar and heating system parameters are given in Table 1.The EIS-CAT heating facility (Rietveld et al., 1993(Rietveld et al., , 2016) ) was operated with a vertical beam at 5.423 MHz with a nominal 80 kW per transmitter, which corresponds to effective radiated power (ERP) in the range between 500 and 580 MW, and X-mode polarization was used with a sequence of 48 s on and 168 s off.The vertical extension of the heater beam extends far beyond the region covered by the radar.Given that the vertical winds and velocity fluctuations of the PMSE observed with EISCAT VHF are within a few meters per second and horizontal winds possibly a few tens of meters per second (Strelnikova and Rapp, 2011), the radar at all times measures PMSEs that are influenced by the heating.
A standard incoherent scatter analysis, GUISDAP (Lehtinen and Huuskonen, 1996), was used to derive the radar data products.It provides the electron density derived from the incoherent scatter spectrum assuming that the electron and ion temperatures are equal (which they are not when the heater is on).The backscatter cross section is proportional to 1/(1 + T e /T i + α 2 ) as is shown by Pinedo et al. (2014), indicating that when the heater is turned on, T e increases and consequently the backscattered power decreases.The actual electron density is assumed to be not affected, so we use the unit of equivalent electron density as was done previously for observations of polar mesospheric winter echoes (PMWEs) (Kavanagh et al., 2006;Belova et al., 2008) and PMSEs (Mann et al., 2016).The post-experiment integration time used throughout this analysis was 24 s for computational reasons except for one of the observations when we compare with simulations.A resolution of 4.8 s was used.We found that choosing a higher time resolution for the overall discussion did not result in additional information.
Overview of observations
The observations were made from 20:00 to 02:00 UT on four nights in August 2018 and 2020.The observations are displayed in Fig. 4 and shown for the entire period with altitudes from 80-110 km, hence including PMSE and the conditions of the surrounding ionosphere.White vertical areas are observation gaps due to operational problems.We identified interesting measurement intervals in each data set we considered for analysis.A closer look at each area is given in the Supplement, and an overview of the time and altitude range of the areas is shown in Table A1 in the Appendix A.
Observation 1: the 11/12 August 2018
PMSE was observed until around 01:30 UT.One can see that the electron densities above and partly below the PMSE are high, showing the typical appearance of particle precipitation.In area 1, the precipitation is strong, and enhanced electron density was observed as low as 80 km, well below the PMSE layer.We considered the following.
-Area 1: PMSE with strong precipitation in the altitude range 83.4-85.6 km from 21:36 UT, lasting about 20 heater cycles; -Area 2: high-altitude and long-lived PMSE layer extending from 86.3-90 km during about 40 heater cycles, starting from 23:06 UT with some precipitation; -Area 3: low-altitude PMSE at 83.4-86.4km from 00:00 UT lasting about 30 heater cycles with some precipitation at the end of the layer.PMSE was observed before midnight and then again at 02:00 UT. at the end of the measurements.The first observed PMSE (area 1) seems to be not influenced by precipitation.
The PMSE observed later (areas 2 and 3) are influenced by moderate precipitation.Modulation is seen in the backscattered power of the lightly ionized portion of the ionosphere from 90-110 km, which can be seen around 20:00-21:00 UT.
We considered the following.
3.6 Observation 3: the 5/6 August 2020 PMSE was observed only before midnight.Some observations (areas 1 and 2) show no apparent influence of precipitation.Before the start of area 1, there is PMSE present.However, this is not included in the analysis due to (most likely) direct interference from the heater caused by arcing, which can be seen as vertical lines extending through all altitudes.
For completeness, we also consider area 3, which displays a layered structure and is influenced by the heating.The height and the shape suggest, however, that this is not PMSE but rather a sporadic E layer.We considered the following.
- 3.7 Observation 4: the 6/7 August 2020 From the fourth observation, we see a low-altitude PMSE layer only slightly influenced by precipitation, a second layer at high altitude influenced by heating that also might be a sporadic E layer, and a third area extending over a long period in time and many altitudes that do not seem to be influenced by particle precipitation.We considered the following.
-Area 1: a long interval of PMSE between 81-88 km partly in the quiet ionosphere and partly influenced by precipitation; -Area 2: sporadic E layer above PMSE height; not included in analysis; -Area 3: a weak PMSE with little apparent precipitation for about 1 h from 21:30 UT at 82-88 km.
We find, in general, that the overshoot effect disappears in the presence of strong or moderate precipitation, as seen in the 15/16 August 2018 observation in Fig. 4.This is better illustrated in the figures given in the Supplement, where each area is enhanced.At the beginning of the observation campaign on 15/16 August 2018 (area 1), a weak PMSE developed under very quiet ionospheric conditions.The echoes are only weakly enhanced in comparison to surrounding areas, the backscattered power is reduced during heating, and an overshoot is also observed (see Fig. S4).
Observed PMSE modulation
First, we discuss two selected cases, one with little or no particle precipitation and one with moderate precipitation.Then we summarize the heating effect and overshoots visible in all the observations, and we discuss these findings in the context of previous observations.Finally, we compare a selected case with simulations of the overshoot cycle and discuss what information we can gain from modulating PMSE with heating.
PMSE modulation under quiet ionospheric conditions
To discuss PMSE modulation under quiet ionospheric conditions, we chose an area with no apparent energetic particle precipitation; we consider area 2 from the 5/6 August 2020 observation (Fig. 4c).The overshoot curves can be assessed using the overall power plot shown in Fig. 5.The beginnings of new heating cycles are marked with dashed lines when the heater is turned on.The dotted line indicates the time when the heater is turned off again.In many cases, the PMSE signal changes at the heater on and off times and during the cycles themselves.The PMSE layer lies within the altitude range of 83-87 km with a maximum extension of 2 km at its widest.
There are clear indications of reduced PMSE power when the heater is on; in many cases, we can see clear overshoots.In Fig. 6, we have selected two altitude sections for a closer look, altitude 85.2 and 85.6 km, where we can see overshoots in many of the cycles.In general, the overshoots are relatively large, with some an order of magnitude larger than the pre-heater value and with some showing no apparent increase in the PMSE power after the heater is turned off.This seems especially true for the top altitude where the PMSE power is at its highest, the lower height has a somewhat lesser PMSE power, and more overshoots are visible.The decline is visible in many of the cycles and is very strong for cycles 40-47.One can also see that characteristics of decline and overshoot often change between adjacent heating cycles and height intervals, e.g., in heating cycle 41.
For a closer investigation, we describe the ratios of the amplitudes during the different phases of the heating cycle.The different power amplitudes are marked in the overall sketch given in Fig. 1.The different amplitudes observed during the heating cycles are plotted in Fig. 7, where the amplitude ratios are considered.We find that during most heating cycles, the signal drops when the heating is switched on (decline R 1 < R 0 , Fig. 7a). https://doi.org/10.5194/angeo-41-93-2023 Ann. Geophys., 41, 93-114, 2023 We assume that the observed signals are PMSE when R 0 > 10 10.5 (which corresponds to around 3.16 × 10 10 ; Ullah et al., 2019), and one can see that in most cases that do not meet this requirement; there is no PMSE modulation visible.However, as we will see later, this condition removes a few instances of low-power modulated PMSE with large overshoots.The same can be said for the green points that show a decline but are below the threshold.They could be showing a decline but also be noise due to random fluctuations from the two measurement points.
The ratio of the amplitudes R 0 and R 3 describes an overshoot (R 0 < R 3 ), and this comparison shows that overshoots and undershoots are equally abundant, independent of the signal strength (Fig. 7b).Comparing the signals at the beginning of subsequent cycles (Fig. 7c) shows no trend and a broad range of values which suggests variation either due to ionospheric conditions or due to neutral turbulence (rather than dust).
One can see in Fig. 7d that for strong signals the amplitude stays constant or decreases slightly during the heateron phase.The change in amplitude during the heating can indicate the charging process of the dust particles, where the faster timescale of diffusion or dust charging dominates (Mahmoudian et al., 2011).According to Havnes et al. (2015), large PMSE structures can cause the diffusion timescale to be longer and, consequently, have a quicker and larger increase in power during the heater-on phase.The comparison of R 2 and R 3 in Fig. 7e describes to what extent the signal increases again when the heater is switched off.This increase is seen in most cases except for the small am-plitudes, which might be either low-power PMSE or random fluctuations.
Finally, in Fig. 7f, the ratio of R 3 and R 4 describes the signal after the heater is switched off (relaxation).One can see a broad scatter symmetrically around the diagonal, indicating that the natural variations in the PMSE power are dominant.Any relaxation after heating is difficult to discern from this since their contribution could disappear due to a significant background increase in PMSE power.This is due to the considerable period between the two points (168 s), which according to Havnes (2004), is enough time for the ionosphere to change or dust to become discharged, whereas 48 s used for the on time is not.
We compile these results in histograms of the amplitude ratios (Fig. 8).The histograms contain only cycles with a value R 0 > 10 10.5 of the PMSE amplitude before the heater is turned off to only include those with PMSE and exclude the cycles that contain noise or are mostly noise.We only include those cycles that show a decline due to heating in all the histograms.In Fig. 8a, we see that 55 % of the ratios are smaller than 1 and thus show a decline (affected by the heater) and that the average value of those ratios that are below 1 is 0.72.This is a reduction of 28 % of the pre-heater value on average when the heater is turned on.We have the overshoot in Fig. 8b.Only 10 % of the cycles show an overshoot with an average value of 0.44.Even though there are not many overshoots for this observation, those observed show an average reduction of more than half, indicating very large overshoots.Figure 8c shows that most (95 %) of the observations show a decrease in power while the heater is on.Figure 8d shows that 66 % of the cycles show an increase in power when the heater turns off, which is similar to the number of cycles that show a power reduction when the heater is turned on.Then in Fig. 8e, we see a general increase in power from cycle to cycle.Thus a general decrease to pre-heater value cannot be determined, most likely due to increasing background PMSE dominating the signal and the histogram, where 87 % of the cycles show an increase in power in subsequent cycles.This can be related to why we see so few overshoots in this observation, and that increase in PMSE power is significant for many of the cycles.The overshoot disappears due to background variations.
PMSE modulation during moderate particle precipitation
Conditions with moderate particle precipitation are observed in area 2 of the observation from 15/16 August 2018 (see Fig. 4b).The overall power plot is shown in Fig. 9.As discussed above, some heating intervals have noticeably very https://doi.org/10.5194/angeo-41-93-2023Ann.Geophys., 41, 93-114, 2023 Only overshoot curves with a minimal background amplitude of R 0 > 10 10.5 are considered.The ratios are chosen so that if we observe an overshoot curve like shown in Fig. 1, all ratios are smaller than 1.Thus, the histograms are clipped at a maximum ratio of 3. The green line and the corresponding number display the mean for all ratios smaller than 1.
strong overshoots (14,15,16,17).One can note that the influence of the heating is most pronounced at the beginning and the very end of the observation interval when there is no apparent particle precipitation.Precipitation occurs in cycles 18 and 19 and then in cycles 24 and 25.When the heater is switched on, there is no reduction in power, and the precipitation dominates the received signal for all altitudes in these cycles.The power plot for two selected height intervals shown in Fig. 10 shows this in more detail, where the modulation entirely disappears in the cycles influenced by precipitation.This is to be expected and has been shown before.One of the reasons why the modulation disappears in the PMSE layer is that the atmosphere below the layer is ionized due to the strong precipitation, and the HF radio wave might be strongly absorbed before it reaches the PMSE layer and thus not be strong enough to heat the electrons appreciably in the layer.The different amplitudes observed during the heating cycles in this area are plotted in Fig. 11.We find that during most heating cycles, the signal drops when the heating is switched on (decline, Fig. 11a).The cases that show no decline are spread over all amplitudes, indicating the cycles that might be influenced by precipitation and thus might show an increase in power when the heater is on.The overshoots and undershoots are equally abundant independent of the signal strength (Fig. 11b).As observed in the area discussed above, there is no trend when comparing the signals at the beginning of subsequent cycles (Fig. 11c).The change of amplitude during the heating (Fig. 11d) is small for most observations.
In most cases, the amplitude increases (Fig. 11e) when the heater is switched off, similar to the heated cycles, which is to be expected.Finally, in Fig. 11f, the ratio of R 3 and R 4 describes the relaxation, showing a large spread around the diagonal with somewhat more observations showing a reduction.This large spread can be attributed to the ionospheric variability due to the large timescale of the off time, as was mentioned previously.
The histograms of the power amplitudes are shown in Fig. 12 with the same criterion as before (also given in the figure text).Here the overshoot is seen in 55 % of the cycles with an average of 0.75 decline ratio (Fig. 12a), similar to the previous observation.Here the overshoot is seen in 31 % of the observations with an average of 0.64 overshoot ratio (Fig. 12b), which is more than the previous observation, even with precipitation.Similar to the previous observation, when the PMSE power increases (and is not influenced by precipitation), we see an influence of the heater but not an overshoot (or a minimal overshoot).For the cycles with a lower PMSE power (like in cycle 15), the overshoot is large, but the background PMSE power is lower; thus, the overshoot is easy to see.During the heating, there seems to be a general decrease in the values, with 76 % of the values showing a decrease during the heater-on phase (Fig. 12c).The recovery (Fig. 12d) ratio shows that 58 % has an increase in power when the heater is turned off, showing similar values to those for when the heater is turned on (decline).Then there seems to be a little over half of the cycles that show a general increase in pre-heater values between cycles (Fig. 12e).
Overall observational discussion
Here we summarize, in Table 2, the decline and overshoot ratios for all the observations (see Figs. S27-S36 in the Supplement for reference).In general, the heating effect is seen in more than half of the heating cycles for each respective area, with most of the average ratios showing values close to 0.75.These calculations show only the observations with a value of R 0 > 10 10.5 to indicate the presence of PMSE and exclude noisy data.
This, however, causes the faintest PMSE to be excluded from the histograms, as is seen for the overshoot ratio for area 1 from 15 August 2018; here, the PMSE power is below the threshold.Thus no cycles are included in the calculation despite 100 % of the cycles showing a decline due to heating.This would suggest manually inspecting low-power PMSE influenced by heating would be a better option or introducing other criteria to include these.
To summarize, we see only overshoots in less than half of the cycles, with many cycles often more influenced by background ionospheric conditions that might overshadow the heating of the PMSE.Ullah et al. (2019) show a more significant occurrence of overshoots in their observations, with around 40 %-70 %, where their observations were during daytime.Havnes et al. (2015) observations had a much larger ratio of cycles with decline present and a slightly higher percentage of overshoots present.
However, in our case, we see a few instances where the overshoot in some cycles is unusually large.Myrvang et al. (2021) found that a higher electron temperature due to heating could be achieved during nighttime compared to daytime, which might help explain some of these large overshoots.However, Kassa et al. (2005) found for their observations that the heating temperature effect observed increased for the observation with the most amount of sunlight (near noon).
Other possible reasons for unusually large overshoots could be a change in the PMSE/NLC season, as is noted by Latteck et al. (2021), when the season is getting longer.Since our observations are in reduced sunlight and close to the end of the season, more varying background conditions might influence our observations than those during the day in June/July.
Comparison of a selected observation to simulation
Here we take a closer look at the approximately 1 h time interval, which is marked as area 2 in the observation from 15-16 of August 2018, shown in Fig. 13; the data cover the heating cycles 12 to 27 and range over seven height intervals of around 360 m each.The ionosphere is influenced by precipitation in cycles 18 and 19 and then again in cycles 24 and 25, and there are no overshoots present in those heating cycles, as mentioned before.The PMSE in intervals marked with A, B, and C in the figure shows a decrease when the heater is on and overshoots when the heater is turned off.Interval A shows relatively low PMSE power but quite high overshoot curves compared to intervals B and C, as we will investigate further.
Individual heating cycles are shown in Fig. 14a for both altitudes from interval A, with PMSE power and measurement error provided by the EISCAT GUISDAP analysis.The corresponding average overshoot cycle for the respective altitude is shown on the right in Fig. 14b; in blue is the correhttps://doi.org/10.5194/angeo-41-93-2023 Ann. Geophys., 41, 93-114, 2023 sponding average overshoot cycle for the respective altitude.
As can be seen, the overshoot is relatively strong for many of the heating cycles, especially the strong overshoot seen in cycle 15 for both altitudes with relatively high but decreasing overshoot on both sides of the cycle.Note the two yaxis scales for the different altitudes, where the heating cycles from altitude 88 km have such a low background PMSE power that the scale is an order of magnitude lower than the altitude below.Both altitudes have relatively low background PMSE power compared to intervals B and C, with the PMSE at 88 km altitude barely present or the irregularities on the limit of being seen by the VHF radar.It is thus interesting to find such large overshoot cycles for this particular interval.Individual heating cycles from intervals B and C are shown in Fig. 15a with their corresponding altitude average on the right-hand side in blue (Fig. 15b; note that here the y-axis scale is the same for all the altitude ranges).They cover heating cycles 21, 22, and 23.The overshoots are present for the lower altitudes but are not as high as in interval A. However, the overshoot does not decline evenly but increases again before reaching the initial signal level.This influence can be seen in the averaged heating curve for altitude 86.7 km, where after about 120 s, the power starts to increase again.This is either because of the beginning influence of particle precipitation on the ionosphere or variation of the PMSE structure due to the long relaxation time (Havnes et al., 2015).This influence is very strong in the subsequent cycle (cycle 24), where the PMSE power increases during the heater-on period.This type of ionospheric variation can influence the observations to the extent that heating effects are less visible.In the same time interval (intervals 21, 22, 23) at the altitude above, the overshoots are small, es- pecially for the first cycle (21), while the PMSE power is relatively low.This is in contrast to the observation made at the higher altitude in interval A where a significant overshoot is observed at low PMSE power.This might indicate that there are different conditions at play for these two cases.Havnes et al. (2015) has mentioned that higher altitudes of PMSE reside in more turbulent conditions, thus a more significant variation in cloud structure and a longer relaxation time after heater turn-off time as a result.
A comparison of the average overshoot curves for each interval (A, B, and C) is shown in Fig. 16a and their corresponding normalized average curves in panel (b).The values are normalized to the initial PMSE power taken as the average of the last five values (24 s) before the heater is turned on.This is chosen to have sufficient data when some meahttps://doi.org/10.5194/angeo-41-93-2023Ann.Geophys., 41, 93-114, 2023 Only overshoot curves with a minimal background amplitude of R 0 > 10 10.5 are considered.The ratios are chosen so that if we observe an overshoot curve like shown in Fig. 1, all ratios are smaller than 1.Thus, the histograms are clipped at a maximum ratio of 3. The green line and the corresponding number display the mean for all ratios smaller than 1.
surement points are missing and to better compare to the rest of the data used in this article which are at a resolution of 24 s.Data were normalized after averaging the cycles from each interval.We can see that the highest normalized overshoot (b) is the one from interval A, which has the lowest background PMSE power (a) and that the lowest normalized overshoot is from interval C, which has the corresponding highest PMSE background power.This high PMSE power is possibly due to an onset of precipitation which becomes apparent in the subsequent cycle 24 right after intervals B and C. -S36) in the Supplement) for the decline (R 1 / R 0 ) and the overshoot (R 0 / R 3 ) ratio when they are smaller than 1 (indicating heating effect and overshoot) for all four observations.These numbers only include observations with minimum background amplitude R 0 > 10 10.5 .A1 refers to area 1 for that observation's date and so forth.We compare these selected overshoot curves to a computational model initially developed at Virginia Tech.It treats the plasma as a fluid including an arbitrary number of charged particles, neutral particles, and dust particles; the dust charging is described in the orbital-motion-limited (OML) approach (see, e.g., Scales and Mahmoudian, 2016).The model's parameters include the electron diffusion time scale, the charging time scale, and the time evolution of electron and ion densities.The dust charging causes electron density depletion, and the amplitude of electron density fluctuations determines the radar backscattered amplitude.The simulations assume an initial plasma temperature of T i = 150 K and a background electron density of 2×10 9 m −3 .Which fits well with the same parameters derived from the IRI model ( 2016) for the time and date of the observation.The simula-tion also assumes a reduced photoemission rate used in the charging equations in line with the experiments being done for conditions with low photoemission.
The resulting simulated overshoot curves are shown in Fig. 17b and for comparison are the averaged and normalized observations from intervals A, B, and C (marked in the same color and symbol as previous figures) shown on the left.The simulations best fit to the observed overshoot curves for 3 nm dust particles.However, there is little difference for similar sizes of dust (e.g., 3-4 nm).This result fits well with the altitude range we measure the observed PMSEs since, in general, we can assume to find smaller particles of dust at higher altitudes (however subject to neutral air movement) as well as the fact that there were no NLCs observed and thus the particles were not optically visible (larger > 20 nm). https://doi.org/10.5194/angeo-41-93-2023 Ann. Geophys., 41, 93-114, 2023 The normalized and averaged data from interval A has a higher overshoot than the simulations can produce, where the simulation has an overshoot of around 8.4.At the same time, the observations show an overshoot of almost 9.9.The timescale of the simulation for interval A runs for 300 s, while the observation has a much quicker equalization toward the "background" PMSE value/undisturbed plasma values.For the simulation to reach such a high overshoot, the ratio between dust and electron number density is only at 35 %, and with a heating ratio increase for electron temperature of 8 times the pre-heater value.This would indicate that the dust density is lower than for the other two intervals and that the heating effect is consequently larger.As discussed later, the electrons gain a higher temperature, and charging onto the dust particles is, therefore, more effective, where some dust particles can gain more than a single charge.
A comparison of observations for intervals B and C and their corresponding simulations show a better agreement where the overshoot and relaxation are very similar.For these overshoots to be produced in the simulation, the ratio of dust to electrons must be higher, with 60 % for interval B and 68 % for interval C. The increase at the end of the relaxation period for both intervals is not reproduced in the simulations; this is assumed to be due to the influence of the precipitation that occurs clearly in cycle 24 and is already increasing the background PMSE power in the previous cycles.Compared to the observations, the simulated signals drop slower during the heater-on phase and rise more slowly to the overshoot when the heater is switched off again.The measured response of the PMSE to the heating is instantaneous within the 4.8 s resolution of the data.A possible explanation for this difference is that the numerical model might have miss- ing parameters or processes to simulate this increase.This is in contrast to the decrease we see in most observations, as was discussed previously.
On the left-hand side in Fig. 18a, we can see the average charge number found for the simulation for each respective interval (marked in the figure).For interval A the average charge number reaches a maximum of about 1.38 charges per dust particle during the heater-on phase.This indicates that to achieve such a high overshoot, the charging efficiency of the dust particles needs to be high and that (due to high electron heating temperature) many dust particles will gain more than one negative charge during the heating cycle.Note the longer timescale shown in the simulation for interval A (300 s), indicating that it takes longer for the overall average charge on the dust particles to equalize back to pre-heater values.As was mentioned before, the dust population is much lower for interval A compared to the other two intervals since the ratio of dust to electrons is lower.Consequently, the significant increase in temperature (by a factor of 8) causes a larger average charge number on the dust particles during the heater-on phase.For the other intervals (B and C), the maximum average charge number is less than one during the on phase of the heater for both cases, with interval B being around 0.9 charges per dust.For interval C, the average dust charge lies at about 0.86.This corresponds well with the observed and simulated overshoot curves from Fig. 17, where the higher overshoot is observed in interval B. Thus the average charge number is consequently higher.So the effective charging of the dust during the heater-on phase for these intervals is less than for interval A, and a smaller overshoot is observed.
On the right-hand side in Fig. 18b, we have the ratio of the diffusion time to the charging time scales for each respective interval.Here we can see the variation between the two timescales and how this changes during the heating cycle.For all the intervals, there is an increase in the ratio when the heater turns on, a relaxation during the heater-on period, a sharp increase when the heater is turned off, and a slow decrease during the heater-off period.The significant increase Table 3. Neutral density for each interval from NRLMSISE-00 atmosphere model (Hedin, 1991) taken at 21:00 UT and the estimated ion-neutral collision frequency (see Ieda, 2020;Cho et al., 1998). in the heater-on time could be understood as the charging timescale becoming smaller with increased electron charging onto dust particles due to the increased electron temperature.This corresponds well with the increased average electron charge on the dust particles seen in Fig. 18a.Here the average dust charge is highest for interval A, and the ratio of timescales is also highest for this interval, which might indicate a faster charging timescale for that interval than for the other two.The increase at heater turn-off time is also due to a decrease in the charging times; more dust is being charged now by the ion portion of the plasma, which drags the electrons along and causes the observed overshoot.Thus for interval A the simulation of the overshoot curve fits best with a lower ratio of dust particles to electron density.Therefore we might argue that there is more plasma than in the other two intervals.This larger plasma population might then charge the dust more quickly, causing a smaller charging timescale and, consequently, a larger overshoot in interval A.
Interval Neutral density (m
Another difference could arise in the diffusion timescales in the respective intervals.The diffusion timescale is proportional to the ion-neutral collision frequency, which decreases with decreasing neutral density.Hence in interval A at a higher altitude and with lower neutral density, the diffusion timescale can be shorter than in the other interval (Havnes et al., 2015).The estimated ion-neutral collision frequencies are given in Table 3, which are derived using neutral density from the NRLMSISE-00 atmosphere model (Hedin, 1991) The timescale that is the fastest is the dominating one.So when the heater turns on, the diffusion timescale might be lower for interval A. So when the heater is turned on, the diffusion timescale decreases even more due to its dependence on the temperature ratio (T e /T i ), and we expect/need a more significant temperature increase for the electrons in interval A to explain such a large overshoot.As the heater is turned on, the charging timescale decreases due to the increase in electron temperature.A larger charging effect is seen in the interval A simulation (average charge number) compared to the other intervals.Consequently, a larger overshoot is seen.
So to summarize, the decreased diffusion timescale for interval A due to reduced neutral density and the significant increase in electron temperature combined help explain the large overshoot seen for interval A. The higher electron temperature could be explained by greater absorption of the heater's energy in the interval.According to Havnes et al. (2015), the amount of electron density per altitude will determine where the heater's energy is absorbed and how much.This generally causes lower altitudes of PMSE to become more heated than higher altitudes.Interval A is at a higher altitude than the other two intervals.Still, the precipitation present in cycle 18 before intervals B and C could cause the altitude regions below these intervals to have a higher electron content and, thus more absorption of the heater's energy below.
Discussion and conclusion
For the presented observations, we find that artificial heating affects the PMSE signals during less than half of all the observed heating cycles with a pre-heated PMSE power R 0 > 10 10.5 ; the average reduction of the power is about 25 % from the pre-heated value.The cutoff, R 0 > 10 10.5 , excludes cycles that do not show PMSE and cycles being highly influenced by noise.With this criterion, we covered most of the PMSE.However, some very faint ones were excluded, and some were affected by heating and showed large overshoots.We find that the heating has little effect on PMSE during ionospheric conditions with particle precipitation which other authors also see.This is especially so for strong and moderate particle precipitation.We assume that under these conditions of higher ionization, the heating waves are mainly absorbed in lower altitudes, thus not causing a heating effect in the PMSE layer.Often the background ionospheric conditions strongly influence the PMSE profile during one heater cycle, and it is thus challenging to derive a correct relaxation time, which would be an interesting parameter because it depends on the dust conditions present in the layer.
As to the shape of the PMSE modulation curves, the variation of the PMSE during the heater-on period (from R 1 to R 2 ) is affected by two competing processes: the charging and the diffusion.For the presented observations, most heating cycles display a signal decrease from R 1 to R 2 .Less than half of the cycles influenced by heating show an overshoot when the heater is turned off.However, observed overshoots are generally high and, in some cases, very high.These high overshoots could be attributed to the dust charge in the presented observations being more strongly influenced by heating, as the influence of photoemission is smaller than during daytime observations.
It is also possible that the size of ice particles and their formation and sublimation rates are different toward the end of the PMSE season; most other heating studies were carried out earlier in the year.A general trend toward a more extended PMSE season (Latteck et al., 2021) and larger particles at PMSE altitudes (at high latitudes) due to increased water vapor content (Lübken et al., 2021) could also cause these recent PMSE observations to show different modulation curves.
The computational overshoot model we considered cannot account for some of the high overshoot cases we observed, and we are unaware of a model that does so.Some processes might need to be included to reproduce these cases of large overshoots.The influence of variation in the ionospheric background with time over the cycles reduces the overshoots and dominates the relaxation phase.We form, however, averaged curves as was done in other studies and compare those to the model calculations.We find that simulations with dust size around 3 nm best fit to all cases considered.
While different electron heating ratios and dust-to-electron densities are needed to match the observational data, a larger temperature heating ratio and a lower dust density are required to best match the large average overshoot observed.The amount of absorption from the heater's energy is impor-tant in how effectively the electrons can be heated.And since there is precipitation between the first interval with large overshoots and the two other intervals, it stands to reason that the altitudes below the PMSE layer have increased electron content after moderate precipitation.This causes a larger absorption of the heater's energy below the PMSE layer.Therefore a combination of decreased heater energy and lower diffusion time can help explain the large overshoot in the first interval.
We conclude that the presented observations during HF heating confirm that high-power radio waves modulate PMSE amplitudes, with the observed modulation varying on short spatial and temporal scales.The presented observations differ from previous studies since they are done late in the PMSE season and during lower solar illumination (dusk/night).In general, we see both an influence of the heating and an overshoot in about half of the heating cycles, which is somewhat lower than previous observations done earlier in the season around midday.We see very high overshoots compared to previous observations and note that increased PMSE power is connected to smaller overshoots.
Figure 1 .
Figure 1.Sketch of the PMSE modulation due to HF heating in a typical overshoot curve; the power amplitudes during different times of the heating cycle are defined.
Figure 2 .
Figure 2.Estimated photon flux for the Lyman-α line for 21 June at 12:00 h (UT) and 15 August at 22:00 and 24:00 h (UT) at 85 km altitude.Solar zenith angles used in the estimation included in the label are from the International Reference Ionosphere (IRI) model (2016).
Figure 3 .
Figure 3. Temperature profiles (blue line) as measured by the Aura satellite at 12 and 16 August 2018 and frost-point temperature profiles (green line) estimated using the Aura water vapor data.Latitude and longitude of the measurement points are given in the figures by φ and λ respectively.
Figure 4 .
Figure 4. Overview of all four observation days with time intervals and dates given in each respective figure.
Figure 5 .
Figure 5. Backscattered power as a function of altitude and heating intervals observed during the night of 5/6 August 2020, in area 2.
Figure 6 .
Figure 6.Backscattered power at altitude 85.2 km (b) and 85.6 km (a) and heating intervals observed during the night of 5/6 August 2020 in area 2. The color of the dots follows the color scale of Fig. 5.
Figure 8 .
Figure 8.Average of (a) decline, (b) overshoot, (c) heating, (d) recovery, and (e) relaxation for the observed data on 5 August 2020 in area 2.Only overshoot curves with a minimal background amplitude of R 0 > 10 10.5 are considered.The ratios are chosen so that if we observe an overshoot curve like shown in Fig.1, all ratios are smaller than 1.Thus, the histograms are clipped at a maximum ratio of 3. The green line and the corresponding number display the mean for all ratios smaller than 1.
Figure 9 .
Figure 9. Backscattered power as a function of altitude and heating intervals observed during the night of 15/16 August 2018, in area 2.
Figure 10 .
Figure 10.Backscattered power at altitude 87.4 km (b) and 87.8 km (a) and heating intervals observed during the night of 15/16 August 2018 in area 2. The color of the dots follows the color scale of Fig. 9.
Figure 11 .
Figure 11.Comparison of the power amplitudes observed on the 15 August 2018 in area 2.
Figure 12 .
Figure 12.Average of (a) decline, (b) overshoot, (c) heating, (d) recovery, (e) relaxation for the observed data on 15 August 2018 in area 2.Only overshoot curves with a minimal background amplitude of R 0 > 10 10.5 are considered.The ratios are chosen so that if we observe an overshoot curve like shown in Fig.1, all ratios are smaller than 1.Thus, the histograms are clipped at a maximum ratio of 3. The green line and the corresponding number display the mean for all ratios smaller than 1.
Figure 13 .
Figure 13.Overview of Area 2 -15 of August 2018, with interesting visible overshoot cycles marked with intervals A, B, and C. Data resolution is 4.8 s.Cycles are marked in the figure (from 12 to 27) as well as their corresponding On/OFF period.
Figure 14 .
Figure14.Individual overshoot curves (a) from interval A (from Fig.13) shown with their corresponding altitude average on the right-hand side (b).Heating cycle numbers are shown at the bottom, and the on-and-off period for the averaged cycles is also shown.Note that the y-axis scale for altitude 88 km is an order of magnitude smaller than for the altitude 87.7 km.
Figure 15 .
Figure 15.Individual overshoot curves (a) from intervals B and C (from Fig. 13) shown with their corresponding altitude average on the right-hand side (b).Heating cycle numbers are shown at the bottom, and the on-and-off period for the averaged cycles is also shown.
Figure 16 .
Figure 16.Average overshoot curves for each respective interval (a) and normalized average overshoot curves (b) for the same intervals.They are normalized with the average of the last five values before the heater was turned on.
Figure 17 .
Figure 17.Comparison of the averaged and normalized heating cycles for each interval (a) to its corresponding simulation of the overshoot cycles (b).Note the longer timescale of simulation of interval A (longer time needed for simulation to return to equilibrium).
Figure 18 .
Figure 18.Simulations of average dust charge number on (a) for each respective interval and the ratio between the diffusion time and the charging time scales for the same intervals on (b).
Table 1 .
Parameters for EISCAT VHF radar operation and EISCAT heating facility.Half of the VHF antenna is used for transmitting, and the entire antenna is used for receiving (beamwidth adjusted accordingly). | 12,577.6 | 2023-02-14T00:00:00.000 | [
"Physics",
"Environmental Science"
] |
Low efficacy of recombinant SV40 in Ugt1a1-/- mice with severe inherited hyperbilirubinemia
In contrast to AAV, Simian Virus 40 (rSV40) not inducing neutralizing antibodies (NAbs) allowing re-treatment seems a promising vector for neonatal treatment of inherited liver disorders. Several studies have reported efficacy of rSV40 in animal models for inherited liver diseases. In all studies the ubiquitous endogenous early promoter controlled transgene expression establishing expression in all transduced tissues. Restricting this expression to the target tissues reduces the risk of immune response to the therapeutic gene. In this study a liver specific rSV40 vector was generated by inserting a hepatocyte specific promoter. This increased the specificity of the expression of hUGT1A1 in vitro. However, in vivo the efficacy of rSV40 appeared too low to demonstrate tissue specificity while increasing the vector dose was not possible because of toxicity. In contrast to earlier studies, neutralizing antibodies were induced. Overall, the lack of a platform to produce high titered and pure rSV40 particles and the induction of NAbs, renders it a poor candidate for in vivo gene therapy.
Introduction
Crigler-Najjar syndrome (CNs), severe unconjugated hyperbilirubinemia, results from the deficiency of UGT1A1, the enzyme that catalyzes the conjugation of unconjugated bilirubin (UCB) with UDP-glucuronic-acid [1]. The conjugation of the hydrophobic UCB results in water soluble bilirubin glucuronides that can be excreted into bile [2]. If not treated effectively the severe form of CNs is lethal in childhood due to UCB accumulation to levels that cause irreversible brain damage [3]. Both lethality and brain damage can be prevented by intensive phototherapy, a cumbersome treatment that becomes less effective overtime [4,5]. Most patients therefore do need a liver transplant at some point in their life, a highly invasive treatment with several challenges, like the need for re-transplantation, toxicities and adverse effects associated with long-term immunosuppression [6]. In addition, because of the limited availability of donor organs the patients are at risk to develop brain damage while on the waiting list. Novel therapies are therefore warranted, and recent clinical studies for other liver diseases, like hemophilia B show the potential of liver directed gene therapy [7]. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 The efficacy of several methods for liver directed gene therapy has been investigated for CNs in UGT1A1 deficient rodent models. Non-viral gene therapy approaches using lipophilic nanoparticles do show potential for delivery of mRNA resulting in effective but transient reduction of serum bilirubin levels in a Ugt1a1 deficient rat [8]. For delivery of DNA the efficiency of this non-viral method still is much lower compared to viral vectors. In addition, several viral vectors, such as rSV40 [9], Adenoviral [10], Lentiviral [11] and Adeno Associated Viral (AAV) [12] vectors and transposons [13] have been tested in the rat model, while only AAV vectors were tested in Ugt1a1 deficient mice [14][15][16]. Of all these viral vectors, AAV vectors are the most advanced and are now tested in clinical trials (NCT03466463 and NCT03223194).
A major challenge for AAV mediated liver directed gene therapy is the presence of preexisting Neutralizing Antibodies (NAbs) in a significant percentage of the CN patients [17]. These NAbs will block hepatocyte transduction hampering effective treatment efficacy. The induction of a high titer of NAbs upon the first AAV administration blocking re-treatment with this vector, is another important hurdle [18]. Re-treating a patient may be required upon loss of correction due to liver growth, when treating juveniles, or due to drug or alcohol induced liver damage, or in patients receiving a sub-optimal vector dose such as those participating in the safety and efficacy studies [19].
In this respect, rSV40 vectors with a low pre-existing immune prevalence seem a promising option [20,21]. Also, the reported absence of a cellular response and absence of neutralizing antibodies upon repeated SV40 administration, render this vector a promising candidate for liver directed gene therapy [22,23]. The recently developed novel production cell line ensures production of batches that are free of large T antigen, a prerequisite for clinical application of this vector [24].
A potential problem of clinical use of rSV40 vectors is the ubiquitous nature of the endogenous SV40 early promoter. This promoter has been used widely in expression studies and is suitable to provide expression of a transgene in many different cell types [25,26]. For in vivo application, this ubiquitous nature is a disadvantage because expression of the therapeutic UGT1A1 protein, for instance in antigen presenting cells, could increase the risk of an adaptive immune response. Restricting the expression of UGT1A1 to the hepatocytes will reduce this risk significantly [12].
In this study an rSV40 vector with a liver specific promotor to drive the expression of a reporter gene and the therapeutic hUGT1A1 gene was developed and its specificity, efficacy and immunogenicity was tested in vitro and subsequently in vivo in a Ugt1a1 deficient mouse model.
Production of viral vectors
To generate liver specific rSV40 vectors a hybrid liver specific promoter (HLP) [27,28] was inserted between the endogenous SV40 early promoter and the luciferase or hUGT1A1 cDNA, using the ClaI and SpeI sites present in Pam310 from AMARNA [24] and rSV-hUGT1A1 [29]. rSV-Luc and rSV-HLP-Luc vectors were produced in using co-transfection with a Cre-recombinase expressing plasmid as reported previously [29] or by removal of the bacteria backbone with Not-I, gel purification and re-ligation [24]. 3 and 6 days after transfection, the medium was collected and viral vector was concentrated using a spin filter (100 KD, Lot R9NA92290, Merck Millipore Ltd, Ireland). The rSV-hUGT1A1 and rSV-HLP-hUGT1A1 vectors used in vitro and in vivo studies were produced in Super-Vero cells by AMARNA Therapeutics (Leiden, The Netherlands) as described [24]. 10 4 cells per well were seeded in Costar white 96 well plates (ref 3610, corning incorporated Kennebunk, ME, USA). The next day cells were transfected using 40 ng of plasmid and 80 ng of PEI/well or transduced with rSV40 vector (400 vg/cell). Luciferase expression was determined 48 hours after transfection or 2-5 days after transduction by adding 100 μl lysis buffer with substrate (ONE-Glo™ Luciferase, Promega), and after a 15 minutes incubation at room temperature, luminescence was measured in the synergy HT, measurement time was 1 second/well. UGT1A1 expression was determined using western blotting at 2 days after transfection or 2-5 days after transduction. The cells were washed once with PBS and lysed with RIPA (50 mM Tris PH 8.0, 150 mM NaCl, 1% Triton X-100, 0.5% Na-deoxycholate, 0.1% SDS) buffer containing Protease-Inhibitor (1:100 dilution) (Roche, Germany) for 20 minutes. 30 μg of cell lysate was loaded on a 10% Acrylamide gel and blotted to PVDF membrane (semi-dry blotting, 1 hour, 0.05 mA per gel). A monoclonal antibody towards UGT1A1 (1:700 dilution) followed by a goat anti-mouse HRP labeled (Dakoplast, the Netherlands) (1: 5,000) was used to detect UGT1A1 as described previously [12]. An anti-UGT1 rabbit polyclonal antibody (Santa Cruz Biotechnology, Santa Cruz, CA) [28] was used to detect UGT1A1 expression in liver of Ugt1a1 deficient mice treated with rSV40 and for comparison, treated with rAAV8 and in wild type mice.
Animal study
FVB/NJ background Ugt1a -/mice were treated with phototherapy (PT) until weaning and housed in a temperature-controlled environment with 12/12-hour light-dark cycle, with a standard diet and water ad libitum in the animal house unit [30]. The Ugt1a -/mice have a one nt deletion in exon 4 of the Ugt1a locus causing a shift in the reading frame introducing a stop codon immediately after the deletion. Due to the absence of the co-substrate binding and transmembrane domains the truncated enzyme is inactive, as published previously [14]. All animal protocols were approved by the Animal Welfare and Ethics Committee of the International Centre for Genetic Engineering and Biotechnology (ICGEB), Trieste, Italy. At 60 days after birth, the mice were randomized and a single administration of rSV-hUGT1A1 (n = 7) or rSV-HLP-hUGT1A1 (n = 8) or a PBS injection control (n = 7) by tail vein injection, 3 mice with dose of 2x10 11 vg/kg, and 4-5 mice with dose of 1.7x10 12 vg/kg, respectively. Intravenous of rSV40 was performed under isoflurane anesthesia. No surgical or other procedures requiring anesthesia were performed. Animals were treated by researchers trained for animal care and monitored daily for general appearance. No signs of stress (lack of appetite, body weight loss, hair loss, stereotypic or aggressive behavior, etc.) or macroscopic alterations of vital functions were observed throughout the experimental phase. Blood was sampled at 1, 3, 5 and 8 weeks after vector administration by facial vein puncture and collected in EDTA-containing tubes. At the time of sacrifice mice were anesthetized with 5% isoflurane, cardiac puncture was performed to obtain blood samples, and sacrificed by cervical dislocation. Plasma was obtained by centrifuging at 400 g for 15 min at room temperature, and stored at -80˚C for further analysis. Total bilirubin determination in plasma was performed following the instructions of the supplier (BQ KITS) as previously described [30]. Organs were directly frozen in liquid nitrogen and stored at -80˚C for further analysis.
Determination of rSV40 neutralizing antibodies
SV40 NAbs titer in mice plasma were determined using the protocol reported for determination of the NAbs titer towards AAV, using serial dilution of plasma in FCS after complement inactivation by incubation at 56˚C for 30 minutes [31]. Briefly, 10 4 COS-1 cells/well were seeded in costar white 96 well plates, in Dulbecco's modified Eagle's medium (DMEM) (Lonza/Westurg) containing 10% fetal calf serum (FCS), 100 U/ml penicillin/streptomycin (Invitrogen) and 20 mM L-Glutamine (Lonza/Westurg) and cultured overnight at 37˚C in 5% CO 2 . The next day plasma samples were diluted in 7 steps on a half-log scale in FCS. Diluted samples were incubated with an equal volume of 2.5x10 9 vg/ml rSV-Luc at 37˚C for one hour. The maximal transduction control, set to 100%, consisted of rSV-Luc incubated with FCS only, the negative control consisted of FCS diluted with medium; both were incubated under the same conditions. After this pre-incubation, 7.5 μl per well of each dilution was added to the COS-1 cells in triplicate. Positive and negative controls were performed in six-fold. The cells were incubated at 37˚C, 5% CO 2 , and after two days the luciferase expression was determined using ONE-Glo™ Luciferase assay system (Promega), according to manufacturer's protocol.
Anti-UGT1A1 antibody assay
An indirect ELISA approach was used to detect the anti-UGT1A1 IgG in the mice serum as previously described [32]. Per well of a 96 ELISA plate, 50 μl of recombinant human UGT1A1 protein (Bio Connect) (1 μg/ml) in coating buffer (15 mM Na 2 CO 3, 35 mM NaHCO 3 ) was added and incubated overnight at 4˚C. The next day the UGT1A1 coating solution was replaced with 100 μl of blocking buffer, 1% gelatine in phosphate-buffered saline (PBS). After 1 hour, the blocking was removed and 50 μl serial dilutions of heparin plasma samples diluted in washing buffer were added to the wells and incubated for 1,5 hours at room temperature. After 3 washings with washing buffer 0.05% Tween-20 in PBS, UGT1A1 binding mouse immunoglobulins were detected with horseradish peroxidase (HRP) conjugated anti-mouse IgG 1:1000 dilution in conjugation buffer containing 4/5 blocking buffer and 1/5 washing buffer followed by o-phenylenediamine (Sigma) conversion.
Statistical analysis
Statistical significance was determined using GraphPad Prism 8.3.0 (GraphPad Software, La Jolla, CA).
Insertion of a liver-specific promoter in rSV40 increases transcriptional activity in liver-derived cells
To restrict the expression provided by an SV40 vector to the liver, a liver specific promoter was inserted between the SV40 early promoter and the luciferase gene generating the rSV-HLP-Luc plasmid. In both human hepatoma cell lines, Huh7 and HepG2, transfection of the rSV-HLP-Luc plasmid provided a higher luciferase expression than the construct containing only the endogenous SV40 promoter, rSV-Luc (Fig 1A and 1B). In non-hepatoma cell lines, COS-1 and HEK293T, the latter was clearly more active (Fig 1C and 1D). This indicates that the liver-specific promoter is functional and significantly increases transcriptional activity in liver-derived cells. To investigate if the tissue-specific promoter remains functional in the viral vector, both plasmids were used to produce recombinant SV40 vector. The hepatoma and non-hepatoma cell lines were transduced using 400 vg/cell of rSV40, and luciferase expression was determined at 2-5 days after transduction. In both hepatoma cell lines transduced with the rSV-HLP-Luc vector, the luciferase intensity was significantly higher than in cells transduced with the rSV-Luc (Fig 2A and 2B). In contrast, in both non-hepatoma cell lines, the rSV-Luc vector provided a higher luciferase intensity (Fig 2C and 2D).
To confirm that insertion of a liver-specific promoter in SV40 gene delivery vector also can improve expression of a therapeutic gene in hepatoma cells, rSV-hUGT1A1 and rSV-HLP-hUGT1A1 vectors were generated and used to transfect human hepatoma and non-hepatoma cell lines. The UGT1A1 expression was measured by western blot at 48 hours after transfection and demonstrated that presence of the HLP resulted in a significant increase in HepG2 cells (Fig 3A and 3B). In Huh7 cells, the difference in UGT1A1 expression between both vectors did not reach significance due to large variation. Again in non-hepatoma cell line HEK293T cells, the expression of HLP-driven UGT1A1 was significantly lower compared to the cells transfected with rSV-hUGT1A1. This increased transcriptional activity in liver-derived cells induced by presence of the HLP promoter was confirmed upon transduction of these cell lines with the rSV vector generated from these plasmids (Fig 3C and 3D).
In vivo, the poor efficacy of rSV vectors is not improved by the insertion of a liver specific promoter
Upon showing that a liver specific promoter enhances transgene expression in hepatocytederived cells, compared to other cell types, the efficacy of both vectors was studied in vivo. 2x10 11 vg/kg rSV-hUGT1A1 or rSV-HLP-hUGT1A1 were injected in the tail vein of 60-dayold Ugt1a -/mice. Upon injection the effect on serum bilirubin was monitored overtime. This resulted in at some time points significant, albeit a minor and far from therapeutic, decrease of serum bilirubin in mice injected with rSV-hUGT1A1 (Fig 4). In the mice treated with the liver specific rSV-HLP-hUGT1A1, the serum bilirubin levels were comparable to that in untreated controls, indicating that inserting the hepatocyte specific promoter does not improve in vivo efficacy of this vector. A possible explanation for the apparent lower in vivo efficacy of the rSV-HLP-hUGT1A1 compared to the rSV-hUGT1A1 vector could be a lack of liver tropism of the SV40 vectors, since the latter would result in transduction of many tissues upon injection into a peripheral vein. The bio-distribution of SV40, determined by the ratio between the human UGT1A1 gene carried by the vector and the mouse β-actin gene, showed presence of rSV vector genomes in several tissues. Although in several animals these levels were below detection, the bio-distribution results did not support a specific liver tropism for rSV40 upon injection into a peripheral vein (Table 1). Subsequently, in a higher dose of 1.7x10 12 vg/kg rSV-hUGT1A1 and of rSV-HLP-hUGT1A1 was tested (Fig 4). At the time of injection, the serum bilirubin levels in this experiment were lower due to the natural variation occurring in this model. Administration of the higher vector dose did not result in therapeutic correction albeit that after one week the levels in the mice receiving rSV-HLP-hUGT1A1 did have lower serum level. The hUGT1A1 expression in liver was determined by western blot showing signals in mice treated with rSV-hUGT1A1 or rSV-HLP-hUGT1A1 comparable to that in PBS treated Ugt1a1 deficient mice (Fig 4D). In contrast administration of a 10 fold lower dose of AAV8 vector, containing hUGT1A1 behind the same liver specific promoter (AAV-L), did show presence of hUGT1A1 while treatment with a comparable dose of this vector (AAV-H), showed a very prominent presence of UGT1A1. Two of the mice injected with rSV-hUGT1A1 died within 24 hours after the injection.
Intravenous injection of rSV40 does induce a neutralizing antibody response
Absence of a humoral response towards rSV40 would allow repeated administrations of this vector. Retreatment in case of loss of efficacy or upon sub-optimal dosing would be a major advance for rSV40 compared to AAV and Adenoviral vectors. To determine the presence of neutralizing antibodies COS-1 cells were transduced with rSV-Luc vector pre-incubated with serum from naïve mice or mice injected with rSV40. After 48 hours transduction efficiency was determined by measuring the luciferase expression in COS-1 cells. Pre-incubation with serum from mice injected with rSV40 did impair luciferase expression, indicating that intravenous administration of rSV40 does induce neutralizing antibodies towards this vector in vivo. Comparing to the mice before administration, the anti-SV40 neutralizing antibody titer increased in rSV40 treated mice but not in PBS treated ones (Fig 5). No humoral response towards the encoded hUGT1A1 was detectable irrespective of the presence of the hepatocytespecific promoter restricting hUGT1A1 expression to the liver or to several tissues when its expression was controlled by the ubiquitous SV40 early promoter (S1 Fig). Fig 3. Insertion of a liver specific promoter in rSV40 provides specific expression of a therapeutic gene. (A) 10 5 cells of HepG2, Huh7, COS-1, or HEK293T cells were transfected with 1 μg of rSV-hUGT1A1 or, rSV-HLP-hUGT1A. 48 hours later, UGT1A1 expression was determined by western blotting and quantified normalizing to calnexin, a membrane protein (B). (C) 10 5 cells of Huh7 and COS-1 cells were transduced with 400 vg/cell of rSV-hUGT1A or rSV-HLP-hUGT1A1. After 2-5 days the cells were lysed and UGT1A1 expression was detected by western blotting and quantified by western blotting normalized to calnexin. All experiments were repeated 3-5 times. � p <0.05, �� p <0.01, ��� p<0.001, ���� p<0.0001. Data represent the mean ± SD. Statistical significance was determined by twotailed Student t test. https://doi.org/10.1371/journal.pone.0250605.g003
Discussion
The data reported in this study show that insertion of a liver-specific promoter results in more specific expression of transgenes encoded by rSV40 vectors in vitro. The insertion of a Presence of UGT1A1 in Ugt1a1 deficient mice treated with 1x10 11 vg/kg of AAV8-hUGT1A1 (AAV L), PBS, 1.7x10 12 vg/kg of rSV-hUGT1A1, 1.7x10 12 vg/kg of rSV-HLP-hUGT1A1, or 1x10 12 vg/kg of AAV8-hUGT1A1 (AAV H), and in a non-treated wild type mouse (WT). Calnexin was used as loading control on the same membrane. Data represent the mean ± SD. Statistical significance was determined by Two-way ANOVA with Dunnett's multiple comparison test.
https://doi.org/10.1371/journal.pone.0250605.g004 hepatocyte specific promoter did not improve the efficacy of the rSV40 vector in vivo. A dose of 1.7x10 12 vg/kg of these vectors did not result in a relevant reduction of serum bilirubin levels. Intravenous administration of both rSV40 vectors did result in the induction of neutralizing antibodies in all animals, rendering effective retreatment with this vector most unlikely.
In vivo gene therapy using rSV40 vector has been reported for unconjugated hyperbilirubinemia [9]. A replication-deficient rSV40 mediated a long-term amelioration of jaundice in Gunn rats. In that study, serum bilirubin levels were lowered by 35% upon a single injection, while in this study the reduction using a similar rSV40 vector in a UGT1A1 deficient mouse model is only about 10%. The different administration route, peripheral vein versus portal vein, may explain the difference in correction since upon peripheral vein injection no clear liver tropism for rSV40 was seen. The bio-distribution showed transduction of many tissues by rSV40 albeit all at a low level ( Table 1). The lack of a liver tropism can also explain the better performance of the non-hepatocyte specific rSV40 vector since UGT1A1 expression in nonhepatic tissues also provides effective conjugation of UCB [15,33,34]. Administration into a peripheral vein of AAV8, a vector that does display a clear liver tropism, provides a much more efficient correction. A dose of 1x10 11 vg/kg results in therapeutic correction of serum bilirubin levels [28] and in contrast to both rSV40 vectors, in clearly detectable levels of UGT1A1 in the liver (Fig 4D). In the Gunn rat, at least three repeated administrations of 3x10 9 I.U./ rat of 200-300 g, resulting in a total dose of 3x10 10 I.U./kg of rSV40 were needed to provide a reduction of serum bilirubin by 70% [9]. In this animal model, AAV does provide therapeutic correction upon peripheral administration [28]. A dose of 5x10 12 AAV8 vg/kg comparable to 2x10 8 Huh7 transducing AAV8 units, results in sustained and complete normalization of serum bilirubin [35]. These data indicate that the liver transduction efficacy of AAV vectors is higher than that of rSV40 vectors. This difference in liver transduction efficacy between AAV and SV40 vectors in rats has also been observed by others [36]. In that study a dose of 3.4x10 9 vg of AAV1-IGF1 provided a comparable expression of IGF-1 as 1x10 11 vg of rSV40-IGF-1 per rat, which is a 30 fold higher dose. Although, comparing doses of different viral vectors is NAbs titers are expressed as the highest serum dilution that inhibited AAV transduction by � 50% compared with the control without serum. Pre represents before vector injection. Data represent the mean ± SD. Statistical significance was determined by Ordinary one-way ANOVA with Dunnett's multiple comparison test.
https://doi.org/10.1371/journal.pone.0250605.g005 complicated by different titration methods, the higher efficacy seen with AAV vector renders that vector more suitable for in vivo gene therapy for inherited liver disorders. Another major hurdle for rSV40 vectors is the absence of a production and purification protocol to generate sufficiently high titered vector batches, which rendered testing the therapeutic efficacy of higher rSV40 vector doses in vivo is not possible because of toxicity. The cause of the toxicity seen in two out of four mice upon injection of the high dose of rSV-hUGT1A1 suggest the presence of impurities in our vector batch.
In the Gunn rat, repeated injections with rSV-hUGT1A1 did not inhibited liver transduction of rSV40-HBs antigen [9]. This indicated that rSV40 did not induce neutralizing antibodies. Others also reported the absence of NAbs upon injection of rSV40 [22]. Absence of a NAbs response would be a major advantage for rSV40 compared to AAV since it renders effective retreatment feasible upon loss of efficacy. Using an in vitro NAbs assay, comparable to that used to test for a response towards AAV, we observed the induction of an immune response that prevented transduction in vitro. At 8 weeks after rSV40 administration, serum from animals treated with rSV40 completely prevented the transduction of COS-1 cells. The observation that in Gunn rats re-administration was effective may be explained by the high dose of rSV40 administered locally to the liver via the portal vein. Presence of NAbs against AAV can, at least in part, be overcome by a high dose of vector or by co-administration of empty particles as a decoy [37]. Upon portal delivery the amount of NAbs present locally may not be sufficient to block the liver transduction completely. Although we cannot exclude that the presence of impurities in the vector batches used may have had an adjuvant effect resulting in a more prominent immune response seen in this study. However, the production system used in these studies is very similar to that used in those Gunn rat studies rendering such an explanation less probable.
Presence of pre-existing immunity towards viral gene therapy vectors is a challenge for AAV mediated in vivo gene therapy for liver disorders. The presence of NAbs renders about 30% of the Crigler-Najjar patients screened not eligible for the ongoing trial [17]. In this respect, the lower pre-existing immunity towards rSV40 is an advantage. In the general population seroconversion towards SV40 immunity is observed at a young age [38]. Most studies report about 10 to 25% of all subject have antibodies to this vector. The percentage of the population that has neutralizing antibodies towards SV40 is much lower [21,39]. Although this shows that pre-existing immunity towards rSV40 is clearly lower than that reported for AAV, these studies do report presence of neutralizing antibodies in the normal healthy population. This presence supports our finding that injection of rSV40 vectors does induce neutralizing antibodies, which argues against the possibility of effective re-administration indicating this remains a hurdle for rSV40 vector also.
Conclusion
Our data indicate that a platform to generate high titered pure rSV40 vector batches is needed to support the potential use of rSV40 as an additional gene therapy vector to treat inherited liver diseases. The current vector batches have, compared to for instance AAV, a much lower transduction efficacy and lack the previously reported immune privilege could not be reproduced. Presently, the only advantage of rSV40 is the lower pre-existing immunity towards this vector in the general population.
Supporting information S1 Fig. Absence of antibodies towards hUGT1A1 in serum of mice treated with rSV hUGT1A1 or rSV-HLP-hUGT1A1. 60-day-old Ugt1a -/mice were treated with 2x1011 vg/kg of rSV-hUGT1A1 vector (n = 3), or rSV-HLP-hUGT1A1 (n = 3) via tail vein injection. At 8 weeks after vector administration the level of anti-hUGT1A1 IgG in serum was determined. Data represent the mean ± SD. (PDF) S1 Raw images. (PDF) | 5,759.6 | 2021-04-23T00:00:00.000 | [
"Biology"
] |
Particle Swarm Optimization Submission for WMT16 Tuning Task
This paper describes our submission to the Tuning Task of WMT16. We replace the grid search implemented as part of standard minimum-error rate training (MERT) in the Moses toolkit with a search based on particle swarm optimization (PSO). An older variant of PSO has been previously successfully applied and we now test it in optimizing the Tuning Task model for English-to-Czech translation. We also adapt the method in some aspects to allow for even easier parallelization of the search.
Introduction
Common models of statistical machine translation (SMT) consist of multiple features which assign probabilities or scores to possible translations. These are then combined in a weighted sum to determine the best translation given by the model. Tuning within SMT refers to the process of finding the optimal weights for these features on a given tuning set. This paper describes our submission to WMT16 Tuning Task 1 , a shared task where all the SMT model components and the tuning set are given and task participants are expected to provide only the weight settings. We took part only in English-to-Czech system tuning.
Our solution is based on the standard tuning method of Minimum Error-Rate Training (MERT, Och, 2003). The MERT algorithm described in Bertoldi et al. (2009) is the default tuning method in the Moses SMT toolkit (Koehn et al., 2007). The inner loop of the algorithm performs optimization on a space of weight vectors with a given translation metric 2 . The standard optimization is a variant of grid search and in our work, we replace it with the Particle Swarm Optimization (PSO, Eberhart et al., 1995) algorithm.
Particle Swarm Optimization is a good candidate for an efficient implementation of the inner loop of MERT due to the nature of the optimization space. The so-called Traditional PSO (TPSO) has already been tested by Suzuki et al. (2011), with a success. Improved versions of the PSO algorithm, known as Standard PSO (SPSO), have been summarized in Clerc (2012).
In this paper, we test a modified version of the latest SPSO2011 algorithm within the Moses toolkit and compare its results and computational costs with the standard Moses implementation of MERT.
MERT
The basic goal of MERT is to find optimal weights for various numerical features of an SMT system. The weights are considered optimal if they minimize an automated error metric which compares the machine translation to a human translation for a certain tuning (development) set.
Formally, each feature provides a score (sometimes a probability) that a given sentence e in goal language is the translation of the foreign sentence f . Given a weight for each such feature, it is possible to combine the scores to a single figure and find the highest scoring translation. The best translation can then be obtained by the following formula: e * = argmax e i λ i log (p i (e|f )) = g p (λ) (1) 2 All our experiments optimize the default BLEU but other metrics could be directly tested as well.
The process of finding the best translation e * is called decoding. The translations can vary significantly based on the values of the weights, therefore it is necessary to find the weights that would give the best result. This is achieved by minimizing the error of the machine translation against the human translation: The error function can also be considered as a negative value of an automated scorer. The problem with this straight-forward approach is that decoding is computationally expensive. To reduce this cost, the decoder is not run for every considered weight setting. Instead, only some promising settings are tested in a loop (called the "outer loop"): given the current best weights, the decoder is asked to produce n best translation for each sentence of the tuning set. This enlarged set of candidates allows us to estimate translation scores for similar weight settings. An optimizer uses these estimates to propose a new vector of weights and the decoder then tests this proposal in another outer loop. The outer loop is stopped when no new weight setting is proposed by the optimizer or no new translations are found by the decoder. The run of the optimizer is called the "inner loop", although it need not be iterative in any sense. The optimizer tries to find the best weights so that the least erroneous translations appear as high as possible in the n-best lists of candidate translations.
Our algorithm replaces the inner loop of MERT. It is therefore important to describe the properties of the inner loop optimization task.
Due to finite number of translations accumulated in the n-best lists (across sentences as well as outer loop iterations), the error function changes only when the change in weights leads to a change in the order of the n-best list. This is represented by numerous plateaus in the error function with discontinuities on the edges of the plateaus. This prevents the use of simple gradient methods. We can define a local optimum not in a strict mathematical sense but as a plateau which has only higher or only lower plateaus at the edges. These local optima can then be numerous within the search space and trap any optimizing algorithm, thus preventing convergence to the global optimum which is desired.
Another problem is the relatively high dimensionality of the search space. The Tuning Task model has 21 features but adding sparse features, we can get to thousands of dimensions.
These properties of the search space make PSO an interesting candidate for the inner loop algorithm. PSO is stochastic so it doesn't require smoothness of the optimized function. It is also highly parallelizable and gains more power with more CPUs available, which is welcome since the optimization itself is quite expensive. The simplicity of PSO also leaves space for various improvements.
PSO Algorithm
The PSO algorithm was first described by Eberhart et al. (1995). PSO is an iterative optimization method inspired by the behavior of groups of animals such as flocks of birds or schools of fish. The space is searched by individual particles with their own positions and velocities. The particles can inform others of their current and previous positions and their properties.
TPSO
The original algorithm is defined quite generally. Let us formally introduce the procedure. The search space S is defined as where D is the dimension of the space and min d and max d are the minimal and maximal values for the d-th coordinate. We try to find a point in the space which maximizes a given function f : S → R.
There are p particles and the i-th particle in the n-th iteration has the following D-dimensional vectors: position x n i , velocity v n i , and two vectors of maxima found so far: the best position p n i visited by the particle itself and the best known position l n i that the particle has learned about from others.
In TPSO algorithm, the l n i vector is always the globally best position visited by any particle so far.
The TPSO algorithm starts with simple initialization: where the function rand(S) generates a random vector from space S with uniform distribution.
The velocity for the next iteration is updated as follows: where U (0, 1) denotes a random number between 0 and 1 with uniform distribution. The parameters w, φ p , φ l ∈ (0, 1) are set by the user and indicate a slowdown, and the respective weight for own vs. learned optimum.
All the following vectors are then updated: The process continues with the next iteration until all of the particles converge to proximity of a certain point. Other stopping criteria are also used.
Modified SPSO2011
We introduce a number of changes to the algorithm SPSO2011 described by Clerc (2012).
In SPSO2011 the global best position l t i is replaced by the best position the particle has received information about from other particles. In the original SPSO2011 this is done in a synchronized fashion: after every iteration, all particles send their best personal positions to m other particles. Every particle chooses the best position it has received in the current iteration and sets its l t i accordingly. This generalization of l t i is introduced in order to combat premature convergence to a local optimum.
To avoid waiting until all particles finish their computation, we introduce per-particle memory of "learned best positions" called the "neighbourhood set" (although its members do not have to be located in any close vicinity). This set of best positions is limited to k elements, each new addition over the limit k replaces the oldest information. To establish the "global" optimum l t i , every particle consults only its set of learned best positions.
The algorithm starts with the initialization of particle vectors given by the equations (4-6). The l 0 i is initialized with the value of p 0 i . The sets of learned best positions are initialized as empty.
Two constants affect computations given below: w is again the slowdown and c controls the "expansion" of examined neighbourhood of each particle. We set w and c to values that (as per Bonyadi and Michalewicz, 2014) ensure convergence: Figure 1: Construction of the particle position update. The grey area indicates P (G, x).
For the update of velocity, it is first necessary to calculate a "center of gravity" G t i of three points: the current position x t i , a slightly "expanded" current best position p t i and a slightly expanded best position known by colleagues l t i . The "expansion" of the positions is controlled by c and directed outwards from x t i : To introduce further randomness, x t i is relocated to a position y t i sampled from the uniform distribution in the area P (G t i , x t i ) formally defined as: Our P (G, x) is a hypercube centered in G t i and touching x t i , see Figure 1 for an illustration. The original SPSO2011 used a d-dimensional ball with the center in G and radius G − x to avoid the bias of searching towards points on axes. We are less concerned about this and opt for a simpler and faster calculation.
The new velocity is set to include the previous velocity (reduced by w) as well as the speedup caused by the random relocation: Finally, the particle position is updated: The optimized function is evaluated at the new position x t+1 i and the particle's best position is updated if a new optimum was found. In any case, the best position p t+1 i together with its value is sent to m randomly selected particles (possibly including the current particle) to be included in their sets of learned best positions as described above. The particle then sets its l t+1 i to best position from its own list of learned positions.
The next iteration continues with the updated vectors. Normally, the algorithm would terminate when all particles converge to a close proximity to each other, but it turns out that this often leads to premature stopping. There are many other approaches possible to this problem (Xinchao, 2010; Evers and Ben Ghalia, 2009), but we choose a simple restarting strategy: when the particle is sending out its new best position and value to m fellows, the manager responsible for this checks if this value was not reported in the previous call (from any other particle). If it was, then the current particle is instructed to restart itself by setting all of its vectors to random initial state. 3 The neighborhood set is left unchanged. The restart prevents multiple particles exploring the same area.
The drawback of restarts is that the stopping criterion is never met. In our first version, we ran the algorithm for a fixed number of position updates, specifically 32000. Later, we changed the algorithm to terminate after the manager has seen 3200 position updates without any update of the global best position. In the following section, we refer to the former as PSO without the termination condition (PSO) and the latter as PSO with the termination condition (PSO-T).
Properties of SPSO2011 have been investigated by Bonyadi and Michalewicz (2014). We use a slightly different algorithm, but our modifications should have an effect only on rotational invariance, which is not so much relevant for our purpose. Aside from the discussion on the values of w and c with respect to the convergence of all particles to the same point, Bonyadi and Michalewicz also mention that SPSO2011 is not guaranteed to converge to a local optimum. Since our search space is discontinuous with plateaus, the local convergence in the mathematical sense is not especially useful anyway.
Implementation
We implemented the algorithm described above with one parameter, the number of particles. We set the size of the neighborhood set, denoted k above, to 4 and the number of random particles receiving the information about a particle's best position so far (m) to 3.
The implementation of our version of the PSO algorithm is built within the standard Moses code. The algorithm itself creates a reasonable parallel structure with each thread representing a single particle.
We use similar object structure as the baseline MERT implementation. The points are represented by their own class which handles basic arithmetic and stream operations. The class carries not only the vector of the current position but also its associated score.
Multiple threads are maintained by the standard Moses thread pools (Haddow, 2012). Every thread ("Task" in Moses thread pools) corresponds to a particle and is responsible for calculating its search in the space using the class PSOOptimizer. There are no synchronous iterations, each particle proceeds at its own pace.
All optimizers have access to a global manager object of class PSOManager, see Figure 2 for an illustration. The manager provides methods for the optimizers to get the best vector l t i from the Figure 2: Base structure of our PSO algorithm neighborhood set, to report its best position to the random m particles (addPoint) and to check if the optimization should still run (cont) or terminate. The method addPoint serves two other purposes: incrementing an internal counter of iterations and indicating through its return value whether the reporting particle should restart itself.
Every optimizer has its own FeatureData and ScorerData, which are used to determine the score of the investigated points. As of now, the data is loaded serially, so the more threads we have, the longer the initialization takes. In the baseline implementation of MERT, all the threads share the scoring data. This means that the data is loaded only once, but due to some unexpected locking, the baseline implementation never gains speedups higher than 1.5, even with 32 threads, see Table 2 below.
This structure allows an efficient use of multiple cores. Methods of the manager are fast com-pared to the calculations performed in the optimizers. The only locking occurs when threads are trying to add points; read access to the manager can be concurrent.
Results
We ran the tuning only for the English to Czech part of the tuning task. We filtered and binarized the model supplied by the organizers to achieve better performance and smaller memory costs.
For the computation, we used the services of Metacentrum VO. Due to the relatively high memory demands we used two SGI UV 2000 machines: one with 48x 6-core Intel Xeon E5-4617 2.9GHz and 6TB RAM and one with 48x 8-core Intel Xeon E5-4627v2 3.30GHz and 6TB RAM. We ran the tuning process on 16 and 64 CPUs, i.e. with 16 and 64 particles, respectively. We submitted the weights from the 16-CPU run. We also ran a test run using the standard Moses MERT implementation with 16 threads for a comparison. Table 1 shows the best BLEU scores at the end of each inner loop (as projected from the n-best lists on the tuning set of sentences). Both methods provide similar results. Since the methods are stochastic, different runs will lead to different best positions (and different scores).
Comparison of our implementation with with the baseline MERT on a test set is not necessary. Both implementations try to maximize BLEU score, therefore any overtraining occurring in the baseline MERT occurs also in our implementation and vice versa. Table 2 shows the average run times and reached scores for 8 runs of the baseline MERT and our PSO and PSO-T, starting with the same n-best lists as accumulated in iteration 1 and 3 of the outer loop. Note that PSO and PSO-T use only as many particles as there are threads, so running them with just one thread leads to a degraded performace in terms of BLEU. With 4 or 8 threads, the three methods are on par in terms of tuningset BLEU. Starting from 4 threads, both PSO and PSO-T terminate faster than the baseline MERT implementation. Moreover the baseline MERT proved unable to utilize multiple CPUs efficiently, whereas PSO gives us up to 14-fold speedup. In general, the higher the ratio of the serial data loading to the search computation time, the worse the speedup. The search in PSO-T takes much shorter time so the overhead of serial data loading is more apparent and PSO-T seems parallelized badly and gives only quadruple speedup. The reduction of this overhead is highly desirable.
Conclusion
We presented our submission to the WMT16 Tuning Task, a variant of particle swarm optimization applied to minimum error-rate training in statistical machine translation. Our method is a drop-in replacement of the standard Moses MERT and has the benefit of easy parallelization. Preliminary experiments suggest that it indeed runs faster and delivers comparable weight settings.
The effects on the number of iterations of the MERT outer loop as well as on the test-set performance have still to be investigated. | 4,188.8 | 2016-01-01T00:00:00.000 | [
"Computer Science"
] |
Prediction of Retail Price of Sporting Goods Based on LSTM Network
Commodity prices play a unique role as a lever to regulate the economy. Price forecasting is an important part of macrodecision-making and micromanagement. Because there are many factors affecting the price of goods, price prediction has become a difficulty in research. According to the characteristics that price data are also affected by other factors except for time series, a multifactor LSTM price prediction method is proposed based on the long-term and short-term memory network (LSTM) deep learning algorithm. This method not only makes use of the memory of LSTM to historical data but also introduces the influence of external factors on price through the full connection layer, which provides a new idea for solving the problem of price prediction. Compared with BP neural network, the experimental results show that this method has higher accuracy and better stability. Analyze the commodity description and commodity price characteristics, find out the commodities similar to the target commodity, complete the commodity price data by using the historical price data of similar commodities, and establish the training set to verify the validity of the proposed method.
Introduction
e market scale of the commodity market is expanding day by day, the trading varieties are becoming richer and richer, and the trading mechanism is becoming more and more standardized. Since the development of the commodity market, it has become an important financial submarket. e price signal spread in the market plays an important guiding role in guiding enterprise production, international trade, and regulating the economy [1]. China's commodity exchanges have also accumulated rich market data in many years of operation and development. Researchers collect, sort out, and analyze the transaction data of various commodities and then add and compile the same commodity indexes. ese indexes can reflect the overall state of relevant commodity price fluctuations and the development trend of the commodity economy and help government functional institutions understand the tail state of the macroeconomy; at the same time, relevant enterprises can also use the rice commodity price information contained in the commodity index to make their own business decisions, reasonably arrange the purchased quantity and minimize unnecessary economic losses [2,3].
Commercial activities occupy a more and more important position in the national economy. Commercial behavior not only makes an important contribution to the national economy in terms of output but also plays an important role in breeding market relations, improving market mechanism, and solving the problem of labor employment [4]. Today, with the high development of information technology, e-retail commerce, as a new retail commodity sales model, has developed rapidly. People can obtain thousands of commodity information through the Internet and contact commodity sellers all over the world for transactions without going out of their homes [5]. e emergence of e-commerce has greatly facilitated the people's life, promoted people's consumption enthusiasm, and increased the vitality of the consumer market; the development of e-retail commerce has driven the prosperity of the business market. Tens of thousands of retail commodities are displayed and sold on the Internet, which reduces the cost of commodity sales and improves the efficiency of sales. At the same time, because e-retail commerce relies on the Internet, it has natural advantages in the acquisition and storage of information and data. Enterprises obtain a large amount of data through information technology. How to mine these data and find valuable laws, so as to guide the business decision-making of enterprises, improve the sales model, formulate effective sales strategies, and finally obtain economic benefits from e-commerce. Many enterprises begin to invest in the direction of commercial data mining and develop their own commercial data mining schemes.
Exploratory data analysis is an important stage of data analysis, which is different from initial data analysis. e focus of preliminary data analysis is on whether the requirements for identifying statistical models and hypotheses are met to ensure the reliability of confirmatory analysis [6][7][8]. In this analysis process, the unqualified data are filled with missing value, data conversion, outlier value discarding, and other processing to enhance the accuracy of the analysis. Exploratory data analysis includes preliminary data analysis, but its starting point is not only to determine data quality but also to discover patterns of data distribution (Patten) and propose new hypotheses from the data. Exploratory data analysis is identified as a key step in the data science workflow that can affect multiple processes. In the data science workflow shown in Figure 1, exploratory data analysis is closely related to other processes.
Each base model trains the input variables of the training set separately.
rough supervised learning, each base model. A weight is established for the predicted value of the training set of each base model by means of a weighted average, and then the established weight value is assigned to the predicted value of the test set of each base model [9,10]. Forecast weight value multiplication and intercept Get model test set of predicted values. Figure 2 shows the construction flow and diagram of the model. From the figure we know that it mainly includes the data preprocessing, the multiple prediction results from multiple LSTM models and the prediction results are integrated and enhanced by weighting multiple parameters.
By analyzing the weight value of each base model of sports products, it is found that the base model with higher prediction accuracy is more likely to get more weight. Figure 3 is the technology roadmap.
Related Work
Price prediction refers to the prediction behavior of dynamic analysis of future price changes according to the historical value and price trend of commodities [11]. e authors of [12] used the recurrent neural network model for high-speed train vibration prediction from time series and achieved good results.
With the development of a simple time series algorithm, the application of simple time series analysis is gradually expanding. At present simple time series analysis algorithm has been in agricultural prices, industrial commodity prices prediction, financial stock price forecasting, and many other fields have a wide range of applications, due to less used in the analysis of data information, at the same time the low effective information analysis to history, lead to predicting the results still cannot meet the needs of social development.
Research status of price forecasting algorithm based on simple time series: according to the periodicity and seasonality of power price fluctuation, Marcjasz et al. [13,14] analyzed the influence of seasonality on future power price change and used NARX neural networks model to predict power price, which achieved good results. In order to obtain good prediction result, the authors of [15] designed a Denoising Aggregation of Graph neural networks by using the principal component analysis. Wang et al. [16][17][18] studied the short-term electricity price forecasting with stacked denoising autoencoders, do the research and application of a hybrid forecasting framework, and proposed the novel hybrid model for air quality index two-phase decomposition technique and modified extreme learning machine (ELM), respectively. Chong et al. [19,20] studied the deep learning networks for stock market analysis and prediction and carried the empirical asset pricing via machine learning, respectively. Nilashi et al. presented an analytical approach for big social data analysis for customer decision-making in eco-friendly hotels and tested the proposed solution on two open datasets [21]. In addition to the above works, Hoseinzade [22] combined the ANN model and CNN model, predicted the rise and fall of Shanghai Futures in the known period and achieved good prediction results. Investors can also make investment decisions with the help of Chen and Ge [23] explored the attention mechanism in LSTM-based Hong Kong stock price movement prediction. Fang et al. do the research on quantitative investment strategies based on deep learning [24]. Based on the above-given discussions, the main contributions of this paper are summarized as follows: (1) As an improved structure of the RNN model, the LSTM model not only inherits the characteristics of the RNN model suitable for dealing with time series data but also further solves the problem of a longterm dependence on time dimension and improves the accuracy of prediction. Its prediction effect is superior to BP neural network, RNN, CNN, GRU, and other neural network models. (2) Grid Search is used to train the model with different parameters and cross-validating each model until the optimal combination of values is found to ensure the best model performance. (3) e optimized model is verified on the test set, and the mean square error is used as the evaluation index to prevent the model from overfitting. e results show that the model achieves low mean square error in both the training set and the test set, and obtains ideal prediction results.
LSTM Network
LSTM neural network was first proposed by Hochreiter et al.
(1997) and further extended after its optimization and improvement by Alex graves. In many practical problems related to sequence data, LSTM has achieved great success and has been widely used, such as natural language processing (NLP), time series prediction, and so on. e traditional recurrent neural network (RNN) can not deal with the long-term sequence problem, to solve this problem, the LSTM neural network is proposed by adding a "gate" structure to control the cell state and output at different times to alleviate the problem of gradient disappearance. e "gate" structure of LSTM includes three types: "forgetting gate," "input gate," and "output gate." e function of the "forgetting gate" is to judge the information transmitted from the previous time to the current time and selectively "forget" some information is shown in Figure 4. In addition, the orthogonal initialization is proposed to avoid gradient disappearance or explosion at the initial stage of training, ReLU (Rectified Linear Unit) activation function can alleviate gradient disappearance, gradient shear can solve gradient explosion, and the LSTM Unit can control gradient disappearance. LSTM has been successfully applied Computational Intelligence and Neuroscience 3 in machine translation, conversation generation, and other fields, showing excellent modeling ability of sequence data. erefore, this paper builds a retail price of a sporting goods prediction model based on the LSTM network unit and can make full use of its feature that any length sequence can be used as input and apply it to online data recognition. LSTM solves the problem that RNN cannot handle long time dependence by introducing [25].
Let the number of input neurons in the whole hidden layer be G, G includes all units and gates and use index G to represent these input neurons. e forward calculation of LSTM is to calculate an input sequence X with a length of time d, whose starting point is t � 1 [26,27]. When the value of time point T increases continuously, the equation will be updated recursively until t � t. Like forward calculation, reverse calculation is an input sequence X with a time length of T, but the starting point of reverse calculation is T � T. When the value of T decreases continuously, the reciprocal of the unit is calculated recursively until T �1. According to the derivatives at each time point above, we can obtain the final weight derivative value.
where l is the loss function used for training. e value of the input gate at time t is where W f and b f represent the weight and bias of the forgetting gate, respectively, while σ represents the Sigmoid function.
Cell State
Update. e cell state is updated according to the calculation results of the input gate and forgetting gate, so as to obtain the cell state update value at the moment [28].
3.3. Output Gate. Control determines which information needs to be output. According to the computed cell status update value C t , the following calculation formula of the output gate can be obtained: Sigmoid activation function and TANH activation function are used in the gated structure of the neural unit structure of the LSTM neural network. Here, the Sigmoid function is tanh function is Before the data are input into the neural network model, the data need to be normalized. Normalization of the features of a numeric type can unify all the features into a roughly identical numeric interval. e normalization of data can eliminate dimensionality, thus avoiding the dependence of data on the choice of units of measurement and help to improve the performance of the model [29].
where x i is the I variable; x is mean of x i ; max (x) and min (z) represent the maximum and minimum values x i . By using different neural network models to train the model on the training set and test the model on the test set, the model performance of different neural network models under different parameter settings is compared. In this paper, MAE, MSE, MAPE, and the correlation coefficient (P) between predicted data and real data were selected as the evaluation indexes of model performance.
e MAE is e MSE is e MSPE is e MAPE � 1 n n i�1 y i − y i y i . (11) e correlation coefficient is where n is the number of test data sets; y i is the true value of the i sample point; y i the model predicted value of the sample point; Y is the true value of sample; Y is the predicted value of the model; COV(Y, Y) is the covariance between y and y; VAR(Y) is the variance of l; VAR(Y) of y is the variance of Y.
Data Processing and Exploration
e main innovation point of this paper is to improve the conventional LSTM model and apply it to forecast the retail price of sports products, so we mainly compare it with the conventional LSTM model. Secondly, because real data sets are precious and difficult to obtain, this paper uses only one data set for the simulation experiment. Because the quality of the data affects the training of the selected model, the collection, analysis, and processing of the data are the key stages before the model training. e data in this paper mainly include two parts: the research object and the characteristic data [30].
General forecast commodity price focuses on sports commodity price. Use Python to conduct descriptive statistical analysis of different sports commodities and draw their closing price charts [31,32], as shown in Figure 5. e distribution is skewed to the right, the peakness is smaller than 3, the tail shape is thin, and does not obey the normal distribution.
Data Noise Reduction.
As the market dynamics are very complex, these data contain infrequent noises, so the library in Python is used for wavelet transformation to remove data noises [33]. It is worth noting that the conventional wavelet change model is used in this paper. It inherits and develops the idea of short-time Fourier transform localization and overcomes the shortcomings of window size not changing with frequency. It can provide a "time-frequency" window changing with frequency, which is an ideal tool for timefrequency analysis and processing of signals. erefore, it is especially suitable for removing noise in financial data. Figures 6 and 7 are the comparisons before and after wavelet transformation.
After the LSTM neural network training is completed, the prediction results and corresponding MSE values of model 1 and Model 2 are given, respectively, as shown in Figure 8, and the prediction results of model 1 and Model 2 are shown in Figure 9.
e basic idea of model training is to fit a set of rules system on the training data set to reveal the rules in the data. In other words, the fit describes how well, or how well, the model can be generalized to the data in the test set. A good Computational Intelligence and Neuroscience model results in good model performance and can be validated with new data outside the training data set, i.e., outof-sample data. In addition to parameters that can be learned, different models require different hyperparameters, which are parameters that do not need to be trained.
Parameters are critical to the model and depend on training data. As part of the training process, the LSTM model is further adjusted and optimized to obtain a better prediction effect by learning parameters from training data through optimization techniques [34][35][36][37]. e process of economic research is generally to explain economic phenomena by constructing economic models. When generally accepted economic phenomena are confirmed by scholars, they need to be listed in the form of typical facts. In the followup research, if the solution of the model in the general equilibrium state is consistent with the typical facts, it can explain that the model is more reasonable to a great extent. In the study of financial problems, especially in the study of financial time return series, some common statistical characteristics can often be observed. According to scholars' typical facts of univariate return series, they are summarized as follows: they often show autocorrelation; It often shows long memory; e slow decay of absolute return autocorrelation; peak thick tail distribution; the distribution shape changes with time; wave aggregation effect; after adjusting the fluctuation aggregation, there is still a conditional thick tail effect. Under the assumption of independent and identically distributed, the performance of the fitted financial time series model is often not optimal, so the distribution characteristics of the data need to be considered when modeling the income series. When the neural network model is used to model the data, the assumption of distribution does not need to be considered. is is because the neural network model has the ability to generalize the structure of input data so that the nonlinear characteristics of financial data can be captured by the neural network.
Grid Search is a method of systematically training a model, using different combinations of hyperparameter values to train the model, cross-validating each model until the optimal combination of values is found to ensure the best model performance. Can through continuous testing parameters of all combinations, a group to find the most appropriate combination of super configuration parameters will be discretization, super parameters according to their own characteristics to select several experience values, and then according to different combinations of the training model, so as to select an optimal combination of configuration, the circumstances of less suitable for super parameters. Random Search is to randomly combine hyperparameters and then select the optimal configuration. It does not make unnecessary attempts on unimportant parameters because, just as regularization coefficients have a limited impact on model performance, learning rates have a greater impact on model performance, so it does not make unnecessary attempts. Random searches are generally more efficient and easier to implement than grid searches. However, these two methods do not consider whether there is a correlation between hyperparameters, so they are relatively inefficient. Bayesian optimization is an adaptive hyperparameter optimization method, which predicts the next possible hyperparameter combination based on the tested hyperparameter combination in order to obtain the maximum utility. Since the cumulative distribution function of Gaussian distribution is an s-type function, the GELU function can be approximated by tanh function or Logistic function as shown in Figure 10. e simplest strategy is to fix a learning rate throughout the training process. Choosing a smaller learning rate allows the optimizer to find a good solution, but it is easy to limit the convergence rate. e relationship between the two can be balanced by taking time to change the learning rate. Figure 11 shows the learning rate for each period.
After many times of adjustment and optimization, the prediction results before and after improvement are shown in Figures 12 and 13, respectively. e final model structure and parameters obtained are as follows: the Sequential length of the test window is 55, and the sequential model consists of three LSTM layers, with the number of neurons in each layer being 100, 100, and 150, respectively. In order to avoid overfitting, two Dropout layers are added with the Dropout layer of 0.2, and the dimension of input data are 5. e dense layer was added to aggregate its dimension into 1, the activation function was linear, and the loss function was set as Mean Squared Error (MSE). Adam was used as the Computational Intelligence and Neuroscience optimization algorithm, and two Epochs were used as the model. Each batch is 32 in size. e conventional price forecasting model is also a deep learning model. Firstly, a conventional price prediction model is constructed; that is, the model only contains market data and financial data. e previous market data and financial data are used to train the deep learning model to get the trained deep learning model to predict the future market; rough natural language processing technology, emotional information extraction and emotional evaluation are carried out on public opinion data. Combined with the conventional price prediction model, the in-depth learning model is trained by using market data, financial data, research reports, and emotional tendency data of financial news, and the model is used to predict the future market. en, the prediction effect of the conventional securities price prediction model is compared with that of the deep learning model based on the natural language processing results of public opinion data.
After determining the number of input nodes, output nodes, and hidden layer nodes of the two models LSTM network, the deep learning model can be trained. After many experiments, it is found that the training times are too few and the model training error is too large, so it is necessary to continuously increase the training times, but with the increase of the training times; the error of model training gradually tends to a stable value. If the training times of the model are increased a lot at this time, the model effect is not improved much. It is the result of training the first model of CCCC. When the training times are less than 200, the error of the model is large. At this time, increasing the training times will quickly reduce the error of model training; When the training times are more than 200 times and less than 1000 times, the model error has been small. At this time, when the training times are increased, the reduction degree of model training error has shown a decreasing law; When training 1000 to 2000 times, the error of the model changes in a small area, and the effect of increasing the number of training times are gradually not obvious. rough many tests and comparisons, it is found that when the number of training times of the model is about 2000 times, it can meet the accuracy requirements of training. If the number of training times is increased, the effect of improving the model training error is small, Moreover, it takes a long time to train the model in the computer, so it is of little significance to increase too many training times. erefore, we set the training times of the in-depth learning model in this study at 2000 times. In the later model, we also verify that it conforms to this law. erefore, the second in-depth learning model with public opinion information data also set the training times at 2000 times.
Conclusion
(1) is paper applies deep learning theory, based on the characteristics of financial time series data, uses LSTM neural network model to predict sports commodity price index, and compares its prediction results with the prediction results using the network model. Experimental results show that LSTM neural network model has the best performance on the test set.
(2) A prediction model is established. On the basis of collated data feature engineering, a neural network prediction model based on long and short-term memory is established, and the model is trained with a training set to predict the price of sports goods. In terms of model optimization, the number of hidden layer neurons, learning rate, batch size, and training wheel were adjusted to achieve the best training results. Computational Intelligence and Neuroscience influences on the prediction effect, so the selection of the data set is also very important.
Although the model proposed in this paper achieves good prediction results, the model does not consider the correlation of data time. Some sliding time window tools can be used in future studies to improve the prediction step size and prediction accuracy of the model.
Data Availability
e data used to support the findings of this study are available from the author upon request.
Conflicts of Interest
e author declares that there are no conflicts of interest. | 5,696.6 | 2022-07-09T00:00:00.000 | [
"Economics"
] |
Abnormally elevated USP37 expression in breast cancer stem cells regulates stemness, epithelial-mesenchymal transition and cisplatin sensitivity
Background Recent studies have indicated that deubiquitinating enzymes (DUBs) are related to the stem-cell pathway network and chemo-resistance in cancer. Ubiquitin-specific peptidase 37 (USP37), a novel DUB, was identified to be a potential factor associated with tumor progression. However, the biological functions of USP37 in breast cancer remain unclear. Methods The distribution of USP37 expression in breast cancer and the correlation between USP37 expression and the overall survival rate were detected by The Cancer Genome Atlas (TCGA) database. Gene set enrichment analysis (GSEA) was utilized to evaluate potential mechanism of USP37 in breast cancer. The USP37 expression in breast cancer tissues and breast cancer cell lines were detected by immunohistochemistry and western blotting. Sorting of breast cancer stem cells (BCSCs) were by using MACS assay. In vitro and in vivo assays were performed to examine the biological functions of USP37 in breast cancer cells. MG132, CHX chase, immunofluorescence staining and co-immunoprecipitation assays were used to test the interaction between USP37 and Gli-1. Results Bioinformatics analysis demonstrated that USP37 gene was elevated in breast cancer tissues and its overexpression was strongly correlated with the increased mortality rate. GSEA analysis showed that USP37 expression was positively associated with cell growth and metastasis while negatively related to cell apoptosis in the TCGA breast cancer samples. USP37 expression was elevated in breast cancer tissues and breast cancer cell lines. Moreover, we also detected that USP37 was overexpressed in BCSCs. USP37 regulated the ability of cell invasion, epithelial-mesenchymal transition (EMT), stemness and cisplatin sensitivity in breast cancer cell lines. Additionally, USP37 knockdown inhibited tumorigenicity and increased anticancer effect of cisplatin in vivo. Knockdown of USP37 significantly decreased hedgehog (Hh) pathway components Smo and Gli-1. Gli-1 was stabilized by USP37 and they interacted with each other. Further studies indicated that USP37 knockdown could inhibit the stemness, cell invasion and EMT in breast cancer via downregulation of Hh pathway. Conclusions These findings reveal that USP37 is highly expressed in BCSCs and is correlated with poor prognosis in breast cancer patients. USP37 can regulate the stemness, cell invasion and EMT via Hh pathway, and decreased USP37 confers sensitivity to cisplatin in breast cancer cells. USP37 is required for the regulation of breast cancer progression, as well as a critical target for clinical treatment of breast cancer. Electronic supplementary material The online version of this article (10.1186/s13046-018-0934-9) contains supplementary material, which is available to authorized users.
Introduction
Breast cancer is the deadliest form of carcinoma affecting women, with nearly a quarter-million cases diagnosed in 2016 [1]. Although there are effective treatments against some types of breast carcinoma, such as those for subtypes with abnormal overexpression of the HER2/Neu oncogene, the majority breast cancers remain incurable. Breast cancer stem cells (BCSCs) possess self-renewal and differentiation capabilities, leading to tumor recurrence, metastasis and therapeutic resistance [2,3]. CD44 + / CD24 − or aldehyde dehydrogenase1 (ALDH1) phenotypes are efficient in the identification of BCSCs from breast cancer populations. However, there is a small overlap between CD44 + /CD24 − and ALDH1 stem phenotypes, as well as less stem markers in differentiation of different breast cancer subtypes [4]. Therefore, it is necessary to identify more discriminatory biomarkers of distinct molecular subtypes for the isolation and identification of the BCSCs subpopulation.
EMT describes the process by which epithelial cells detach from neighboring cells and are transferred to other tissue sites via dissolution of basement membrane and passage through the extra-cellular matrix [5,6]. EMT could also facilitate the generation of cancer stem cells from more differentiated cancer cells [7]. The EMT process during breast carcinogenesis is considered to be controlled by a series of signaling pathways, including Notch, Wnt/β-catenin and Hedgehog [8,9]. The Hedgehog (Hh) pathway is responsible for the maintenance of stem cells and EMT, which can contribute to the evolution of breast cancer [10].
Ubiquitination describes a highly conserved and reversible modification process of protein degradation, which is involved in nearly all aspects of cell biology [11]. Deubiquitinating enzymes (DUBs) can prevent ubiquitin-mediated degradation of target proteins [12]. Importantly, dysregulated DUBs expression is frequently associated with the tumorigenesis process, specifically cell self-renewal, apoptosis and EMT [13,14]. It has been confirmed that DUBs are essential for the regulation of stem cell-related markers and controlling various steps of metastatic progression, including invasion, dissemination and eventual metastasis to distant organs [14,15]. Ubiquitin specific peptidase 37 (USP37), a novel DUB, is a member of ubiquitin-specific processing proteases family. Human USP37, localized mainly in the cytoplasm, is composed of 979 amino acids harboring three ubiquitin-interacting motifs between the Cys box and His box of the primary sequence [16,17]. The function of USP37 was initially identified as a potent regulator of the cell cycle where it could accelerate the G1/S transition with exceptional high expression [18,19]. Previous studies have found that USP37 could regulate the stem cell-related marker SOX2 by binding with its promoter region at the transcriptional level [20]. Pan et al. reported that high levels of USP37 gene expression in lung cancer promoted cell viability as well as the Warburg effect via deubiquitination and stabilization of pluripotent factor c-Myc protein [21]. These advances implicated that USP37 gene may be associated with the stemness of tumor cells facilitating cancer progression. Recent studies indicate that USP37 is a potential factor involved in breast cancer progression [16]. However, the biological function of USP37 in the direct regulation of BCSCs and EMT remains unexplored.
In this study, we found that USP37 expression was upregulated in breast cancer tissues compared with surrounding tissues and its overexpression was significantly correlated with increased rates of mortality. We demonstrated that USP37 was highly expressed in BCSCs. The knockdown of USP37 could inhibit the stemness, cell invasion and EMT via downregulation of Hedgehog pathway. USP37 also interacted with and stabilized glioma-associated oncogene 1 (Gli-1) protein. Additionally, USP37 knockdown enhanced the sensitivity of breast cancer cells to cisplatin in vitro and in vivo. We postulate that USP37 may represent a novel molecular target for breast cancer treatment.
Bioinformatic analysis
Gene data extracted from invasive breast tumor samples was obtained from the TCGA Data Portal. Results of analysis were used to create figures via GraphPad Prism. According to PAM50 gene expression signature, information about four breast cancer subtypes (basal-like, Luminal A, Luminal B, and enriched Her-2) were classified. A scatter plot diagram, where each dot indicated an individual sample, was synthesized using the All Complete Tumors of Breast Invasive Cancer dataset [22,23]. For survival analysis, clinical data related to invasive breast carcinoma were downloaded from the TCGA database [22]. Kaplan-Meier curves were analyzed by GraphPad Prism.
Gene set enrichment analysis (GSEA) was utilized to evaluate potentially biological mechanism associated with USP37 mRNA expression levels in the TCGA breast cancer samples. GSEA software which was obtained from the Broad Institute draws the result pictures automatically.
BALB/c Nude mice (6 to 8 weeks old) were purchased from Vital River Laboratory Animal Technology Company (Beijing, China) and reared according to the Animal Care and Use Committee of Dalian Medical University.
siRNA, shRNA, lentivirus
The design and synthesis of siRNAs were completed by RiboBio company (Guangzhou, China), and transfection of siRNAs against USP37 was performed with Lipofectamine 2000. MCF-7 and MDA-MB-231 cells were transfected with USP37 siRNAs or a negative control (NC) siRNA. Lentivirus vectors including short hairpin RNA against USP37 (shUSP37#2) and the negative control (shScramble) were purchased from GenePharma Company (Shanghai, China). The transfected cells were treated with puromycin (Clontech, USA) for selection. To overexpress USP37 in MCF-7 cells, the pEZ-M35-USP37 plasmid (2 μg) were infected into MCF-7 cells via Lipofectamine 2000. Expression was confirmed by RT-qPCR and western blotting. The siRNA or shRNA sequences were listed in Additional file 1: Table S1.
Colony formation assay and cell viability assay
Colony formation assay of breast cancer cells was carried out by plating infected cells at a density of 1000 cells/well in a 6-well plate and then different concentrations of cisplatin were added. After 2 weeks of incubation, cells were washed three times with phosphate buffered saline, fixed with methanol and stained with 1% crystal violet. Visible colonies were stained violet and counted for data analysis. The detection of cell viability was performed in accordance with the Cell Counting Kit 8 assay (Dojindo Laboratories, Kumamoto, Japan).
Cell invasion assay
Cells were similarly cultured in MEM Alpha Modification medium for 24 hours. Briefly, 1 × 10 5 cells were seeded without serum into 24-well insert Transwell chambers (8 μm pore size, Corning, USA) and pretreated with Matrigel (BD, Bioscience, San Jose, CA, USA). Medium supplemented with 20% serum was added into the lower chamber. After 12-16 hours of seeding, cotton swabs were used to clean the upper cells. The cells on the other side of the membrane were stained with 1% crystal violet. A randomly selected area was counted with an optical microscope.
Wound healing assay
Cells were incubated in 12-well plates. When cellular density reached nearly 100%, the cell monolayer was wounded with a 200 μl micro-pipette tip. The wound areas were washed three times with phosphate buffered saline (PBS). Then the medium was changed to MEM Alpha Modification without FBS. The wound areas were micrographed at 0, 12, 24 and 48 h. All assays were performed in triplicate.
RNA extraction and real-time quantitative PCR analysis
Total RNA was extracted from cultured cells with Trizol reagent (Transgen, China). In total, 1 μg of total RNA was reverse-transcribed with the All-in-one First-Strand cDNA Synthesis SuperMiX kit (Transgen, China). The mRNA expression levels of USP37, ALDH1, CD24 and CD44 were quantified by real-time PCR. RT-qPCR was analyzed with the iCycler™ Real Time System and a SYBR Premix EX Tag Master mixture kit (Transgen, China) according to the manufacturer's instructions. The relative expression levels of mRNA were evaluated by using the 2 −ΔΔCt method. Primer sequences are listed in Additional file 1: Table S1.
CHX chase assay and co-immunoprecipitation (co-IP) assay
For the CHX chase assay, cells were treated with CHX (50 μg/ml) and harvested at the indicated timepoints. Treated cells were lysed, and the lysates were analyzed by western blotting with anti-USP37 or anti-GAPDH antibodies. For the Co-IP assay, 8 × 10 6 MCF-7 cells were harvested and lysed in IP lysis buffer containing protease inhibitor. The experiment based on manufacturer's protocols (IP Kit, Proteintech Group) was performed.
Mouse xenograft mode
Animal experiments and procedures were conducted in accordance with the Guide for the Care and Use of Laboratory Animals (NIH). Twenty BALB/C nude mice were randomly divided into four groups. The negative control group (shScramble) and shUSP37#2-transfected MCF-7 cells (5 × 10 5 ) were resuspended in 100 μl PBS and injected into mammary fat pads. When the size of tumor reached approximately 100 mm 3 , animals were randomly treated with cisplatin (2 mg/kg) or 0.9% saline. Mice were then retreated with their assigned treatment once every 2 days. The tumor sizes were measured with a Vernier caliper and recorded every other day. The tumor volume was calculated using the formula: After the inoculation of tumors for 3 weeks, xenografted tumors were excised from sacrificed mice then analyzed by immunohistochemistry and western blotting.
Immunohistochemical analysis
Tumor tissues were obtained from mouse xenograft mode. Paraffin-embedded tissue was cut into 5 μm thick slices that were fixed onto glass slides. All human breast cancer tissue arrays were purchased from Shanghai Outdo Biotech (Shanghai, China), including 60 cases of cancer tissues and surrounding tissues (HBre-Duc060CS-03; HBre-Duc060CS-04). These tissue sections were immunostained with corresponding antibodies, and then deparaffinized in xylene and rehydrated with ethanol. Tissue sections were preincubated with 10% normal goat serum, followed by incubation with primary antibody solution overnight at 4°C. After washing with PBS, slides were incubated with the secondary antibody at 37°C for 10 min, then cleaned with cold PBS and treated with peroxidase conjugated-biotin streptavidin complex for 10 min. Finally stains were examined with 3,3′-diaminobenzidine and hematoxylin. As previously described [23], the immunostained tissues were scored by multiplying the intensity (0-3) and extent (0-100) of staining.
Ethical approval
This study was conducted with the approval of the Ethical Committee and Institutional Review Board of Dalian Medical University.
Statistical analysis
Statistical analysis was performed with SPSS software version 11.0. Data are expressed as mean ± SD. Differences between two groups were evaluated by Student's t-test. One-way ANOVA was used when comparing multiple groups. P < 0.05 was considered statistically significant. Clinical data analysis of survival and relevant correlations were performed with GraphPad Prism.
USP37 is commonly overexpressed in breast cancer
USP37 has been previously confirmed to be overexpressed in lung cancers cells and tissues [21]. In order to investigate the role of USP37 in tumorigenesis, we examined the breast cancer database of The Cancer Genome Atlas (TCGA) to evaluate the differential expression of USP37 [22]. Analysis of the TCGA database indicated that cancer with USP37 transcripts (n = 517) had a significantly higher expression level than normal samples (n = 29) (P < 0.0001) (Fig. 1a). Similar results were also found via immunohistochemistry analysis (Fig. 1g). To further confirm the overexpression of USP37 gene in breast cancer, we utilized a series of human breast cancer cells (MCF-7, MDA-MB-231, BT549 and T47D) and human normal breast epithelial cells (MCF-10A) to examine USP37 protein expression by western blotting (Fig. 2a). These results confirmed that USP37 was also overexpressed in breast cancer cells compared to normal breast epithelial cells.
For investigation of the correlation between USP37 gene and the breast cancer heterogeneity, we also tested USP37 expression in the four cell subtypes using PAM50 gene expression profiling. First, we observed an starkly different tendency within USP37 gene expression among different pathological subtypes of breast cancer cells, including the normal-like subtype having the lowest, and Luminal B type endowed with the highest expression levels of USP37 (p < 0.0001) (Fig. 1b). We next estimated the effect of USP37 as an oncogenic biomarker for overall survival of patients diagnosed with breast cancer. Clinical data from the TCGA database were divided into two groups according to the differential expression of USP37 gene. The results indicated that cancer with higher expression levels of USP37 was significantly correlated with the elevated rates of mortality (P < 0.05) (Fig. 1c). Samples with high USP37 expression also had shorter survival duration than those with low USP37 expression. Moreover, GSEA analysis showed that high USP37 expression was positively associated with metastasis, cell growth and anti-apoptosis in the TCGA breast cancer samples (Fig. 1d-f ).In brief, these data confirmed that USP37 gene could act as a cancer biomarker in predicting a worse outcome for breast cancer patients. Collectively, these data suggest that USP37 is abnormally overexpressed in human breast cancer patients and cell lines.
USP37 is highly expressed in breast cancer stem cells CD24 − /CD44 + cells and ALDH1 + cells are widely considered to be breast cancer stem cells [24]. To validate the expression levels of the USP37 gene in breast cancer stem cells, we isolated CD24 − /CD44 + cell populations from MCF-7 cell lines by magnetic activated cell sorting (MACS). We detected that USP37 was significantly overexpressed in CD24 − /CD44 + cells compared to their counterpart CD24 + or CD24 − CD44 − cells in regards to mRNA levels (Fig. 2c). Immunofluorescence staining further confirmed the upregulation of USP37 expression in BCSCs (Fig. 2d). Additionally, we used breast cancer cell lines to perform sphere-formation experiments and measured the differential expression of the USP37 gene. We confirmed that protein levels of USP37 were evidently elevated in spherical cells compared to their counterpart adherent cells (Fig. 2b). Taken together, these results indicated that elevated expression of USP37 was enriched in breast CSCs and was a novel feature of breast cancer stem cell-like subpopulation.
USP37 knockdown suppresses breast cancer cell migration and invasion by promoting mesenchymalepithelial transition
Currently, the underlying biological mechanism accounting for the elevated expression of USP37 in breast tumors remains unclear. To investigate the effects of USP37 on breast cancer progression, we used siRNA oligonucleotides to knockdown endogenous USP37 in MCF-7 and MDA-MB-231 cells. USP37 mRNA expression levels significantly decreased with two different siR-NAs treated breast cancer cells compared to levels in the control group cells (Fig. 3b). Moreover, USP37 protein was also reduced significantly with siRNA#2 and siRNA#3 treatment of MCF-7 and MDA-MB-231 cells as evidenced by western blotting (Fig. 3a). These results confirmed that USP37 gene expression could be effectively downregulated by siRNA#2 and siRNA#3.
The expression of USP37 gene has been demonstrated to be elevated in patients with a recurrence of cancer, indicating that USP37 levels may be closely related to breast cancer distant metastasis [16]. In order to elucidate the function of USP37 in the migration and invasion of breast cancer cells, we conducted wound healing and Transwell assays to detect the role of USP37 in cell migration and invasion. Knockdown of USP37 evidently inhibited cell migration capacity in both MCF and MDA-MB-231 cells (Fig. 3f, g). The transwell assay revealed that downregulation of USP37 obviously decreased the invading capacity of MCF-7 and MDA-MB-231 cells. In addition, depletion of USP37 in MCF-7 and MDA-MB-231 cells displayed a Fig. 1 Overexpression of USP37 in human BC was correlated with poor-disease outcome. a Cancer with USP37 transcripts apparently had higher expression (n = 517) than normal breast tissues (n = 29) from the TCGA database [22] (***P < 0.0001). The median and interquartile range are represented by black lines. b USP37 transcripts among different subtypes are divided into five distinct molecular subtypes. The parameters were analyzed with the PAM50 gene expression profiling [22]. The median with interquartile range is represented by black lines. c Breast cancer (n = 738 patients) were evaluated with USP37 mRNA levels and results were correlated with overall survival over 25 years. The red line indicates patients with high USP37 transcript (n = 369) and blue line indicates patients with low USP37 transcript (n = 369). P value was analyzed by Kaplan-Meier analysis using GraphPad Prism. d-f GSEA analysis showed that USP37 expression was positively associated with metastasis (d) and cell growth (e) while negatively related to cell apoptosis (f) in the TCGA breast cancer samples. g The USP37 protein level in breast cancer tissues and surrounding tissues are shown by immunohistochemistry (IHC) (Brown: USP37). Scale bars: 100 μm. h USP37 IHC staining scores in breast cancer tissues (n = 60) and surrounding tissues (n = 60) are shown. **P < 0.01 greater reduction of invasion capacity compared to that of the control cells (Fig. 3h). Epithelial-mesenchymal transition (EMT) is a dynamic process in which epithelial cells acquire enhanced mobility and invasive properties by losing cell-cell adhesion structures and polarity. To further detect the biological mechanism of USP37 on cell migration and invasion, we focused on the identification whether USP37 affected EMT. We examined typical EMT markers (E-cadherin, N-cadherin, Snail1 and Vimentin) by western blotting. As shown in Fig. 3c and d, knockdown of USP37 significantly decreased Snail1, N-cadherin and Vimentin expression, but increased E-cadherin expression levels. These results indicated that downregulation of USP37 induced the mesenchymal-epithelial transition (MET) process in breast cancer cell lines. Similar results were also found via immunofluorescence analysis (Fig. 3e).
Conversely, USP37 overexpression in MCF-7 cells promoted a mesenchymal phenotype with the downregulation of E-cadherin and upregulation of N-cadherin, Vimentin and Snail1 as evidenced by western blotting (Fig. 3i). Immunofluorescence assay showed that overexpression of USP37 further inhibited E-cadherin expression and upregulated N-cadherin expression (Fig. 3j). Furthermore, upregulation of USP37 markedly promoted invasion and migration of breast cancer cells (Fig. 3k and l). Based on this experimental data, we were able to further demonstrate that USP37 is involved in regulating EMT in breast cancer progression and promotes the migration and invasion capacity of breast cancer cells.
USP37 is essential for the maintenance of cell self-renewal and chemo-resistance Since USP37 was highly expressed in breast cancer stem cells, we further examined whether USP37 played a role in breast cancer stem cells behavior. We knocked down USP37 in MCF-7 and MDA-MB-231 cells via infection with lentivirus expressing USP37#2 shRNA or control lentivirus (shScramble); we confirmed USP37 gene expression by examining protein levels (Fig. 4a, b). The Hh pathway is considered to control the self-renewal of CSCs in breast cancer [25]; therefore, we tested the expression of representative Hh signaling factors (smoothened and Gli-1) and main stem cell markers (ALDH1 and OCT4) in USP37 knockdown cells and control cells by western blotting analysis. We observed that expression of smoothened, Gli-1, ALDH1 and OCT4 were decreased in USP37 knockdown cells compared with their counterpart cells, indicating that inhibited stemness was accompanied by a decrease in USP37 expression ( Fig. 4a and b). In addition, USP37 knockdown significantly inhibited the formation of spheroids, as well as their size and volume, compared to the control cells in MCF-7 and MDA-MB-231 cells (Fig. 4c and d). With regard to chemical sensitivity, the cytotoxic effects of different concentrations of cisplatin on USP37 knockdown cells and their respective control cells were detected after 48 h treatment of cisplatin by CCK-8 assays. The results indicated that USP37 knockdown cells had increased sensitivity to cisplatin-induced growth inhibition and as evidenced by decreased colony formation compared with the control MCF-7 and MDA-MB-231 cells (Fig. 4f-i, k and l). Furthermore, the expression of Smo, Gli-1, ALDH1 and OCT4 were elevated in USP37 overexpression plasmid transfected MCF-7 cells compared with the control cells (Fig. 5a). USP37 overexpression also increased the formation of spheroid and chemo-resistance to cisplatin-induced growth inhibition (Fig. 5b, d).
To further explore the mechanism involved in cisplatin-induced apoptosis caused by USP37 knockdown, we detected apoptosis-related markers Bcl-2, Bax and cleaved caspase-9 in treated cells. We observed the upregulation of Bax and cleaved caspase-9 and downregulation of Bcl-2 in MCF-7 cells and MDA-MB-231 cells after transfection with USP37 shRNA and/ or treatment with cisplatin for 48 h (Fig. 4e, j). In the Bcl-2 protein family, the Bcl-2/Bax ratio has been shown to affect apoptosis induction [26,27]. Our data indicated that USP37#2 shRNA combined with cisplatin induced the cell apoptosis with an underlying decrease in the Bcl-2/ Bax ratio. Consistent with our hypothesis, an increase in the Bcl-2/Bax ratio might be involved in anti-apoptosis activity induced by USP37 upregulation treatment with cisplatin in MCF-7 cells (Fig. 5c). These data demonstrated that USP37 knockdown induced drug sensitivity by dislocating intracellular apoptosis-related proteins.
Downregulation of USP37 inhibits stemness, cell invasion and EMT via hedgehog signaling pathway in breast cancer
It has been verified that tightly controlled Hh pathway ensures proper development and averts tumor formation in the mammary gland [10]. In ER-positive breast cancer, Hh pathway contributes to the maintenance and regulation of CSC invasion and EMT [28]. We were able to further determine the effect of USP37 on the stemness, cell invasion and EMT via Hh signaling pathway.
Purmorphamine (PM) is considered to be an Hh agonist.
Smo, Gli-1 and USP37 expression post-PM treatment was examined by western blotting analysis and immunofluorescence analysis. Interestingly, our data indicated that there was a time-dependent increase in the expression of Smo, Gli-1 and USP37 expression following 24 h, 48 h PM treatment (0.5 μM) (Fig. 6a). Immunofluorescence assays showed similar results (Fig. 6b). These findings are further supported by the ability of USP37 to participate in Hh signaling pathway in breast cancer. Figure 6c shows that silencing of USP37 expression could reverse the positive effects of PM on the Hh pathway. Moreover, the effect of downregulating USP37 on CSC traits, including the formation of spheroid, BCSCs markers and cell invasion, was impaired after PM (See figure on previous page.) Fig. 4 Downregulation of USP37 weakens cell self-renewal and chemo-resistance to cisplatin. a, b MCF-7 and MDA-MB-231 cells were transfected with shUSP37#2 and shScramble, and expressions of USP37, Smo, Gli-1, ALDH1 and OCT4 were confirmed by western blotting. *P < 0.05, **P < 0.01. c, d Comparison of mammosphere formation among inoculated cells in MCF-7 and MDA-MB-231 cells (original magnification, 4× or 20×).*P < 0.05, ***P < 0.001. e MCF-7-ShScramble and MCF-7-shUSP37#2 cells were treated with cisplatin (1 μg/ml) for 48 h, then treated cells were harvested to detect the levels of Bcl-2, Bax and cleaved caspase 9. GAPDH was examined as a loading control. j MDA-MB-231 cells transfected with shScramble or shUSP37#2 were exposed to 3 μg/ml cisplatin for 48 h, and the levels of Bcl-2, Bax and cleaved caspase 9 were detected via western blotting. f, g, k, l CCK-8 assay and colony formation assay showed that cell viability was decreased in MCF-7 and MDA-MB-231 cells after USP37 knockdown. Quantitation of colony formation in (h) MCF-7 and (i) MDA-MB-231. GAPDH was examined as a loading control. *P < 0.05, **P < 0.01, ***P < 0.001 Fig. 5 Overexpression USP37 elevates cell self-renewal and chemo-resistance to cisplatin. a Expressions of USP37, Smo, Gli-1, ALDH1 and OCT4 were detected in MCF-7-CTL or MCF-7-USP37 cells by western blotting. *P < 0.05, **P < 0.01. b Comparison of mammosphere formation among MCF-7-CTL or MCF-7-USP37 cells. *P < 0.05. c MCF-7-CTL and MCF-7-USP37 cells were treated with cisplatin (1 μg/ml) for 48 h, then treated cells were harvested in order to detect the levels of USP37, Bcl-2 and Bax. GAPDH was examined as a loading control. *P < 0.05. d CCK-8 assay showed that cell viability was enhanced in MCF-7 after USP37 upregulation compared to the control group. *P < 0.05, **P < 0.01, ***P < 0.01 (Fig. 6c, d and e and Additional file 2: Figure S1). Silencing of USP37 expression induced MET including upregulated E-cadherin expression and downregulated N-cadherin expression, while PM reversed MET in silenced USP37 cells. Similar results were observed with immunofluorescence assay (Fig. 6f). Collectively, these data indicate that downregulation of USP37 expression was involved in the attenuation of the BCSCs stemness, cell invasion and EMT by suppression of the Hh pathway.
USP37 interacts with and stabilizes Gli-1 protein
Given that USP37 expression was significantly associated with the Hh pathway, we detected whether endogenous USP37 could also regulate Gli-1 stability. Endogenous USP37 was depleted by lentivirus expressing USP37#2 shRNA in MCF-7 and MDA-MB-231 cell lines. The effect of USP37 knockdown on Gli-1 was evidently reversed by MG132, which is a protease inhibitor ( Fig. 7a and b). Furthermore, knockdown of USP37 expression significantly decreased the stability of Gli-1 protein both in MCF-7 and MDA-MB-231 cells (Fig. 7d and e). Additionally, the cycloheximide (CHX) chase assay was used to assess MCF-7 breast cancer cell line for the effect of USP37 upregulation on endogenous Gli-1 protein. As shown in Fig. 7f, upregulation of USP37 expression consistently stabilized Gli-1 protein. Immunofluorescence staining assays showed that USP37 could regulate Gli-1 expression (Fig. 7c). Co-IP assay indicated that USP37 interacted with Gli-1 (Fig. 7h). We also found overlapping expression of exogenous USP37 and Gli-1 by immunofluorescence staining (Fig. 7g). Together, these data suggest endogenous USP37 may regulate Gli-1 protein stability in breast cancer.
USP37 knockdown inhibits tumorigenicity and increases sensitivity to cisplatin in vivo
To further assess the anti-tumor effect of USP37 downregulation on tumor growth and cisplatin sensitivity in vivo, we established a xenograft tumor models by subcutaneously injecting breast cancer MCF-7 cells, which had been transfected with USP37#2 shRNA or scramble shRNA, into the mammary gland of nude mice (Fig. 8a). Tumor volume and weight measurements were analyzed. Compared with the scramble shRNA control group, the volume and weight of USP37 shRNA group increased at slower rate (Fig. 8b, c). The average tumor volume and weight of cisplatin-treated USP37#2 shRNA tumors was evidently smaller than that of cisplatintreated tumors or USP37#2 shRNA tumors. We found that lower expression level of USP37 obviously impaired the expression of Hh targets (Smo and Gli-1) and cell proliferation marker Ki-67 in the tumor tissues as seen by immunohistochemical staining (Fig. 8e). Consistent with these results, tissues formed from cisplatin-treated USP37 knockdown cells showed the lowest expression levels of Smo, Gli-1, ALDH1, OCT-4 and Bcl-2 than other tissues by western blotting analysis (Fig. 8d). These results indicated that USP37 downregulation attenuates breast cancer progression and enhances sensitivity to cisplatin in vivo.
Discussion
DUBs have been shown to participate in ubiquitin cleaving from ubiquitin conjugated protein substrates [29,30]. USP37, a novel DUB, contains an insert between its catalytic lobe and its ubiquitin-binding lobe while its function could prevent 14-3-3γ degradation, which might contribute to malignant transformation by MAPK signaling [31,32]. Subsequently, it was verified that USP37 expression modulated the oncogenic fusion protein PLZF/RARA stability and cell transformation potential in PLZF/RARA-associated acute promyelocytic leukemia [33]. Recently, clinicopathological analysis confirmed that USP37 was a poor prognostic factor in breast cancer [16]. However, there was no direct evidence to identify the carcinogenic mechanism of the USP37 gene in breast cancer. In this study, we have demonstrated overexpression of USP37 in Luminal B subtype breast cancer was a predictor of poor outcomes in breast cancer. Supporting our results, previous clinicopathological analysis has demonstrated that overexpressed USP37 is considered to be a poor prognosis in breast cancer [16]. Additionally, GSEA analysis on the TCGA dataset implicated that USP37 expression was positively associated with metastasis, cell growth and anti-apoptosis. Therefore, USP37 levels could potentially serve as a specific oncogene involved in breast carcinoma progression. Cancer stem cells (CSCs) are endowed with stem cell properties, which include the ability to self-renew and differentiate by symmetrical or asymmetrical cell division. CSCs self-renewal of the cellular population and generation of progenitor cells have been shown to resist radiation and chemotherapy. Therefore, these cells are commonly deemed to be crucial target for cancer therapy [34,35]. It was verified that CSCs could be maintained in an undifferentiated status and induce tumor-sphere formation in defined serum-free medium [36]. However, limited markers were utilized for the identification of breast cancer stem cells. While researchers have demonstrated that CD44 and ALDH1 are critical biomarkers to identify BCSCs from breast cancer populations [37,38], our data suggested that the USP37 gene was significantly associated with CSC properties, such as self-renewal, treatment resistance and EMT phenotype as well. Interestingly, elevated mRNA expression of USP37 was detected in CD24 − /CD44 + cells and ALDH1 + cells; the protein expression of USP37 was obviously elevated in spheres compared to adherent cells. Knockdown of USP37 suppressed mammospheres formation and inhibited cancer stem markers, such as ALDH1and OCT4 in breast cancer cells. It has been reported that USP37 participates in regulation of CSCs-related proteins, such as SOX2 and c-myc [20,21]. Our data further indicated that knockdown of USP37 by constructed lentiviral system weakened the stemness in MCF-7 and MDA-MB-231 cells via inhibiting the expression levels of ALDH1 and OCT4. CSCs are considered to be "bad seeds" due to their drug resistance caused by imbalanced pathway and epigenetics in cancer [39]. As shown above, we detected that knockdown of USP37 in breast cancer cells could promote the sensitivity to cisplatin-induced cell death by the CCK-8 assay and colony formation assays. The Bax/ Bcl-2 ratio is a commonly used method to determine whether intracellular apoptosis system is activated [40]. We detected that USP37#2 shRNA combined with cisplatin treatment induced cell apoptosis with an underlying decrease in Bcl-2/Bax ratio. On the contrary, the upregulation of USP37 reversed this phenomenon. USP37 knockdown suppressed stemness and chemoresistance, which may assist clinical oncologists in designing and testing novel therapeutic strategies. In summary, we suggest that USP37 gene expression confers the stemness and potentially acts as a critical marker of CSCs in breast cancer.
The process of EMT was involved in the acquisition of aggressive cellular traits, including motility, invasiveness and anti-apoptosis, resulting in the dissemination of cells and colonization in distance tissues [6,41,42]. Here, we showed that USP37 siRNA treatment stimulated the expression of epithelial marker (E-cadherin) but decreased the expression of three known inducers of EMT (Snail1, N-cadherin and Vimentin) (Fig. 6a). Moreover, we found that knockdown of USP37 suppressed cell migration and invasion in breast cancer cells. On the contrary, upregulation of USP37 promoted EMT, migration and invasion. Accumulation evidence has indicated that tumor cells undergoing EMT process are endowed with the trait of cancer stem-like cells [43], which further speculated USP37 as a CSC marker of breast cancer. In summary, our data showed that USP37 could regulate the migration, invasion, EMT of breast cancer cells.
The Hh pathway was found to be required for the maintenance of breast CSCs traits, tumor formation and the EMT [10]. Normally, the ligand binding of Patched (Ptch1), a 12-pass transmembrane receptor of Sonic Hedgehog (SHH), can activate zinc finger transcription factor (Gli-1). It has been suggested that ectopic expression of Gli-1 upregulates expression of the transcription factor Snail1 accompanied with a decrease in E-cadherin, a characteristic of EMT [24]. In fact, Snail expression promotes EMT via repressing E-cadherin [44]. Hh signaling pathway modulates EMT indirectly via Snail [45]. Therefore, it is a promising research to investigate whether USP37 promoted the activation of the Hh pathway.
Our findings suggested that the post transcriptional levels of Smo and Gli-1 were decreased in vitro and in vivo. Upregulation of USP37 could enhance protein expression levels of Smo and Gli-1. In order to understand the mechanism of USP37-induced EMT and stemness via the Hh pathway, we treated MCF-7 cells with PM. We found that activation of Hh signaling pathway was accompanied by elevated expression of USP37 gene as visualized by western blotting and immunofluorescence assay. Furthermore, the effect USP37 downregulation on CSCs traits including the formation of spheroids, BCSCs markers and cell invasion were impaired after PM treatment. Meanwhile, knockdown of USP37 reversed the effect of PM on (See figure on previous page.) Fig. 8 USP37 knockdown inhibits tumorigenicity and increases sensitivity to cisplatin in vivo. a, b, c MCF-7 cells transfected with shScramble or shUSP37#2 were subcutaneously injected into BALB/c nude mice treated with 0.9% saline or cisplatin. Tumors were excised from nude mice at day 28. a Images of tumor xenografts and b tumor weight for tumor xenografts treated with 0.9% saline or cisplatin are shown above. *P < 0.05, **P < 0.01. c Tumor growth curves were drawn every other day up to day 25. d Western blotting analysis of USP37, Smoothened, Gli-1, ALDH1, OCT4 and Bcl-2 in tumor xenografts. GAPDH was used as a loading control. *P < 0.05, **P < 0.01. e HE staining and immunohistochemical detection of USP37, Smoothened, Gli-1, Ki-67 in tumor xenografts. Scale bar: 100 μm the EMT markers. These data confirm that USP37 mediates breast cancer stem-like properties, cell invasion and EMT via the Hh pathway. Gli-1 acts exclusively as a transcriptional activator in the Hh pathway and signal outcomes are determined by the balance of activated and inhibitive Gli proteins [46]. Remarkably, we found that USP37 could regulate and stabilize the protein level of Gli-1 (Additional file 2 : Figure S1). As a deubiquitinase, USP37 is involved in the regulation of multiple proteins by deubiquitination, including P27, Cdt1, PLZF/RARA and 14-3-3γ [31,33,47,48]. Future work should aim to determine whether USP37 could stabilize Gli-1 through deubiquitination.
Conclusions
In conclusion, our research demonstrated that USP37 was highly expressed in breast CSCs and was correlated with poor prognosis in breast cancer patients. Knockdown of USP37 expression hampered cell invasion, stemness, EMT and also resulted in the drug sensitivity to cisplatin. Mechanically, USP37 could orchestrate the stemness, cell invasion and EMT via activation of Hedgehog pathway. Further studies detected that USP37 also interacted with and stabilized Gli-1 protein, which is the main activator of Hedgehog target gene. These results indicate that USP37 is essential in the regulation of breast cancer progression suggesting an experimental basis for its use as a cancer biomarker and expands its potential clinical value.
Additional files
Additional file 1: Table S1. Sequences of primer, siRNA and shRNA. (DOCX 14 kb) Additional file 2: Figure S1. Protein level of OCT4 as detected by western blotting after the NC siRNA group or the USP37 siRNA#2 group was treated with 0.5 µM purmorphamine for 48 h. GAPDH was examined as a loading control. (DOCX 80 kb) | 8,366.4 | 2018-11-27T00:00:00.000 | [
"Biology",
"Medicine"
] |
Leveraging Global Partnerships to Achieve SDG 7: A Case Study of Pakistan
Energy is fundamental to socio-economic growth for the world; but how this energy is generated and distributed would determine whether the world could achieve a sustainable future. There is no denying the fact that energy is one of the primary sources of greenhouse gas (GHG) emissions. Thus, addressing the climate crisis, sustainable energy lies at the core of both the United Nations’ Sustainable Development Goals (SDG) and the Paris Agreement. Among other goals, SDG 7 calls for affordable, reliable, sustainable and modern energy for all by 2030 (UNDP, 2017). This work utilizes the guidelines established by the Global Tracking Framework on SDG 7 to assess the interventions needed to meet SEforALL targets for Pakistan using a bottom-up approach. The Global Tracking Framework is a joint effort of multilateral development agencies to track the world’s progress towards the three SEforALL goals to be achieved by 2030 (ESMAP, 2016). In this process, national and provincial stakeholders were identied to provide their inputs on creating a reliable, affordable and sustainable energy market. This study on policy with and governments, interviews of private sector stakeholders, review of plans and policies of different energy sector institutions, nancial and technical analysis of nancial institutions, and evaluation of international donor agencies programs in As an a comprehensive and integrated National Action Plan (NAP) was developed with technical and nancial support of United Nations Development Program to accelerate transition towards SEforAll goals for Pakistan in short, medium and long term goals
Progress towards addressing the climate change has been modest over the past decades despite the ever-increasing signi cance for action against global warming. Climate change is fundamentally driven by the emissions of Green House Gases (GHGs) into the atmosphere, such as carbon dioxide (CO2), methane (CH4) and nitrious oxide (N2O) (US EPA, 2016). Therefore, there is a need to halt anthropogenic climate change with radical transformation of GHG emitting sectors. Worldwide, energy sector that drives the global economy is responsible for over 70% of global GHGs (Mengpin & Johannes, 2020). The energy sector includes transportation, electricity and heat, buildings, manufacturing and construction, fugitive emissions and other fuel combustion. Thus, it necessitates that for drastic transformational change in the energy system with a vision to achieve sustainable development and to improve the well being of billions who lack access to reliable and clean energy to meet their basic needs.
Energy plays a critical role in enabling sustainable development, as highlighted at the Rio+20 Sustainable 1. Ensuring universal access to clean and reliable energy Three billion people lack access to either electricity or clean fuels for cooking, or both; this has severe, adverse implications for human health (WHO, 2009). In practice, ensuring universal access means providing electricity to remote and poor rural areas, as well as the substitution of traditional biomass such as solid fuels or rewood by cleaner and modern energy services and appliances using e cient cookstoves, natural gas, or lique ed petroleum gas (LPG).
Doubling the share of renewable energy in the global energy mix
This requires increasing the share of renewables in global nal energy mix, which is nal energy available to actual users from 15% to 30% by 2030 (IRENA, 2017). For example, the wind, solar, hydropower, biomass and geothermal power are all sources of renewable energy.
3. Doubling the rate of improvement in energy e ciency Energy e ciency of the global economy is interpreted as a goal to reduce the amount of energy required to provide products and services. This objective has been translated to an average improvement rate for global energy intensity, which is measured in units of nal energy per gross domestic product of a nation.
More speci cally, this signi es the target of achieving a global energy intensity improvement by an average rate of 2.4% per year between 2015 and 2030 in comparison to historical rate of 1.2% annually (Djaheezah, 2017).
Sustainable Energy For All (seforall) Initiative In Pakistan
This work utilizes the guidelines established by the Global Tracking Framework on SDG 7 to assess the interventions needed to meet SEforALL targets for Pakistan using a bottom-up approach. The Global Tracking Framework is a joint effort of multilateral development agencies to track the world's progress towards the three SEforALL goals to be achieved by 2030 (ESMAP, 2016). In this process, national and provincial stakeholders were identi ed to provide their inputs on creating a reliable, affordable and sustainable energy market.
This study is based on consultations, policy dialogues with federal and provincial governments, interviews of private sector stakeholders, review of plans and policies of different energy sector institutions, nancial and technical analysis of nancial institutions, and evaluation of international donor agencies programs in Pakistan. As an outcome, a comprehensive and integrated National Action Plan (NAP) was developed with technical and nancial support of United Nations Development Program (UNDP) to accelerate transition towards SEforAll goals for Pakistan in short, medium and long term goals (UNDP, 2019).
Energy is one of the most important contributors that drives economic growth for Pakistan. Access to clean and reliable sources of energy is essential to achieve desired socio-economic development in Pakistan. Pakistan is blessed with an enormous amount of renewable energy resources. The country has an excellent solar insolation; an extremely good wind energy potential; 1054 KM long coastal line with a resource of tidal energy; large livestock population and agricultural waste to generate bioenergy (AEDB, 2015). Despite the massive energy resource potential, energy infrastructure in Pakistan is still underdeveloped and energy resources remain underutilized. The existing energy infrastructure is insu cient to meet the energy needs of the people, particularly those living in the remote locations (Ichord, 2020).
As a developing country, Pakistan faces a number of challenges related to overcoming poverty and improving the health, education, and employment opportunities for its large and lower income population. The energy sector is vital in addressing these problems, however, the major constraints are inadequate investment, unreliable energy supplies, weak governance, and poor scal management.
Pakistan's energy sector remains one of the main obstacles to economic growth. Although Pakistan has managed to increase power generation since 2013 and mitigate power blackouts that plagued the country over the past decade. After spending decades tackling electricity shortages, Pakistan now faces a new and unfamiliar problem: too much generation capacity (Mangi, 2021). In past decades, the country's energy sector relied on expensive fuel sources mostly on imported energy products, chronic natural gas and electricity shortages, major debt in the power sector, and aging and insu cient transmission and distribution systems have prevented the sector from growing and modernizing.
Objective 1: Access to Modern and Clean Energy in Pakistan
There are many variants and de nitions of energy access. However, we take the energy access as percentage of the households with an access to either electricity connection or gas pipeline network. It is also important to understand that access to the grid or piped network does not guarantee an adequate or reliable supply of energy because many households face intermittent or low voltage/pressure of supply of energy.
In Pakistan, 73% of population has an access to grid electricity (Syed Aziz, et al., 2019). This means 55 million people or 8.8 million homes with an average household size of 6.354 are living without access to electricity. Insu cient generating capacity and lack of grid network for rural areas are one of the major challenges to electricity access. There are 32,266 villages in the country which will remain without grid access (NEPRA, 2020). These villages have sparsely distributed population and are in remote locations which is making expansion of grid nancially unviable and technically challenging.
It is pertinent to note that the connection to the grid does not equate to the availability of electricity. Most of the villages o cially listed as electri ed, continue to experience long hours of blackouts; thus being forced to spend more than half of the day without electricity. According to the Solar Consumer Perception study conducted by IFC Lighting Pakistan program (2015), 73% of the country's population experiences few hours of blackouts. Thus, SEforALL goal is to achieve 100% electricity access in all the regions and to ensure 24 hours of uninterrupted supply of electricity to the customers.
In terms of access to natural gas network, only 25% of population has an access to natural gas network. This indicates that more than two-third of population have to rely on ine cient and unhealthy recourses such as fuel wood, dung and other biomass to meet their basic energy needs for cooking, space heating and water heating etc. Thus, it requires scaling up the off grid-solutions and other renewables in large part to bring energy access to millions of people. The high capital cost of laying gas pipelines and due to depleting reserves of natural gas, households continue to switch to more expensive LPG cylinders (transported from down-country) or those who cannot afford LPG ful ll their heating and cooking requirements through rewood and coal.
Traditional fuels like rewood, dung and residues currently contribute a major share in meeting the everyday energy requirements of rural and low-income urban households in Pakistan. Our analysis demonstrated that in Pakistan, almost 60% of the country's population uses wood for cooking, followed by gas (22%), crop residuals (11%) and dung (7%).
Scaling up of of renewable energy technologies would be the most preferred and cost effective route to ensure improved energy access in rural or remote areas of Pakistan, as it eliminates the need of extending new power and gas network to those far off areas which would ultimately result in huge savings in capital expenditures.
Worldwide, the costs of these clean energy technologies are constantly declining and their e ciencies are improving; some of these technologies have already reached grid parity. Given this trend of improved economics and technical viability, renewable energy technologies are the most suitable options for generating clean electricity for both the large scale -grid connected plants and also an excellent choice for the off-grid applications.
Also, it is uneconomical to connect these remote areas to electricity grid mainly because of low energy demand in remote regions where population density is low, towns/villages are located far away from each other and most importantly, the limited paying capacity of people living in those areas. Therefore, we have to rely on solar and other renewables in large part to bring energy access to millions of Pakistanis.
Objective 2: Doubling the share of renewable energy in Pakistan Pakistan has tremendous potential to harness renewable energy, which includes wind potential of over 50,000 MW (Renne, 2007) in several wind corridors, and an exceptionally good solar resource with an annual average Global Horizontal Irradiance (GHI) of 2322 kWh/m2 [1] and; large livestock population and agricultural waste to generate bioenergy.
However, Pakistan has been transitioning at a much slower pace in terms of renewable capacity additions of solar and wind both at utility and distributed generation scale. Although hydropower has traditionally been the most prominent source of renewable energy in Pakistan -making up almost a third of electricity generation with current installed capacity of 9.8 gigawatts (GW) in 2020 (NEPRA, 2020).
Despite this massive resource potential, cumulative contribution of solar, wind and biomass stood less than 3% [2] in total electricity generation with current installed capacity of 1248 MW wind, 530 MW solar and 369 MW in 2019-20 (NEPRA, 2020).
Seemingly, Pakistan has undertaken a number of policy interventions to promote renewables in the country. With promulgation of its rst ever Policy to promote Renewable Energy in 2006; outlining incentives to spur green growth -adding up roughly 2,000 MW of solar and wind to the national grid over the course of 14 years. Apparently, these are not impressive numbers and do not correspond with growth potential of renewables that exist in the country.
Pakistan's total installed power generation capacity stands at 39,000 MW, of which 66% of energy comes from fossil fuels, 24% from hydro, and 6% from renewable sources (wind, solar and bagasse) and 4% from nuclear. In the current scenario, renewable energy (RE) resources can play an important role in closing the de cit. Government of Pakistan has set the target of 30% of total power generated from renewable energy sources by 2030 (AEDB, 2019).
SEforALL target for Pakistan outlines doubling the share of renewable energy in the total nal energy consumption (TFEC) from 7% in 2018 to over 15% by 2030. TFEC is the secondary energy that is received after the transformation processes and related losses, e.g. electricity, petrol, heat and natural gas if used for heating and cooking. The current share of renewables in total nal energy consumption in Pakistan is about 7% and it needs to be doubled to 14% by 2030. In order to meet the policy objectives of achieving 30% of renewables and SEforALL target, solar and wind capacity additions would have to increase to 24,000 MW by 2030 from the existing capacity of 902 MW in 2018. Another important consideration is the rapid innovation in technology, which is driving down the combined cost of renewables plus electricity storage at a signi cant rate, while improving their conversion e ciencies. If RE market continues to achieve such a drastic level of reduction in costs and improvements in battery technology, this would be a game changer especially for the off-grid and unelectri ed regions of the country.
Objective 3: Doubling the Rate of Improvement of Energy E ciency Rising costs, increasing energy demand and declining resources are one of the key reasons, the world is on the path towards achieving higher energy e ciency. The energy intensity of some developed countries is already declining due to increased measures in the energy e ciency. The capital cost of new energy supply is often higher than the cost of investing in energy e ciency improvements. Therefore, increased energy e ciency measures can be one of the easiest and least-cost pathways for Pakistan to reduce the demand-supply gap. There is huge potential for energy e ciency in Pakistan as substantial amount of energy is wasted in supply side (production, transmission, distribution) and also in the demand side (consumption) of various energy items like oil, gas and electricity in the various sectors of economy. Energy intensity measures the e ciency of an economy in which a given country converts energy into production. It is expressed as the ratio of total energy consumption per unit of economic output or Gross Domestic Product. Countries with a high level of energy intensity use more energy to create a unit of GDP than countries with lower levels of energy intensity. Whereas, the energy e ciency is the ratio of energy input per unit output, typically measured in physical terms and relatively challenging to measure across various countries. Thus, using energy intensity as an indicator for energy e ciency works well when making country-wide comparisons.
Energy intensity measured is in uenced by a variety of factors, including the e ciency of underlying processes, factors such as changes in sectoral structure of GDP, annual variations in weather, climatic conditions and geographical sizes of the countries. It is important to note that the energy intensity of GDP actually measures energy productivity and it is not a de nitive indicator of energy e ciency from a technical perspective, as it takes into account various effects and factors that are not directly linked to energy e ciency of a country.
The target for energy e ciency under SEforALL is to double the rate of improvement of energy e ciency. Therefore, it is important to identify and assess the current rate of improvement of energy e ciency which has to be doubled by 2030. Currently, the rate of improvement of energy e ciency has been improving at the rate of 1.7% annually since 2000 as shown in the table 1 (i.e., 27% over the last 14 years) (EnerData, 2014). In other words, the primary energy intensity of Pakistan has decreased by 1.7% annually since 2000. In order to achieve the SEforALL target, this rate of improvement needs to be doubled by 2030, which means that there has to be reduction in the primary energy intensity by 3.4% annually or by about 50% over the next 14 years.
Broadly, the SEforAll NAP underscores these high impact action areas, in order of following priority:
Investment needs to nance SEforALL
There is a global consensus that nancing needs to achieve SEforALL target will be huge. Morgan Stanley (2020) estimates the nancing cost to achieve SDGs could be over $50 trillion mark. Meeting SDGs will require the global community to increase development nancing from "billions" to "trillions," which implies a substantial nancing gap (International Finance Cooperation (IFC, 2019). McKinsey estimates required global infrastructure spending to be on the order of $6 trillion a year through 2050. The Global Commission on Adaptation (2019) concluded that adapting to the climate change that cannot be avoided will be a multi-trillion-dollar expense. A recent analysis in Carbon Tracker (2020) estimates that replacing the global fossil fuel infrastructure will cost about $22 trillion.
Providing energy access requires considerable capital expenditures in building the infrastructure to support energy supply, as well as on-going operations and maintenance. We identify that overall nancing requirement to meet SEforALL goal in Pakistan by 2030 across renewable energy, energy e ciency and universal access -is estimated to be over US$ 66 billion.
For access to clean energy and reliable energy, we underline that millions of people in Pakistan live in rural communities with no grid connection. They are widely dispersed with very low levels of density. So, reaching these households is capital intensive and time consuming. On top of it, there is a challenge affordability, which means project developers are unable to charge full cost-recovery prices for the infrastructure, and services. This means that in order to achieve universal energy access, off-grid solutions must be scaled up to a certain level so that the projects become nancially viable and economical. We identify high impact areas to improve clean energy access in Pakistan with investment requirement of about US$ 30 billion as illustrated in the Table 3 below. While progress is being made to scale-up nancing, current annual nancing required to achieve these targets are signi cantly lower in Pakistan. We identify that there is a considerable annual energy nancing gap with billions of dollars, and the available volumes of nance is not su cient to meet set targets. This is in line with global trends which underline a widening gap between required and actual investment to achieve universal energy access in high-impact countries in sub-saharan Africa and Pakistan (Sustainable Energy for All, 2020). Hence, interventions from global nancial institutions must be targeted to support and mobilise creation of viable commercial markets, as the scale-up of private sector nancing will play a central role in nancing.
A further challenge for nancing SEforALL objective is high-cost nancing for low carbon technologies.
Given low-carbon energy's upfront capital intensity, low-carbon energy is penalized from high nancing costs compared to the conventional energy projects. Such high nancing costs can re ect a range of lowcarbon energy investment risks that exist in early-stage markets. Debt and equity capital provider price these risks into their cost of nancing. These kind of barriers limit the availability of capital and nancing opportunities to promote sustainable energy markets in developing countries.
Financing for sustainable energy involves many stakeholders, including public and private, domestic and international. Public actors include domestic governments and international actors (bilateral and multilateral agencies, development banks, and climate funds). Private nance involves a wide range of players: businesses, banks, capital markets, institutional investors and philanthropy. However, private nancing for renewable and large scale energy infrastructure projects have not been so prominent in Pakistan.
Renewable energy nancing requirements to meet SDG 7 by 2030 are estimated at US 18 billion as illustrated in the Table 4 below. It is important to recognize that many of the developed nations today made rapid industrialization, which have been primarily fueled by coal and other fossil fuel in their early stages of growth and development. Thus, they have an important role in supporting the deployment of sustainable energy interventions in the developing countries.
It requires transfer of resources, funds, and technological knowledge from developed countries is a crucial part of sustainable energy deployment in the developing world, and must be a collaboration involving research, business, and government organizations. So industrialized nations not only have to accelerate their actions, but they must also support the developing countries as they make the transition to low-carbon economic growth. Table 6 presents the high impact areas with potential investment for energy e ciency towards SEforALL. The energy e ciency and renewable energy interventions mentioned above will clearly not occur without mobilizing the adequate nancial resources. An effective implementation of the SEforALL objectives does not necessarily mean the cost-optimal path to climate protection; and in terms of who pays for the transformational change required globally, there is still no clear consensus, particularly among the industrialized world. How the investment burden for SEforALL objective has to be distributed over countries and regions is not essentially a scienti c question, but rather a political one.
Conclusion
While addressing the climate crisis, sustainable energy lies at the core of both the United Nations' Sustainable Development Goals (SDG) and the Paris Agreement. SDG 7 particularly calls for affordable, reliable, sustainable and modern energy for all by 2030. For developing countries like Pakistan, it is critical to meet SEforALL objectives in order to achieve sustainable economic development. Achieving the three SEforALL targets could put entire world on a path towards global climate protection. However, meeting those targets would require massive investment.
As a developing country, Pakistan faces a number of challenges related to overcoming poverty and improving the infrastructure, health, education, and employment opportunities for its large and lower income population. The sustainable energy sector is vital in addressing these problems, however, the major constraints are inadequate investment, unreliable energy supplies, weak governance, and poor scal management.
Many of the developed nations today made rapid industrialization, which have been primarily fueled by coal and other fossil fuel in their early stages of growth and development. Thus, industrialized nations have an important role to play in supporting the deployment of sustainable energy interventions in the developing countries.
We determine the investment needs of over US$ 66 billion for Pakistan to achieve its SEforALL objectives. It is clear that Pakistan like many other developing countries would not be able to make this huge investment on its own and requires leveraging external support from the developed and industrialized nations. Hence, Pakistan would require substantial amount of nancial support from the industrialized nations to mobilize enough nancial resources to materialize the SEforALL objectives.
China is expanding its in uence in the region and has made Pakistan a top focus of its Belt and Road Initiative through CPEC, with its heavy investment in power and infrastructure projects. Given the growing climate change impacts and stresses on water resources, it is critical for Pakistan to elevate both mitigation and adaptation in its policies and budget and investment priorities.
Therefore, transfer of resources, funds, and technological knowledge from developed countries is a crucial part of sustainable energy deployment in the developing world, and must be a collaboration involving research, business, venture capital, and government organizations. Hence, industrialized nations or developed world not only have to accelerate their actions, but must also support the developing countries as they make the transition to low-carbon economic growth.
Declarations
Funding | 5,492.2 | 2021-04-20T00:00:00.000 | [
"Environmental Science",
"Political Science",
"Economics",
"Business"
] |
Double Diffusion in Square Porous Cavity Subjected to Conjugate Heat Transfer
Double diffusion inside a square porous cavity being subjected to conjugate heat transfer arising because of the square solid block at the center is studied under homogeneous properties. The double diffusive equations along with the conjugate effect due to square block are subjected to solution through finite element method (FEM). The right vertical surface of cavity remains at cool temperature (Tc) and concentration (Cc), whereas the hot temperature (Th) and higher concentration (Ch) is maintained at the left surface of the cavity. The investigated parameters considered are the Rayleigh number (Ra), Thermal conductivity ratio (Kr), Lewis number (Le), Buoyancy ratio (N) and the height of solid block (B). The outcomes are presented in terms of contours of concentration, isotherms and streamlines, along with local Sherwood number (Sh) and Nusselt number (Nu). The fluid is found to rush around the top left corner of solid. The mass transfer has sharp rise around the top left corner of the solid block.
INTRODUCTION
The double diffusion has drawn the attention of several researchers due to its immense importance and vital applications in the field of science and engineering. Prominent authors [1][2][3][4][5][6] have discussed thoroughly the behaviour of fluid flow and heat transfer encompassing various relevant issues pertaining to porous media. Some of the studies like heat transfer, mass transfer, viscous dissipation, magneto hydrodynamic etc. are studied extensively. The literature reveals that various heat transfer characteristics of porous cavities have been investigated and reported by many researchers. For instance, the heat transfer has been investigated in porous cavities and porous annulus by applying hot temperature at left vertical wall and maintaining cold temperature at the right wall [7][8][9][10][11][12][13]. This included the investigation of heat transfer along with difficult boundary conditions or various other phenomenon such as viscous irregular cavity [7], viscous dissipation [8][9][10], Soret and Dufor effect [11], heat and mass transfer [12] , the effect of heat generating strip placed inside the porous cavity [13], wavy tube heat exchanger [14], flow of nanofluid inside the porous media [15][16][17] etc.
The addition of any solid or subtraction of some section of porous domain from square cavity can affect the heat transfer quite substantially. For instance, presence of a hollow section at the center of cavity can lead to multiple variations in the heat transfer along the hot surface [18][19].
Similarly, an addition of any solid substance in the porous domain is known to produce different heat transfer characteristics depending upon the solid and its location in the domain. Arasteh et al. [20] while studying the heat transfer of metal foam reported that the metal foam addition to the channel increases heat transfer and pressure drop. Nazri et al. [21] analysed the mixed convection in porous cavity. They found that the volume fraction increment of nanoparticles and reduction in Richardson number improved the temperature distribution.
The process of heat transfer across the porous medium via a solid substance that emanates complex boundary conditions between solid and fluid is coined as conjugate hear transfer. This concept did not draw much attention of the researchers due to its complex phenomenon when any solid obstruction arises in the porous medium that substantially changes the flow behaviour leading to change in heat and mass transfer characteristics. Lamnato et al. [22] reported that the drying process of porous media produced an extreme mass transfer at the surface of porous sliced wood with the aid of upstream flow divider. It is reported that [23] the wall thickness and ratio of wall to fluid conductivity, helps in enhancing the heat transfer characteristics. In an identical work Tao [24], found that heat capacity ratio of the fluids affects the conjugate heat transfer in finned tube. Saeid [25] investigated the effect of natural convention due to presence of solid vertical wall with emphasis on conduction in the solid wall. Amiri etc al [27] studied the steady state heat transfer with conjugate effect in porous cavity. Heat transfer analyses of solid on inner radius of a vertical annulus was analyzed by Ahmed et al [28] which was further extended to analyze the heat and mass transfer [29]. Alhashash et al [30] found that the radiation intensity enhances the fluid circulation and so does the Nusselt number. Al-Farhany et al [31], have shown in their work that the Darcy number is inversely related with average Nusselt number. Badruddin et al [32] analyzed the role of varying conductivity ratio as well as the relative thickness of two solid surfaces in an annulus. They reported that the conductivity ratio as well as the relative thickness of participating solids have impacts on heat transfer characteristics. Recent investigations have shown that the size and location of a rectangular solid block inside the porous domain would have significant influence on the double diffusion [33]. Ghalambaz et al. [34] investigated the conjugate heat transfer using local thermal non-equilibrium with a porous enclosure with hybrid nano-particles. Ghalambaz et al. [35] also studied the conjugate heat transfer in a cavity using thermal nonequilibrium model. The mixed convection conjugate heat transfer due to nanofluid in inclined flat plate about a porous regime was analysed by Khademi et al [36].
The current work emphasizes the analysis of the heat and mass transfer in presence of a solid block at the centre of porous cavity. It was noted that the solid block at the cavity centre would produce unique heat transfer behaviour [37]. However, the literature needs to be updated concerning the double diffusive flow in a cavity containing a solid block. The present work is an effort to fulfil this missing information.
MATHEMATICAL METHOD
A square cavity as depicted in figure 1 with a square solid placed at the center was considered. The equations governing heat and fluid flow for this particular problem are given as [38]: Subjected to the boundary conditions as shown in table 1: Using stream function ψ as: The above equations are dimensionless as: Substituting equations (6) into equation (2)(3)(4) yields: The corresponding dimensionless Boundary Conditions are given in table 2: The Nusselt number (Nu) is calculated as: The Sherwood number (Sh) is given by: The above-mentioned 4 partial differential equations (7-10) with boundary conditions given in Table 2 are simplified using Finite Element Method (FEM), following the procedure as given in [39][40][41][42]. The domain of the porous cavity is discretized using three nodded triangular elements. This results in a simpler form of three algebraic equations for each of the elements. We achieved the solution by utilizing the algorithm as suggested by Badruddin et al [43]. The domain is meshed by utilizing the triangular elements. The solution is obtained by an iterative method with a convergence criterion of 10 -6 for temperature as well as concentration and 10 -7 for stream function. The mesh independency is ensured to have good accuracy of results. [19]. The mass transfer is significantly affected by the increased size of solid as represented by iso-concentration lines of figure 2. As the solid size increases (owing to increased B ), the larger area of cavity is influenced by concentration. This happens because of the obstruction to mass transfer posed by solid block. The impact of changing buoyancy ratio from N=0.1 to N=1 is depicted in figure 3, having the constants as
RESULTS AND DISCUSSION
Le=5, Ra=50, Kr=2. As expected [33], the effect of N is negligible on isotherms as compared to the mass transfer. The increased buoyancy ratio generates better mass transfer, which in turn allows the concentration lines to shift towards the lower side of the hot surface. This is consistent with available literature covering the topic of heat and mass transfer. The magnitude of the streamlines rises with the increase in N and surrounds the square block. Figure 4 draws the conclusion of Lewis number (Le) at 375 . 0 = B , Kr=2, Ra=100, N= 0.2. As observed in case of heat and mass transfer [12], Lewis number influences the mass transfer substantially. This is obvious from present case (figure 4), that shows highly distorted iso-concentration lines as compared to isotherms, when Lewis number (Le) is increased from 1 to 25. The isoconcentration substantially comes closer to the hot surface that creates a large concentration gradient, which in turn should increase the mass transfer in the domain. Figures 5-10 shows the influence of distinct parameters on local Nusselt number (Nu) and Sherwood number (Sh) with respect to different parameters. The Nu decreases along the height of cavity, which is in line with that of cavity having no solid [7]. The Nu is found to decrease with increase in the solid size in the porous medium, which acts as a barricade. This is because the fluid movement gets restricted due to increase in solid size that reduces the heat carrying capability thus reducing the Nu at hot surface of the domain. The Nu is found to be higher for higher value of buoyancy ratio (N) until almost half of the hot surface as shown in figure 7. The Lewis number is found to have lesser impact on heat transfer as shown in figure 9, which follows the trend as in other studies [29]. The local Sh too decreases with the rise in cavity height but not for smaller dimensions of solid. However, there is peculiar behavior of mass transfer for where a sharp rise and fall of Sherwood number is observed at cavity height of about 0.7. This happens due to the reason that the larger size of solid makes concentration distribution split into two halves at around cavity height 0.65, as observed in figures 2-4.
The presence of larger sized solid block forces the fluid to confine within a narrow vertical channel thus reducing the mass carrying capacity from the hot surface. However, the fluid rushes towards the cold surface when it gets the opportunity at the end of solid block, which leads to sharp rise in mass transfer as reflected in terms of increased Sherwood number. This is further vindicated by the streamline flow pattern as shown in figures 2-4. This behaviour is consistently seen for Sherwood number with respect to all other parameters such as buoyancy ratio (figure 8), Lewis number (figure 10), Thus it can be conveniently said that the mass transfer has a unique behaviour at higher size of solid that shoots up the mass transfer at a particular position of cavity height. Such behaviour is not reported in literature for solid being placed at other positions. It can be said that the Nu increases with increase in buoyancy ratio (N) (figure 7), However, Nu diminishes as Lewis number (Le) increases ( figure 9). This is consistent with the findings of other researchers [36][37].
CONCLUSION
The problem of double diffusion in a porous square cavity with a solid square block is analysed numerically. The main concluding points of this work are: The heat and mass transfer in porous media deteriorated with increasing the solid height. This significantly depends on the Buoyancy ratio, the local Nusselt number increases for higher values of Buoyancy ratio (N), along the height of the cavity, but decreases at a certain point of the cavity.
The Sherwood number shows a sharp increase at . In all cases, the most common effect of Buoyancy ratio was that the Nusselt number becomes the increasing function of N except the Lewis number | 2,650.2 | 2020-01-01T00:00:00.000 | [
"Physics"
] |
A Portable and Wireless Multi-Channel Acquisition System for Physiological Signal Measurements †
We propose a portable and wireless acquisition system to help consumers or users register important physiological signals. The acquisition system mainly consists of a portable device, a graphic user interface (GUI), and an application program for displaying the signals on a notebook (NB) computer or a smart device. Essential characteristics of the portable device include eight measuring channels, a powerful microcontroller unit, a lithium battery, Bluetooth 3.0 data transmission, and a built-in 2 GB flash memory. In addition, the signals that are measured can be displayed on a tablet, a smart phone, or a notebook computer concurrently. Additionally, the proposed system provides extra power supply sources of ±3 V for the usage of external circuits. On the other hand, consumers or users can design their own sensing circuits and combine them with this system to carry out ubiquitous physiological studies. Four major advantages in the proposed system are the capability of combining it with a NB computer or a smart phone to display the signals being measured in real time, the superior mobility due to its own independent power system, flash memory, and good expandability. Briefly, this acquisition system offers consumers or users a convenient and portable studying tool to measure dynamic vital signals of interest in psychological and physiological research fields.
Introduction
Most fitness and healthcare devices are capable of dynamic monitoring; for instance pedometers and heart rate monitors. These wearable devices must have the function of an independent power supply and the ability of wireless transmission. Typically, they measure and deal with these physiological signals by means of a microcontroller unit (MCU) or send these signals to a tablet or a smart phone to be further processed. Therefore, in the development of wearable healthcare devices, it is important to explore innovative measurement technologies and signal processing methods to overcome the instability relevant to the acquisition of dynamic physiological signals. For these researchers in hardware development, they must not only develop novel analog driving circuits, but also build a standard acquisition system which is supposed to incorporate a power system, an analog-to-digital converter, and a wireless transmission scheme. Moreover, they also need to build a display and control interface in a software environment. It is a really time and energy-consuming task for those researchers Device is an analog tri-axial accelerometer integrated chip applied for object activity measurement [15]. As the sensor modules are employed to detect a variety of physiological signals, an acquisition system with multiple channels is required to show and record these signals. However, among the current commercial products, merely a few devices can not only be controlled with a tablet or a smart phone, but can also store data on them. Thus, the goal of this study was to develop a portable and wireless multi-channel acquisition system for the physiological signal measurements. It has eight analog channels and can be controlled by a NB computer or a smart device. The measured signals can either be shown on a NB computer or a smart device in real time, or be stored on the flash memory of the portable acquisition device. A sampling frequency of the portable signal acquisition device is 500 Hz, which is enough to conform to the Nyquist frequency of some physiological signals, such as the ECG, electroencephalogram (EEG), electrooculogram (EOG), galvanic skin response (GSR), and photoplethysmogram (PPG), since most of them do not have a large bandwidth [16]. With a TI MSP430 F5438A as its MCU, the portable acquisition device has a compact size, uses a lithium battery (350 mA) to supply the needed power, employs a BT3.0 module to transmit data, and a 2 GB flash memory to store the signals being measured. Moreover, the portable acquisition device can offer dual power levels, ±3 voltage, so that the external sensor modules may connect with this device to measure different physiological signals. The real-time measurements can be displayed on a NB computer or smart phone. Thus, consumers or researchers can confirm the stability and accuracy of the measured signals during the experiment.
The rest of the paper is organized as follows: Section 2 describes the structure of the multi-channel acquisition system, and its software commands on both smart devices and NB computers. Section 3 describes the hardware and firmware designs of the portable signal acquisition device. Section 4 presents the results, and the discussion. Conclusions are drawn in Sections 5 and 6.
Measurement System
The structural diagram of the portable and wireless multi-channel acquisition system, which includes three parts, a portable acquisition device, a graphic user interface (GUI), and an application program (APP) for a NB computer and a smart device, is shown in Figure 1. There are eight analog channels available to connect with the external sensor modules indicated with an orange arrow. The black arrows represent the control commands which are sent out by a smart device or a NB computer. The control commands can trigger the portable acquisition device to either start or stop the measurement, write the data to the flash memory, or clear the data on the flash memory. The data on the flash memory of the portable acquisition device is downloaded to the NB computer by means of the BT or a universal serial bus (USB). Then, the downloaded data can be displayed on the NB computer. The blue arrows represent the measured physiological signals that can be shown on the smart device or the NB computer in real time. When the measured signals are transmitted from the portable acquisition device to the NB computer through either the BT3.0 or the USB; the sampling rate is 500 Hz. When the portable acquisition device transmits one signal to the smart device, the sampling rate is reduced to 100 Hz. The software commands on smart devices and NB computers are described in the subsections below.
Application Program
In this study, Cordova and Eclipse were chosen to develop an APP for smart devices, such as a tablet or a smart phone. As we know, Cordova is an open source software and many users can employ it to discuss how to create new APPs online. Eclipse is also a well-known cross-platform which belongs to a free integrated development environment, with lots of external programs. Thus, it can be utilized flexibly and conveniently.
We selected, in this work, JavaScript as the programming language to develop an APP. In addition, both the built-in BT module and BT plug-in of the Android system are employed to connect with the acquisition system and to send out commands, such as triggering the analog to digital converter (ADC), stopping the ADC, writing data to the flash memory, and finishing the writing of the data. As shown in Figure 2, the commands inside the red block belong to the APP.
Graphic User Interface
As shown in Figure 2, the NB GUI not only has the same commands as the APP, but also has the "Load" and "Clear" commands. The loading command is used to download the data stored on the flash memory to the NB computer, while the clear command is used to erase the data on the flash memory. There are four bytes in each command package. The first two bytes of the package are the head bytes, 0 × 66. Six commands are proposed in this acquisition system and their corresponding codes are shown in Table 1.
Application Program
In this study, Cordova and Eclipse were chosen to develop an APP for smart devices, such as a tablet or a smart phone. As we know, Cordova is an open source software and many users can employ it to discuss how to create new APPs online. Eclipse is also a well-known cross-platform which belongs to a free integrated development environment, with lots of external programs. Thus, it can be utilized flexibly and conveniently.
We selected, in this work, JavaScript as the programming language to develop an APP. In addition, both the built-in BT module and BT plug-in of the Android system are employed to connect with the acquisition system and to send out commands, such as triggering the analog to digital converter (ADC), stopping the ADC, writing data to the flash memory, and finishing the writing of the data. As shown in Figure 2, the commands inside the red block belong to the APP.
Hardware and Firmware of the Portable Signal Acquisition Device
In this section, the major parts of the portable signal acquisition device are described, respectively, including the power circuit, the communication protocol of the BT module, the storage arrangement of the flash memory, and the principal settings of the TI MSP430 F5438A.
Graphic User Interface
As shown in Figure 2, the NB GUI not only has the same commands as the APP, but also has the "Load" and "Clear" commands. The loading command is used to download the data stored on the flash memory to the NB computer, while the clear command is used to erase the data on the flash memory. There are four bytes in each command package. The first two bytes of the package are the head bytes, 0 × 66. Six commands are proposed in this acquisition system and their corresponding codes are shown in Table 1.
Hardware and Firmware of the Portable Signal Acquisition Device
In this section, the major parts of the portable signal acquisition device are described, respectively, including the power circuit, the communication protocol of the BT module, the storage arrangement of the flash memory, and the principal settings of the TI MSP430 F5438A.
Power Circuit
In the power circuit of the portable signal acquisition device, the Texas Instruments BQ24072 battery charger is employed. The interatged circuit (IC) (TPS78233, IT) is applied to provide a voltage of 3.3 V, which is supplied from a lithium battery, while the IC (TPS60400, IT) is used to provide a voltage of −3.3 V. This power scheme also offers the dual power, ±3.3 V, which can be utilized by external sensor circuits.
Communication Protocol of Bluetooth
The BT 3.0 module (BTM-204B) was chosen to connect with the Universal Asynchronous Receiver/Transmitter (UART) of the MSP430 F5438A. There are 18 bytes in the transmission package. First, the two bytes, 0 × 66, are used as the head bytes of the package data, and the ADC value for each channel is divided into high bytes and low bytes. The purpose of the two head bytes is to mark the start of the data entry. After the starting information is received, the deciphering operation includes the multiplication of the high bytes of ADC value by 0 × 100. Then, this value is added to the low bytes as the true value. The UART port of the MSP430 F5438A is selected as a virtual USB port by the PL2303HXD IC for communicating with the NB computer.
Storage Format of Flash Memory
Since the built-in 256 kb flash memory of the MSP430 is insufficient to record a large amount of data, external add-on memory is required for the portable signal acquisition device. A NAND flash memory (GD5F2Q4UBFI) of 2 GB is employed in the proposed system. The array of the flash memory consists of 64 pages in total. The first 2 bytes of the first page specify the file number, which is followed sequentially by the start page, the last page, and the last bytes. There are 4096 bytes in total, for the first page. Since the file name occupies the first 2 bytes, six bytes are left. Therefore, the 2 GB flash memory can be used to store 682 files. Figure 3 shows the structure of one data access file in the 682 files. NAND Flash uses the SPI (serial peripheral interface) communication approach for data transfers to the MSP430 microcontroller, in which the clock (CLK) frequency is 12 MHz and the data transmission speed is nine kilobits per second at a sampling frequency of 500 Hz. There are eight Sensors 2019, 19, 5314 6 of 12 channels, each channel signal being assigned 2 bytes, with a transmission rate of 8 kilobits per second. Therefore, the access rate of the portable signal acquisition device is good enough.
Sensors 2020, xx, x FOR PEER REVIEW 6 of 12
The Settings of MSP430F5438A
The MSP430 F5438A, a 16-bit MCU used in this study has one flash memory of 256 kilobytes and eight ADCs of 12 bits each. Each ADC channel with a sampling frequency of 500 HZ and an internal flash memory is used to record the number and the size of a file accessed in the external flash memory. The firmware flowchart for the MSP430 microcontroller is shown in Figure 4. There are two UART sets in the 16-bit MCU. One is used for the BT module and the other for the USB port. When the measurement signals are sent to a smart device to determine whether they are saved to the flash memory or not, the sampling frequency is reduced to 100 Hz. The main clock rate of the MCU is set at 24 MHz. Clock A0, at 500 Hz, is employed as the ADC sampling frequency and to set up the UART and SPI clocks. The eight channels are made available for the ADC operation. After initial settings are completed, time interruption is enabled and the MCU will wait for a command, such as an "ADC on," a "Write data" to flash memory, and so on.
The MCU determines whether the flash memory has enough free space based on two conditions: (1) there are 682 files in the flash memory, and (2) the last page of the flash memory has been used. When either of these two conditions is satisfied, the data can no longer be written to the flash memory.
The MCU controls a set of four LEDs indicating: (1) the status of the connection to the USB port, (2) the status of the BT connection, (3) the status of the 3.3 V power, and (4) the status of the flash memory, respectively. If the MCU is in the reading mode, the LED will remain on. However, when in the writing mode, a blinking LED is displayed.
The Settings of MSP430F5438A
The MSP430 F5438A, a 16-bit MCU used in this study has one flash memory of 256 kilobytes and eight ADCs of 12 bits each. Each ADC channel with a sampling frequency of 500 HZ and an internal flash memory is used to record the number and the size of a file accessed in the external flash memory. The firmware flowchart for the MSP430 microcontroller is shown in Figure 4. There are two UART sets in the 16-bit MCU. One is used for the BT module and the other for the USB port. When the measurement signals are sent to a smart device to determine whether they are saved to the flash memory or not, the sampling frequency is reduced to 100 Hz. The main clock rate of the MCU is set at 24 MHz. Clock A0, at 500 Hz, is employed as the ADC sampling frequency and to set up the UART and SPI clocks. The eight channels are made available for the ADC operation. After initial settings are completed, time interruption is enabled and the MCU will wait for a command, such as an "ADC on," a "Write data" to flash memory, and so on.
The Settings of MSP430F5438A
The MSP430 F5438A, a 16-bit MCU used in this study has one flash memory of 256 kilobytes and eight ADCs of 12 bits each. Each ADC channel with a sampling frequency of 500 HZ and an internal flash memory is used to record the number and the size of a file accessed in the external flash memory. The firmware flowchart for the MSP430 microcontroller is shown in Figure 4. There are two UART sets in the 16-bit MCU. One is used for the BT module and the other for the USB port. When the measurement signals are sent to a smart device to determine whether they are saved to the flash memory or not, the sampling frequency is reduced to 100 Hz. The main clock rate of the MCU is set at 24 MHz. Clock A0, at 500 Hz, is employed as the ADC sampling frequency and to set up the UART and SPI clocks. The eight channels are made available for the ADC operation. After initial settings are completed, time interruption is enabled and the MCU will wait for a command, such as an "ADC on," a "Write data" to flash memory, and so on.
The MCU determines whether the flash memory has enough free space based on two conditions: (1) there are 682 files in the flash memory, and (2) the last page of the flash memory has been used. When either of these two conditions is satisfied, the data can no longer be written to the flash memory.
The MCU controls a set of four LEDs indicating: (1) the status of the connection to the USB port, (2) the status of the BT connection, (3) the status of the 3.3 V power, and (4) the status of the flash memory, respectively. If the MCU is in the reading mode, the LED will remain on. However, when in the writing mode, a blinking LED is displayed. The MCU determines whether the flash memory has enough free space based on two conditions: (1) there are 682 files in the flash memory, and (2) the last page of the flash memory has been used. When either of these two conditions is satisfied, the data can no longer be written to the flash memory. The MCU controls a set of four LEDs indicating: (1) the status of the connection to the USB port, (2) the status of the BT connection, (3) the status of the 3.3 V power, and (4) the status of the flash memory, respectively. If the MCU is in the reading mode, the LED will remain on. However, when in the writing mode, a blinking LED is displayed.
Results
In this section, the circuit structure and the electric performance of the portable signal acquisition device are depicted. The GUIs on the notebook and smart phone designed to operate the function of the portable signal acquisition device and the display of measured signals are also presented. Figure 5a shows the block diagram of the portable acquisition device, Figure 5b shows its printed circuit, and Figure 5c shows its real photograph. Without the battery, its weight is only 15 g, and its size is 12 × 12 cm. In the upper right-hand corner, two voltage sources, 3.3 and −3.3 V, provide the power for external sensor devices with two maximum currents of 400 and 60 mA, respectively. The USB port is used either to charge the lithium battery or to download the data to the hard disk in the NB computer. In Figure 5a, the black lines represent the power connection, and the blue lines represent the data transmission. The 5 V line and ground line of the USB port were connected to the input pins of the charge IC, TI BQ24072. The two data lines of the USB port were connected to the USB-to-serial bridge controller, PL 2303HX. The USB socket is placed at the upper left-hand corner of the printed circuit board. The MCU, BT module, and flash memory consume the largest amount of power in this portable acquisition device. Since the active current of the MSP430 F5438A is 165 µA/MHz, it requires the currents of 3.96 mA and 2.6 µA under the maximum running conditions and the low power mode, respectively. The BT model needs the currents of 37 mA and 70 µA under the transmission and standby modes, respectively. Additionally, the flash memory needs the currents of 40 mA and 70 µA under the operation and standby modes, respectively. Therefore, when the portable acquisition device makes use of a smart device to display the signals, it requires a maximum current of about 43 mA. But, if users want to display signals and write data to the flash memory at the same time, the maximum current conumed will be 73 mA. When the input of ADC is connected to the power source, +3.3 V, and ground, 0 V, the codes of ADC with the ten samples are shown in Table 2. The statistical codes for the positive power source and ground source are 4095 ± 1 (mean ± standard deviation) and 1 ± 1, respectively.
Notebook Computer GUI
As the communication interface software, GUI codes for the NB computer were written in C#. The main function of the NB computer is the system control and downloading data from the portable signal acquisition device. The GUI software was developed for the connection with the portable signal acquisition device, and for the control operations, including starting the ADC, stopping the ADC, writing data to the flash memory, loading and clearing the data of the flash memory, and triggering the other actions.
The portable acquisition devices connected with three sensor modules include the KL-74002A (ECG), KL-74006A (PPG), and KL-74005 (heart sound). These sensor modules were commercial products designed by the K&H Co., Ltd. Company, Taiwan. In the GUI, the first operation is to select one of the common ports (COM PORT) to link the portable signal acquisition device, as shown in Figure 6. Once this connection has been made, the "Connect" button on the display is indented. If the portable signal acquisition device and the NB computer have been successfully linked, the Connect button will change to Disconnect, and the other commands such as the ADC ON, Load Flash, and Clear Flash will appear, and any one of them can be used. After the ADC ON button has been pressed, the measurement signals will be immediately sent to the NB computer. Then, users can choose which channels will be used to display the signals, as shown in Figure 6. There are three kinds of measurement signals respectively, from the channels 1, 2, and 4 that can be displayed on the screen-ECG, PPG, and heart sound signals. To stop the ADC operation, press the ADC OFF button. If users want to record the signals, they can press the Write button. These signals will be recorded on the flash memory in real time. For each new recording event, the file number in the flash memory is automatically increased by one, and the data will be recorded on a new page of the flash memory.
Notebook Computer GUI
As the communication interface software, GUI codes for the NB computer were written in C#. The main function of the NB computer is the system control and downloading data from the portable signal acquisition device. The GUI software was developed for the connection with the portable signal acquisition device, and for the control operations, including starting the ADC, stopping the ADC, writing data to the flash memory, loading and clearing the data of the flash memory, and triggering the other actions.
The portable acquisition devices connected with three sensor modules include the KL-74002A (ECG), KL-74006A (PPG), and KL-74005 (heart sound). These sensor modules were commercial products designed by the K&H Co., Ltd. Company, Taiwan. In the GUI, the first operation is to select one of the common ports (COM PORT) to link the portable signal acquisition device, as shown in Figure 6. Once this connection has been made, the "Connect" button on the display is indented. If the portable signal acquisition device and the NB computer have been successfully linked, the Connect button will change to Disconnect, and the other commands such as the ADC ON, Load Flash, and Clear Flash will appear, and any one of them can be used. After the ADC ON button has been pressed, the measurement signals will be immediately sent to the NB computer. Then, users can choose which channels will be used to display the signals, as shown in Figure 6. There are three kinds of measurement signals respectively, from the channels 1, 2, and 4 that can be displayed on the screen-ECG, PPG, and heart sound signals. To stop the ADC operation, press the ADC OFF button. If users want to record the signals, they can press the Write button. These signals will be recorded on the flash memory in real time. For each new recording event, the file number in the flash memory is automatically increased by one, and the data will be recorded on a new page of the flash memory. If users need to download the data from the flash memory to the NB computer, they may press the Load Flash button. An access window will be displayed immediately. Select a location and enter a file name first; then, press the Save button to allow all the data on the flash memory to be saved to a file in a "text" format, and the data of each experiment can be recorded in an independent file. The information associated with the saved files will be displayed when the data on the flash memory has been successfully saved, as shown in Figure 7. Such file information includes the file numbers, the downloading bytes, the ratio of the flash memory used, and the downloading condition. The data of the eight channels in the first file are shown in the left column of Figure 7. Figure 8 illustrates the APP interface where the numbers with the red color, are used to indicate the processing steps. First, as BT Search is initiated, the corresponding BT device can be found by name. After establishing a successful connection, users may then press the ADC_ ON button to begin the signal measurement. But, only the signal of the first channel will be displayed in the APP, as shown in Figure 6. When the Flash_Write button is pressed, the signals of the eight channels are recorded on the flash memory. The Flash_Stop button is used to stop the recording. Finally, users may press the ADC_Off button to turn the ADC operation off. If users need to download the data from the flash memory to the NB computer, they may press the Load Flash button. An access window will be displayed immediately. Select a location and enter a file name first; then, press the Save button to allow all the data on the flash memory to be saved to a file in a "text" format, and the data of each experiment can be recorded in an independent file. The information associated with the saved files will be displayed when the data on the flash memory has been successfully saved, as shown in Figure 7. Such file information includes the file numbers, the downloading bytes, the ratio of the flash memory used, and the downloading condition. The data of the eight channels in the first file are shown in the left column of Figure 7. If users need to download the data from the flash memory to the NB computer, they may press the Load Flash button. An access window will be displayed immediately. Select a location and enter a file name first; then, press the Save button to allow all the data on the flash memory to be saved to a file in a "text" format, and the data of each experiment can be recorded in an independent file. The information associated with the saved files will be displayed when the data on the flash memory has been successfully saved, as shown in Figure 7. Such file information includes the file numbers, the downloading bytes, the ratio of the flash memory used, and the downloading condition. The data of the eight channels in the first file are shown in the left column of Figure 7. Figure 8 illustrates the APP interface where the numbers with the red color, are used to indicate the processing steps. First, as BT Search is initiated, the corresponding BT device can be found by name. After establishing a successful connection, users may then press the ADC_ ON button to begin the signal measurement. But, only the signal of the first channel will be displayed in the APP, as shown in Figure 6. When the Flash_Write button is pressed, the signals of the eight channels are recorded on the flash memory. The Flash_Stop button is used to stop the recording. Finally, users may press the ADC_Off button to turn the ADC operation off. Figure 8 illustrates the APP interface where the numbers with the red color, are used to indicate the processing steps. First, as BT Search is initiated, the corresponding BT device can be found by name. After establishing a successful connection, users may then press the ADC_ ON button to begin the signal measurement. But, only the signal of the first channel will be displayed in the APP, as shown in Figure 6. When the Flash_Write button is pressed, the signals of the eight channels are recorded on the flash memory. The Flash_Stop button is used to stop the recording. Finally, users may press the ADC_Off button to turn the ADC operation off.
Discussion
In the last decade, wearable devices have been applied in various fields, especially in the healthcare sector [17]. Among the physiological measurements, the ECG, EMG, EEG, and PPG signals are the most popular in modern wearable devices for daily healthcare monitoring. However, one of the challenges with designing such wearable devices is how to faithfully acquire these physiological signals under an actual environment. Requirements for the wearable healthcare devices are somewhat different from those for the traditional devices in clinical locations. A big difference is that a wearable device always not only detects the dynamic physiological signals, but needs to display these signals in real time. In recent studies, some acquisition systems have employed BT, Zigbee, WiFi, or RF transceiver to transmit the measurement signals to a NB computer or a smart device, and to display or store them on those smart devices [5][6][7][8]. The characteristic comparison among the four wireless techniques is shown in Table 3. Although the WiFi has the fastest transmission rate and the longest transmission distance, its power consumption also is the largest. The BT does not have the best performance of these wireless techniques. But, nowadays, both NB computers and smart phones have WiFi and BT communication functions. Nevertheless, there are two main disadvantages in those acquisition systems, including low mobility and low sampling rates.
Technique
Distance Power consumption Transmission rate BT 3.0 (HC-05) [18] < 10 m 30 mA 1M (bits/s) WiFi (ESP8285) [19] > 10 m 80 mA 54M (bits/s) Zigbee (Xbee ® ) [20] > 10 m 45 mA 250k (bits/s) RF (BC9824) [21] < 2 m 10 mA 2M (bits/s) In this study, a portable and wireless multi-channel acquisition system for physiological signal measurements was fully established. This acquisition system is primarily comprised of a portable acquisition device, a GUI, and an APP to display the signals on a NB computer or a smart device. Since the power consumption of the BT technique is relatively low compared with WiFi, we finally chose the BT technique to perform the connection with the NB computer or the smart device in order to reduce the total power consumption of the portable acquisition device. Three major advantages exist in the present multi-channel acquisition system. First, this acquisition system allows users to combine it with a NB computer or a smart device to display the measurement signals in real time and to easily control the functions of the portable acquisition device. Therefore, users are permitted
Discussion
In the last decade, wearable devices have been applied in various fields, especially in the healthcare sector [17]. Among the physiological measurements, the ECG, EMG, EEG, and PPG signals are the most popular in modern wearable devices for daily healthcare monitoring. However, one of the challenges with designing such wearable devices is how to faithfully acquire these physiological signals under an actual environment. Requirements for the wearable healthcare devices are somewhat different from those for the traditional devices in clinical locations. A big difference is that a wearable device always not only detects the dynamic physiological signals, but needs to display these signals in real time. In recent studies, some acquisition systems have employed BT, Zigbee, WiFi, or RF transceiver to transmit the measurement signals to a NB computer or a smart device, and to display or store them on those smart devices [5][6][7][8]. The characteristic comparison among the four wireless techniques is shown in Table 3. Although the WiFi has the fastest transmission rate and the longest transmission distance, its power consumption also is the largest. The BT does not have the best performance of these wireless techniques. But, nowadays, both NB computers and smart phones have WiFi and BT communication functions. Nevertheless, there are two main disadvantages in those acquisition systems, including low mobility and low sampling rates. [19] >10 m 80 mA 54M (bits/s) Zigbee (Xbee ® ) [20] >10 m 45 mA 250k (bits/s) RF (BC9824) [21] <2 m 10 mA 2M (bits/s) In this study, a portable and wireless multi-channel acquisition system for physiological signal measurements was fully established. This acquisition system is primarily comprised of a portable acquisition device, a GUI, and an APP to display the signals on a NB computer or a smart device. Since the power consumption of the BT technique is relatively low compared with WiFi, we finally chose the BT technique to perform the connection with the NB computer or the smart device in order to reduce the total power consumption of the portable acquisition device. Three major advantages exist in the present multi-channel acquisition system. First, this acquisition system allows users to combine it with a NB computer or a smart device to display the measurement signals in real time and to easily control the functions of the portable acquisition device. Therefore, users are permitted to check the stability and accuracy of the recorded signals in real time during the experiment. When the correctness of the measured signals is carefully confirmed, users can begin to record the signals on the flash memory. Second, the portable signal acquisition device possesses superior mobility. Users may take the device to any place due to its own independent power system and adequate memory. Third, the acquisition system has good expandability. For instance, it offers ±3.3 voltage for the use of external or additional circuits. Furthermore, with eight channels, it can be applied to simultaneously register various physiological signals. Because the resolution of ADC is 80.6 uV, it could be used to properly measure several kinds of physiological signals-ECG, EMG, PPG, body acceleration signal, and so on.
The commercial products, such as the BioPac Bionomadix smart center [3] and Bioradio TM wireless measurement system [4], have the exclusive sensor modules and must connect with the NB. These measurement systems are just like gateways to transmit the signals to the NB computer. Therefore, their compatibilities are not higher than our proposed system. Users need to carry a NB computer to record these signals during the experimental course. For our system, users only use a smart phone to check the measurement signals and to see whether they are stable and accurate or not when doing the experiment. Once the signals are properly measured, users can control the portable acquisition device to record the signals onto the flash memory in real time. Moreover, since our proposed system is intended to be employed for signal acquisition, it does not need to do the calibration of the sensor modules.
Conclusions
In summary, we designed the portable and wireless multi-channel acquisition system which offered a convenient and portable studying tool to measure the dynamic physiological signals in healthcare research fields. Users can watch the measurement signals in real time and control the functions of this multi-channel acquisition device with a smart phone or NB computer. Additionally, researchers can design their own sensing circuits and combine them with this acquisition system. Then, they can rapidly perform their experiments, and with a clear mind, collect physiological signals of interest in any scenario. Our proposed acquisition system is currently a prototype. However, in the future, we will modify this system as a commercial product according to the feedback of the intended audiences. | 8,435 | 2019-12-01T00:00:00.000 | [
"Computer Science"
] |
The HTLV-1 Virological Synapse
Human T-lymphotropic virus-1 (HTLV-1) spreads efficiently between T-cells via a tight and highly organized cell-cell contact known as the virological synapse. It is now thought that many retroviruses and other viruses spread via a virological synapse, which may be defined as a virus-induced, specialized area of cell-to-cell contact that promotes the directed transmission of the virus between cells. We summarize here the mechanisms leading to the formation of the HTLV-1 virological synapse and the role played by HTLV-1 Tax protein. We propose a model of HTLV-1 transmission between T-cells based on the three-dimensional ultrastructure of the virological synapse. Finally, in the light of recent advances, we discuss the possible routes of HTLV-1 spread across the virological synapse.
Introduction
Human T-lymphotropic virus type 1 (HTLV-1) is a complex retrovirus, classified in the Deltaretrovirus genus of the subfamily Orthoretrovirinae. The majority (about 95%) of infected individuals develop no associated disease (asymptomatic carriers; ACs). However, in a small fraction of individuals, infection results in one of two types of disease: about 2-5% develop adult T cell leukemia/lymphoma (ATLL) [1], and a further 0.5-3% develop one or more of a range of OPEN ACCESS inflammatory diseases. The most common of these conditions is HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP), an inflammatory disease of the central nervous system [2,3].
Until recently, it was generally believed that HTLV-1 is latent in vivo [4][5][6]. However, there is significant evidence to the contrary [7,8]. The presence of high frequencies of activated HTLV-1specific cytotoxic T lymphocytes (CTLs) [9,10] in the peripheral blood supports the hypothesis that the virus is not latent, i.e., there is ongoing viral transcription, and that this is greater in patients with HAM/TSP than in ACs. Direct evidence for selective proliferation of HTLV-1-infected T cells was obtained by Asquith et al. (2007) by in vivo labeling with deuterated glucose [11].
In this review, we consider the mechanisms of cell-to-cell spread of HTLV-1. Following the discovery of the virological synapse (VS) in 2003, there have been significant advances in the understanding of the mechanism of formation of the synapse and in the locus of transfer of virions from cell to cell. We conclude that HTLV-1, as appears to be the case for HIV-1 and MLV, can be transferred from cell to cell both at sites of budding at the closely apposed plasma membranes at the VS and by lateral movement of preformed virions at, or near, the periphery of the cell-to-cell contact, where they may be protected in a 'biofilm' of extracellular matrix.
HTLV-1 cell tropism, cell-to-cell spread and the VS
HTLV-1 can infect a wide range of human cell types in vitro [12], but in vivo the virus is almost confined to the CD4 + T lymphocyte subset [13][14][15][16]. Furthermore, most of the malignancies induced by HTLV-1 are tumors of CD4 + T lymphocytes [17]. CD8 + T lymphocytes can also carry the virus, but at a consistently lower frequency than CD4 + T cells [18,19].
The conjunction of two observations led to the postulation of the VS. First, direct cell-to-cell contact is necessary for efficient transmission of HTLV-1 from an infected cell to a new host cell, both in vitro [20,21] and in vivo [22], where transmission depends on transfer of infected lymphocytes in breast milk [23][24][25], semen [26] or transfused blood products [27,28]. HTLV-1 virions are typically undetectable in the serum of infected individuals by RT-PCR. Virions are produced only by certain continuous in vitro T cell lines: fresh, naturally infected lymphocytes do not produce cell-free HTLV-1 particles. Furthermore, of the cell-free HTLV-1 virions that are produced by transfected T cells or continuous producer T cell lines, only one in 10 5 to 10 6 is infectious [29]. Second, HTLV-1-specific T cells are themselves infected more frequently with HTLV-1 than are T cells specific to other antigens. This preferential infection was evident in both CD8 + T cells [18] and CD4 + T cells [30]. These two observations raised the possibility that HTLV-1 transmission was assisted by the process of T cell antigen recognition. More precisely, HTLV-1 might spread across the 'immunological synapse' [31], the specialized area of contact that is formed between a lymphocyte and another cell in which distinct protein microdomains mediate adhesion, antigen recognition and secretion of cytokines or lytic granules. Confocal microscopy of conjugates formed spontaneously between ex vivo CD4 + cells from an HTLV-1-infected person and autologous (or allogeneic) lymphocytes revealed a structure at the cell-cell junction, which indeed resembled the immunological synapse [32]. Polarization of the adhesion molecule talin and the microtubule organizing center (MTOC) to the cell-cell junction was accompanied by accumulation of the HTLV-1 core protein Gag and the HTLV-1 genome at the cell-cell junction. After 2 h, both the Gag protein and the HTLV-1 genome were transferred from the infected to the uninfected cell [32].
A critical observation revealed the distinction between the immunological synapse and the structure formed between an HTLV-1-infected cell and another cell. In an immunological synapse, the MTOC in the responding T cell is polarized towards the antigen-presenting cell, such as a virus-infected cell. This polarization is triggered by engagement of the T-cell antigen receptor [33,34]. In contrast, in the cell-cell conjugates formed with an HTLV-1-infected cell, the MTOC was polarized inside the virusinfected cell, not towards it. The results are shown in Table 1 [32]. Table 1. HTLV-1-infected cells polarize their MTOCs to the cell-cell junction in CD4+ Tcell conjugates. Two experiments were performed, each with fresh ex vivo CD4+ T cells from an unrelated HTLV-1-infected subject. Conjugates were allowed to form for 30 min (Subject 1) or 60 min (Subject 2), then were fixed and stained for HTLV-1 Gag p19 and tubulin alpha. Only conjugates containing two cells were counted. The figures denote the number (percentage) of cells whose MTOC was polarized to the cell-to-cell junction. Odds ratio (a) This observation showed that the mechanisms triggering the cytoskeletal polarization differed from the immunological synapse, and immediately suggested that the polarization was induced by HTLV-1 itself, perhaps in order to transmit viral material to the uninfected cell.
Inhibition of polarization and cell-to-cell transfer of HTLV-1 Gag protein at the cell-cell junction by the microtubule inhibitor nocodazole (30 µM) confirmed that the donor cell's microtubule cytoskeleton plays a central part in cell-to-cell transmission of HTLV-1 [35,36]. Recently, Mazurov et al. [37] used an elegant new system of luciferase-expressing, replication-dependent vectors to quantify the infectivity of HTLV-1 and HIV-1 and the contribution of Tax protein to cell-to-cell transmission.
These authors found that cell-to-cell transmission of HTLV-1 was >10 4 times more efficient than cellfree virus, whereas cell-to-cell transmission of HIV-1 was only two-fold more efficient. Disruption of the microtubules with nocodazole reduced HTLV-1 co-culture infectivity by 85%, whereas HIV-1 infectivity was reduced by 60%. In addition, the authors showed that the induction of an immunological synapse between a Raji cell and a CD4 + target cell did not increase infection with either HIV-1 or HTLV-1 virus-like particles, suggesting that cell-to-cell infection requires the formation of a specialized VS [37]. These results confirm the exceptionally strong dependence of HTLV-1 transmission on cell-cell contact and the importance of cytoskeletal remodeling in the cell-tocell transmission of both HTLV-1 and HIV-1.
Molecular triggers and intracellular pathways that cause the cytoskeleton polarization in HTLV-1-infected cells have now been identified (see below). Because both the mechanisms and the function of the HTLV-1-induced structure were distinct from the immunological synapse, the term 'virological synapse' was used [32].
Definition of the VS
A VS may be defined as a virus-induced, specialized area of cell-cell contact that promotes the directed transmission of the virus between cells. Many viruses are known to spread efficiently by cellcell contact, but such spread typically takes place across normal, pre-existing cellular contacts. In contrast, a VS is actively induced by contact between an infected cell and another cell. The viruses that benefit most strongly from transmission by a VS are therefore those that infect mobile cells, such as leukocytes. The VS maximizes the efficiency of transmission and limits the exposure of the virus to host defense mechanisms, both in time and space.
Polarization of viral proteins at the VS
The distribution of the HTLV-1 proteins Gag, Env and Tax can be visualized by confocal microcopy (Figure 1), and electron microscopy ( Figure 2) in naturally infected cells [32,35,36,[38][39][40][41]. In a single isolated cell (not making contact with a target cell) Gag is detected in the cytoplasm, forming large cytoplasmic inclusions whose precise nature remains unknown [42]. These inclusions appear to be randomly distributed around the periphery of the cell under the plasma membrane. Virions may also accumulate on the outside of the plasma membrane of an isolated cell, in the extracellular matrix ( [93]: see Section 9 below). Env protein is uniformly distributed around the plasma membrane in an isolated cell; Tax protein is mostly nuclear but a significant fraction of Tax is present around the MTOC in association with the cis-Golgi compartment [35,[43][44][45]. The distribution of Tax protein between the nucleus and the cytoplasm depends on the balance between ubiquitylation and sumoylation [44,46]. In cell conjugates formed between an HTLV-1 infected T-lymphocyte and autologous cells, the Gag and Env proteins are characteristically polarized toward the region of contact formed with the target cell [32,35,36,41]. In about 30% of cell-cell conjugates, a fraction of Tax also appears at the inner side of the plasma membrane in the region of cell-to-cell contact [35]: the function of Tax protein in this location is unknown. The mechanisms by which the HTLV-1 proteins are transported, sorted and delivered to the VS also remain unclear.
Polarization of the cytoskeleton to the VS
The formation of the VS is accompanied by the polarization of the MTOC in the infected cell toward the cell contact formed with the target cell ( Figure 1). This polarization is a good indication of the establishment of a VS [32,35,36,47]. The polarization depends on the integrity of both actin and microtubule components of the cytoskeleton and requires the activity of the small GTPases Rac1 and Cdc42 [35].
As described above, the observation that the MTOC polarizes towards the cell-cell junction inside the HTLV-1-infected cell, rather than in a T cell responding to an antigen, revealed the distinction between the VS and the immunological synapse. The trigger that causes the microtubule polarization in the VS therefore cannot be delivered by the T-cell antigen receptor. It was thus postulated that two signals were required to trigger the MTOC polarization observed at the VS: one from cell-cell contact and one from the viral infection itself. Specifically, it was hypothesized that engagement of a receptor on the surface of the infected cell acted in synergy with a signal from an intracellular HTLV-1 protein to trigger the observed polarization.
Cell surface trigger of MTOC polarization at the VS
To identify the signal from cell-cell contact, latex beads coated with monoclonal antibodies were used to cross-link certain molecules on the surface of the infected T cell. As expected, cross-linking of the T cell receptor with an anti-CD3 monoclonal antibody caused efficient polarization of the MTOC towards the coated latex bead [47]. In fact, cross-linking of a number of other cell-surface molecules also triggers MTOC polarization [47], demonstrating the unusual mobility of the T cell's microtubule cytoskeleton. However, engagement of either of two molecules, ICAM-1 (CD54) or the IL-2 receptor alpha chain (CD25), caused a significantly higher frequency of MTOC polarization in HTLV-1infected cells than in uninfected cells. The importance of ICAM-1 in triggering MTOC polarization was corroborated by two further observations. First, a soluble cyclic peptide derived from the region of LFA-1 that normally binds to ICAM-1 selectively inhibited the observed preferential MTOC polarization in HTLV-1-infected cells towards an uninfected cell. Second, the selective MTOC polarization in HTLV-1-infected cells was abolished when the target cell lacked LFA-1, but was enhanced when the target cell expressed the constitutively actively form of LFA-1 [47]. The reason for selective polarization towards CD25 is less obvious: we have suggested that this results from the known physical association between CD25 and ICAM-1 in the T cell plasma membrane [48].
Intracellular trigger of MTOC polarization at the VS
The signal from ICAM-1 cross-linking was especially effective in triggering cytoskeletal polarization in cells infected with HTLV-1. What was the signal from HTLV-1 infection that acted in synergy with the ICAM-1 signal? The Tax protein of HTLV-1 was a strong candidate, because of its early and abundant expression during HTLV-1 replication and its known capacity to transactivate genes via the CREB, NF-κB and SRF pathways [44,46,[49][50][51][52][53]. By transfecting a Tax expression plasmid into Jurkat T cells, we showed [35] that Tax is indeed sufficient to explain the observed preferential cytoskeletal polarization in HTLV-1-infected cells. We then used a series of mutants of the Tax protein to investigate the relationship between the signaling pathways activated by Tax, the intracellular localization of the Tax protein, and its ability to trigger MTOC polarization.
Mazurov et al. [37] have recently quantified the contribution of Tax to the efficiency of HTLV-1 transmission across the VS. Their results showed that Tax protein increased the efficiency of HTLV-1 transmission by more than 10-fold, confirming the earlier observations [35]. Remarkably, Tax also increased the efficiency of cell-to-cell transmission of HIV-1 by more than 10-fold, and the authors concluded that HTLV-1 Tax protein is a major determinant of the difference between HTLV-1 and HIV-1 transmission.
The cross-linking of ICAM-1 (reviewed in [54]) activates two distinct signaling pathways: one is RhoA-dependent [55], and the other involves Ras-MEK-ERK activation [56]. We previously showed that MTOC polarization in HTLV-1-infected cells is independent of RhoA GTPase activation [35]. Our recent data, obtained both in naturally infected lymphocyte and transfected Jurkat T-cells, confirms that MTOC polarization induced by the cross-linking of ICAM-1 depends on the Ras-MEK-ERK pathway [36]. This pathway is distinct from that activated by TCR engagement [36]. Indeed, HTLV-1 infection significantly reduced the frequency of MTOC polarization caused by cross-linking TCR (CD3) on the cell surface [47], which suggests a competition between the two pathways.
The identification of the major triggers involved in causing preferential microtubule polarization towards the VS provided a possible resolution of the long-standing paradox that HTLV-1 infection is almost confined to T cells in vivo, whereas it can infect almost any nucleated mammalian (and some avian) cells in vitro. First, the role of ICAM-1 can explain preferential transmission to LFA-1-positive cells, the great majority of which are T cells. Second, the role of microtubule polarization suggests that HTLV-1 has evolved specific mechanisms to subvert the unusual mobility of the T cell cytoskeleton in order to propagate. This in turn can explain the observation that, whereas HTLV-1 can infect other cell types in vitro such as epithelial cells and fibroblasts [29,57], it is not possible to propagate HTLV-1 in these cell types. That is, HTLV-1 can enter other cell types, but the infected cell is unable to pass the virus on to other cells.
In view of the observed sensitivity of the T cell's microtubule cytoskeleton to engagement of a range of cell-surface molecules, it is likely that several ligand-receptor interactions contribute to cellto-cell transmission of HTLV-1 [58,59]. However, the conclusion that ICAM-1 plays a particularly important role is reinforced by the observations that cross-linking of ICAM-1 can increase HTLV-1 protein expression [60], and conversely, HTLV-1 infection upregulates expression of ICAM-1 on the infected cell surface [61][62][63][64].
Transfer of viral protein and genome to target cells
Confocal microscopy was used to examine the distribution of HTLV-1 Gag and Env proteins and the HTLV-1 genome in fresh, unstimulated peripheral blood mononuclear cells (PBMCs) isolated directly from HTLV-1-infected individuals [32]. CD4 + or CD8 + T cells isolated from a HAM/TSP patient were allowed to form conjugates with T cells from a healthy uninfected donor for 120 min: in addition to accumulation of Gag p19 staining at the cell-cell junction, there was frequent Gag p19 staining in the cells derived from the uninfected donor [32,35,41]. In addition to Gag protein, HTLV-1 RNA was also transferred from infected cell to uninfected cell [32]. The HTLV-1 Gag p19 was observed to transfer from CD4 + T cells and CD8 + T cells to both CD4 + and CD8 + allogeneic T cells. This process may represent the initial establishment of HTLV-1 infection in a newly infected individual, which involves contact between allogeneic lymphocytes. Polarization of Gag complexes to the cell-cell junction and transfer to the uninfected cell were also observed in conjugates between CD4 + T cells and both B cells and NK cells [32]. The kinetics of HTLV-1 Gag transfer showed a peak intensity between 90 and 210 minutes after initiation of conjugate formation [36]. This kinetics closely resembles the kinetics of the cell-to-cell transfer of HIV-1-Gag recently reported [65]. HTLV-1 transfer is significantly reduced within this time frame either by depolymerization of the cytoskeleton (microtubules or microfilaments) or by inhibition of ERK phosphorylation and CREB activation in the infected donor T-cell [36].
Role of microtubule cytoskeleton in the cell-to-cell transfer of Gag protein
The microtubule organizing center (MTOC) is typically oriented to the area of cell-to-cell contact in lymphocyte conjugates where it lies immediately adjacent to the accumulation of HTLV-1 Gag protein [32,35,36]. This close apposition of polarized Gag molecules to the MTOC suggested that the microtubule cytoskeleton affected the polarization of Gag. In conjugates formed between infected and uninfected CD4 + T cells, depolymerization of the microtubule network blocked the polarization and transfer of Gag protein [32,35,36]. There is a significant association between MTOC polarization and Gag positivity in conjugates between autologous CD4 + and CD8 + T cells from an infected individual [32].
Recently, the HTLV-1 accessory protein p12 was reported to reduce the expression of ICAM-1/2, which in turn prevents the destruction of HTLV-1-infected cells by NK cells [66]. In addition p12 has been shown to induce LFA-1 clustering on the surface of infected T cells via a calcium-dependent signaling pathway, which is thought to promote HTLV-1 spread between T cells [67]. HTLV-1 Tax expression was also shown to strongly up-regulate ICAM-1 expression [61]. Thus, HTLV-1 appears to regulate the expression of adhesion molecules at the surface of infected T-cell using both Tax and p12 protein. However, the early expression of Tax protein and its powerful effect on ICAM-1 suggests that Tax will dominate -and so upregulate ICAM-1 -in the early stages of the infectious cycle, so promoting the formation of the VS.
The association of Tax protein with the MTOC region in the infected T cell and its ability to activate CREB pathway are both required for the polarization observed during the formation of the HTLV-1 VS [36]. Tax also binds directly to the protein kinase MEKK1 [68], and this interaction might play a part in the synergistic induction of MTOC polarization by Tax and ICAM-1. However, Tax protein interacts with a large range of cellular proteins, including Ras p21 proactive 2 and cdc42/Rac effector kinase [69], which are a part of the Ras activation pathway.
Ultrastructure of the VS: Electron tomography
We investigated the three dimensional ultrastructure of the HTLV-1 VS using electron tomography (Figure 2
) of immunostained conjugates formed between CD4 + T cells from HTLV-1-infected individuals and conjugates between MS9 cells (a chronically HTLV-1-infected cell line) and Jurkat cells [41].
Enveloped HTLV-1 particles were identified in an intercellular space or "synaptic cleft" in the VS formed between a naturally infected lymphocyte and an autologous target lymphocyte (Figure 2). Each synaptic cleft is characteristically bounded by the tightly apposed membranes of the infected and target cells, with a membrane-membrane distance of ~20 nm (median 25.7 nm). This distance is consistent with measurements of intercellular separation between lymphocytes and target cells, and with data on the conformation of ICAM-1/LFA-1 binding from electron microscopy and X-ray crystallography [70][71][72]. However, the ultrastructural studies of the VS show that in the case of HIV-1 the intercellular space is less tight than in the HTLV-1 VS and is therefore likely to be accessible to inhibitors both during and after virus assembly [73]. The HTLV-1 VS typically contains more than one synaptic cleft, spatially distinct and separated by areas of close membrane-membrane apposition (Figure 2, A-C). The membrane of the synaptic pockets presents sites of virus budding, and the pockets contain enveloped HTLV-1 particles [41]. This contrasts with the immunological synapse where secretory lysosomes are secreted into a single synaptic pocket in the cytotoxic synapse [74,75]. The size of HTLV-1 particles that originate from naturally infected CD4 + T cells has a wide range (62 nm-173 nm), with a peak at about 100 nm similar to the variability in size of HIV-1 particles [76,77]. The directional nature of the process is indicated by the polarization of Gag and Tax proteins, and the MTOC [32,35,47]. The accumulation of mitochondria near the VS (Figure 2, D) again resembles the immunological synapse [41,75].
The ultrastructure of the HTLV-1 VS reconciles the requirement for cell contact and HTLV-1 Env protein for the spread of the virus, with the lack of detection of cell-free virions in the serum. Whereas most viruses spread by releasing large numbers of virions from the infected cell, HTLV-1 uses the mobility of the host cell to propagate from cell to cell. Sequestration in the synaptic cleft presumably allows efficient transfer of small numbers of virions, and may give the virions a degree of protection from components of the immune response (complement system and antibodies) (Figure 3, A1).
Recent advances
New routes have been recently described that may contribute to HTLV-1 spread in vivo. Jones and co-authors investigated the role of human dendritic cells (DCs) in the transmission of HTLV-1 [78]. DCs are potent antigen-presenting cells that play a central role in initiating the immune response. However, infection of DCs can impair their ability to mount an appropriate immune response. Indeed, many viruses infect DCs to facilitate their transmission including the retrovirus HIV-1 and the mouse mammary tumor virus [79][80][81]. Some viruses use the trafficking proprieties of DCs to facilitate their transport from the periphery to lymph nodes where they infect target cells.
Jones et al. have shown that myeloid DCs (mDC) and plasmacytoid DCs (pDC) are efficiently and reproducibly infected in vitro with cell-free HTLV-1 virions released by chronically infected T cell lines (MT2, DB1) [78]. The infected DCs can rapidly transfer HTLV-1 to autologous primary CD4 + T-cells, resulting in a chronic productive infection of CD4 + T cells ex vivo. The DC-T cell transmission of HTLV-1 is reduced by blocking Neuropilin-1 (NP-1) and heparin sulfate proteoglycans (HSPGs) [78], two molecules involved in the initial interaction of HTLV-1 with CD4 + T-cells [82][83][84][85] and in the DC-T cell interaction [86,87]. DC-specific ICAM-3-grabbing nonintegrin (DC-SIGN) was also shown to facilitate the interaction of dendritic cells with HTLV-1-infected cells [88]. Recently it was reported that blocking any one of four molecules HSPGs, NRP-1, GLUT-1 or DC-SIGN led to reduction of virus binding to the cell; however, HTLV-1 transmission from DCs to T cells was mediated primarily by DC-SIGN [89]. Based on their findings and on the studies reporting the presence of viral proteins [90] and genome [91,92] in DCs isolated from HTLV-1-infected individuals, Jones et al. suggested that cell-free virus can also infect DCs in vivo (Figure 3, A2). It is possible that DCs play an important part in the early stages of HTLV-1 infection of a new host, by efficiently acquiring the virus and disseminating it to T cells. This system provides a new approach for dissection of the early events in HTLV-1-induced transformation and HTLV-1 infection.
Figure 3. Illustration of possible routes of HTLV-1 spread between cells in vivo.
(A) cell-to-cell transmission via intercellular synaptic cleft surrounded by a tight cell-cell contact between the donor and recipient cell, 1. Cell-free particles can be internalized by dendritic cells and are then transferred to lymphocyte only by cell-to-cell contact, 2. Virus particles can be retained on the cell surface in a biofilm-like-structure before lateral transfer to the recipient outside the cell-cell contact region, 3. (B) In an isolated HTLV-1infected lymphocyte, the viral proteins are not polarized.
Pais-Correia et al. have reported evidence that HTLV-1 can also be transferred from cell to cell at the VS via biofilm-like extracellular viral assemblies [93]. By using electron and light microscopy analysis they detected extracellular viral clusters at the cell surface of HTLV-1 infected cell after overnight incubation in vitro. These viral assemblies are carbohydrate-rich structures, suggesting the involvement of matrix-linker proteins in the adhesion and attachment of HTLV-1 to the surface of infected cells. The authors concluded that the extracellular matrix component and cellular lectins together generate a cocoon-type structure that concentrates virions in a confined protective environment to escape the immune response. Consistent with the tomographic studies of HTLV-1-VS [41] the authors observed that infected T cells formed a tight contact with the target cell. However, their microscopic analysis suggested that many virions cluster overlapped the cell contact, bridging the gap between the two T cell surfaces [93]. Extensive washing or heparin treatment of chronically infected cells and primary infected CD4 + cells reduces their capacity to infect reporter target cells [93]. In contrast with an earlier report [78], the cell-free HTLV-1 particles recovered from the cell surface clusters by extensive washing and heparin treatment were able to infect target T cells, although less efficiently than by cell-to-cell contact at the VS.
Recent studies on Murine Leukemia Virus (MLV), a retrovirus related to HTLV-1, have quantified the sequential assembly and transmission events for individual viral particles in living cells [94,95]. The authors showed that MLV Gag protein is present in 10-fold greater amounts at the cell-cell contact area compared to the periphery and that the de novo assembly of virions is highly polarized toward the zone of cell-to-cell contact [94]. They also observed that a fraction of 30% of MLV particles is retained at the infected cell surface after completion of assembly, and they named this fraction "surfacing viruses". These virions are competent for transmission to infected target cells after the establishment of a physical cell-to-cell contact [95]. This fraction may represent a reservoir of infectious virus particles that can be sequestered and later passed on to uninfected target cells upon the establishment of cell-to-cell contacts.
We conclude that HTLV-1, like other retroviruses (notably HIV-1 and MLV), can be transferred between cells at more than one locus at the VS. The relative magnitude of cell-to-cell transmission of virions that bud within the cell contact zone and preformed virions at the periphery of the cell contact zone remains to be quantified (Figure 3, A3); this ratio may depend on the time interval between the onset of viral antigen expression and the formation of a cell-cell contact.
To identify and quantify the mechanisms involved in the formation and function of the HTLV-1 VS, experimental work in vitro has focused chiefly on two-cell conjugates. However, it is likely that, in vivo, an HTLV-1-expressing T-cell will make contact with more than one cell simultaneously in the low-velocity, cell-rich environment of the lymphoid circulation. HIV-1 has indeed been shown to spread in culture from one infected cell to more than one recipient cell: the authors called the multiple contact a polysynapse [96]. These rosette-like structures, displaying Gag accumulation at each cell contact, were observed in infected lymphocytes and DCs. Microscopic analysis suggests that the cell contacts in the polysynapse can be formed simultaneously rather that one after another. Multifocal capping of assembled budding virions required for polysynapse formation is promoted by tetraspanins and plasma membrane mobility via lipid rafts. The actin and microtubules are also involved because nocodazole, cytochalasin D and latrunculin B inhibit the formation of polysynapses. The description of the virological polysynapse recalled the studies of multiple immunological synapse formation by T cells that interact simultaneously with cells presenting different antigenic stimuli [97]. These studies showed that the MTOC moves repeatedly between the different cell-cell contacts; eventually the secretory machinery becomes selectively polarized toward the antigen-presenting cell that provides the strongest stimulus. The intensity of molecular translocation at the different synapses reflects the strength of signals received via the T-cell receptor and accessory molecules [97].
The present studies carried out in vitro and ex vivo indicate that HTLV-1 may use more than one route to spread between CD4 + T-cells. However, it is clear that efficient propagation of HTLV-1 depends on the close contact whose formation is triggered by the virus infection in the VS. While it is difficult to completely exclude the possibility of cell-free virion transmission of HTLV-1 in vivo, cell-free HTLV-1 particles are typically undetectable in serum even by RT-PCR, and transfusion of plasma or the cell-free blood product does not transfer HTLV-1 infection. It is possible that transfer of viral particles at the periphery of the VS involves the same molecular mechanisms that cause the polarization of HTLV-1 protein and the cytoskeleton, as described above. However, this remains to be tested.
Mechanism of HTLV-1 entry: Comparison of two possible mechanisms of HIV transmission
The mechanism by which HTLV-1 penetrates the target cells is yet to be established. Endocytosis is an obligatory entry step for enveloped viruses whose fusion proteins are activated by acidic pH [98]. However, HIV can mediate fusion between adjacent target cells ("fusion from without") and HIV Env expressed on effector cells promotes fusion with target cells at neutral pH [99].
Fusion from without at the cell membrane in intercellular pockets at the VS
Certain enveloped viruses such as herpes simplex virus 1 (HSV-1), Sendai virus, and many retroviruses, including HIV, have pH-independent fusion proteins and can therefore penetrate into cells by fusing directly with the plasma membrane. It is generally assumed that fusion events at the plasma membrane lead to productive infection, although this is difficult to prove because virus particles are also continuously endocytosed [98].
Penetration of enveloped viruses occurs by membrane fusion catalyzed by fusion proteins in the viral envelope. The machinery involved is rather simple, at least when compared to the apparatus needed for intracellular membrane-fusion events. One reason for simplicity is that viral fusion factors are used only once. Fusion activity is triggered by cues in the form of receptor binding or low pH. They induce, as a rule, irreversible conformational changes. Membrane fusion is an elegant and effective way to deliver viral capsids into the cytosol. No macromolecular assemblies need to pass through a hydrophobic membrane barrier. The underlying principle is the same as in intracellular membrane traffic; the viral envelope is a "transport vesicle", and the capsid is the cargo [100].
Endocytosis
The endocytic entry route gives many advantages to the virus: (i) the endocytic vesicles are designed to traverse the barriers imposed by the cortical cytoskeleton and the highly structured cytoplasm, so endocytosis gives the virus efficient access into the cytoplasm; (ii) the incoming viruses are exposed to compartmental environments that differ from the extracellular milieu, in particular the mildly acidic pH in endosomes provides an essential signal that triggers penetration and uncoating; (iii) if the penetration is lytic, endosomal membrane lysis is likely to be less damaging to the cell than lysis at the plasma membrane. However, it is possible that endocytosis leads to delivery of the virus to the lysosome, a degradative compartment and a dead-end for most viruses. This is why viruses have adapted to carefully adjust the threshold pH for activation to match that of early (pH 6 to 6.5) or late endosomes (pH 5 to 6) [100]. The early and late endosomes constitute distinct entry sites: this has been confirmed with dominant negative mutants of endosome-associated small guanosine triphosphatases (GTPases) [101]. A constitutively inactive mutant of Rab5 (early endosome) blocked the entry of both Semliki Forest virus (pH 6.2) and influenza virus (pH 5.4), whereas the corresponding Rab7 mutant (late endosomes) only blocked influenza virus entry. The progress of individual virus particles through endocytic compartments can be tracked with real-time video microscopy [65,102]. Individual fluorescent virus particles can be observed to bind to the cell surface, diffuse along the membrane, get trapped in coated pits or caveolae, enter by endocytosis, and move along microtubules. With the use of specific fluorescent dyes, the acidification of virus particles and the fusion of the viral envelope with cellular membranes can also be monitored.
1.
A VS may be defined as a virus-induced, specialized area of cell-cell contact that promotes the directed transmission of the virus between cells.
2. HTLV-1 relies almost exclusively on cell-to-cell transmission to spread, both within the host and between hosts. Only dendritic cells can be efficiently infected with cell-free HTLV-1: this route may be important during initial acquisition of infection. However, DCs can infect T cells only by cell contact.
3) The HTLV-1 VS has a large area of close (about 26 nm) apposition of the plasma membranes of the infected cell and the target cell; there is no evidence of fusion between the plasma membranes of the two cells. The viral proteins Gag, Env and Tax are polarized at the cell-cell junction.
4) The integrity and function of both actin and microtubule cytoskeletons are required for the formation of the HTLV-1 VS and for transfer of HTLV-1 between cells. The MTOC is polarized to the cell-cell junction. 5) HTLV-1 Tax protein acts in synergy with crosslinking of ICAM-1 on the infected cell surface, to cause polarization of the infected cell's microtubule cytoskeleton to the VS. This polarization requires the presence, in the vicinity of the MTOC, of a Tax molecule competent to activate the CREB pathway. The ERK pathway is also required for MTOC polarization at the VS. 6) An HTLV-1-infected cell can form more than one VS simultaneously (a "polysynapse"). The multiple contacts strongly resemble the multiple immunological synapses previously demonstrated. As in the multiple immunological synapses, it is likely that the dynamic microtubule cytoskeleton of the T cell moves between the respective cell contacts in a multiple VS.
7) The closely apposed cell membranes at the VS are interrupted by multiple intercellular clefts or 'pockets'. Immunoelectron tomography has shown HTLV-1 Gag staining localized under the plasma membrane in these pockets. The pockets contain Gag-staining particles within the intercellular clefts consistent with the size and morphology of virions. Virions may be transferred to the target cell either across these clefts, or at the periphery of the cell contact, or (probably) both. 8) Env protein is required for infectivity of HTLV-1. Both endocytosis and fusion from without (virion fusion with the plasma membrane) may contribute to infection of the target cell; the relative importance of the two routes remains unknown.
9) The HTLV-1 VS appears to differ from the HIV-1 VS in the closeness of the cell-to-cell contact and the enhancing role of HTLV-1 Tax protein. | 8,082.2 | 2010-07-01T00:00:00.000 | [
"Biology",
"Medicine"
] |
A sample preparation workflow for adipose tissue shotgun proteomics and proteogenomics
ABSTRACT Animals with large adipose stores, such as marine mammals, may provide insights into the evolution and function of this multifunctional tissue in health and disease. In the absence of sequenced genomes, molecular information can be rapidly obtained by proteomics and transcriptomics, but their application to adipose tissue is hindered by low nucleic acid and protein yields. We sequenced and compared proteomes isolated from the blubber of four elephant seals using phenol and guanidine thiocyanate (Qiazol) or detergent (sodium deoxycholate) buffer. Qiazol recovered more subcellular proteins such as metabolic enzymes, in addition to extracting RNA, facilitating proteogenomic analyses of small lipid-rich tissue biopsies. We also compared proteomics data analysis platforms and found that de novo peptide sequencing improved protein identification sensitivity compared to database search alone. We report sample preparation and data analysis workflows for proteogenomics and a proteome of elephant seal blubber containing 2678 proteins, including many of interest for further functional studies. This article has an associated First Person interview with the first author of the paper.
INTRODUCTION
Adipose is a complex organ that participates in energy storage, thermogenesis, immunity and regulation of metabolic homeostasis. Specialized fat deposits arose early in the metazoan lineage and have supported evolutionary adaptations such as migration, hibernation and lactation, among others (Birsoy et al., 2013;Pond, 2017). While adipose tissue has been studied extensively in humans and laboratory animals due to the emergence of a global obesity epidemic, non-model organisms can provide fundamental information on metabolic adaptations in animals and potentially novel insights into mechanisms by which adipose function is dysregulated in disease (Grabek et al., 2015;Houser et al., 2013). Such insights can be rapidly obtained via non-targeted approaches such as transcriptomics and proteomics in the absence of available genomes (Nesvizhskii, 2014).
With some of the largest subcutaneous adipose stores (modified as blubber) in the animal kingdom, marine mammals hold valuable information about rapid fat accrual and loss, lipid-based metabolism and the physiological effects of lipophilic pollutants (Bossart, 2011;Houser et al., 2013). Several recent studies have used omics technologies to profile gene and protein expression in blubber (Brown et al., 2017;Kershaw et al., 2018;Khudyakov et al., 2017;Van Dolah et al., 2015). However, widespread application of these approaches to marine mammal systems is hindered by the technical challenge of obtaining sufficient quantities of nucleic acids and proteins from small biopsies of lipid-rich tissues with low nuclear and cytoplasmic content. Indeed, most proteomics studies of marine mammals to date have used tissue matrices other than blubber (Neely et al., 2015a, b;Sobolesky et al., 2016).
Nucleic acids and proteins are commonly isolated from cells and tissues using separate pipelines, with phenol and guanidine thiocyanate solutions (e.g. Trizol®) used for RNA and DNA extraction and detergents (e.g. sodium dodecyl sulfate, SDS) for protein isolation (Feist and Hummon, 2015;Zhang et al., 2013). The Trizol® reagent was originally developed for simultaneous extraction of DNA, RNA, and proteins (Chomczynski, 1993), but its effectiveness for RNA and protein extraction from adipose, and proteome completeness and compatibility with tandem mass spectrometry (MS/MS), have not yet been described. In this study, we compared two lysis buffers for shotgun proteome sequencing of adipose tissue: (1) a detergent buffer containing sodium deoxycholate (SDC method), and (2) Qiazol®, a solution similar to Trizol® that was developed for nucleic acid isolation from lipid-rich samples (QIA method). We used blubber samples collected from four juvenile northern elephant seals (Mirounga angustirostris), a fasting-adapted marine mammal study system frequently used in comparative metabolic physiology studies, including analyses of cellular responses to stress and fasting (Khudyakov et al., 2017;Martinez et al., 2018). Marine mammal blubber is vertically stratified by fatty acid composition and function; the outer layer plays a role in thermoregulation while the inner layer is more metabolically active (Strandberg et al., 2008). We used the outer half of blubber biopsies in this study, as this layer is typically sampled by remote biopsy dart from many marine mammals (Hunt et al., 2013).
We show that a larger number of unique proteins, including those involved in metabolism and protein translation, can be identified in samples lysed using QIA with the added benefit of RNA isolation, and that de novo peptide sequencing (PEAKS Studio) combined with database search increases sensitivity of protein identification compared with database search alone (SEQUEST in Proteome Discoverer). We report the first elephant seal outer blubber layer proteome containing a number of metabolic enzymes and adipokines of interest to The study was conducted using four biological replicates of blubber tissue. Proteins were isolated from one half of each sample using the SDC method and RNA and proteins were isolated from the other half using the QIA method. SDC yielded 1.85-fold more protein than QIA ( paired t-test, t=4.48, P<0.05). Mean protein yields per mg wet mass tissue for SDC and QIA were 13.7 µg/mg (s.d.=1.8) and 7.4 µg/mg (s.d.=2.2), respectively. Lower protein yields have been reported for Trizol® compared with detergent-based buffers (Yamaguchi et al., 2013). However, QIA also recovered mean 4.6 µg (s.d.=4.2) of total RNA per sample. Variability in RNA yields between samples (range: 1.78-10.89 µg) could be due to decreased efficiency of homogenization and silica column-based RNA purification with higher tissue inputs. RNA isolated by QIA from blubber samples had high purity and integrity. The mean 260/ 280 and 260/230 ratios were 1.99 (s.d.=0.07) and 1.75 (s.d.=0.25), respectively. The mean RNA integrity number (RIN) was 7.67 (s.d.=0.39) and mean rRNA ratio was 0.81 (s.d.=0.12) (Fig. S1). RNA integrity was above the threshold (RIN 7) commonly recommended for RNA sequencing (Gallego Romero et al., 2014).
Protein identification
After processing using standard methods [ Fig. 1; Bodzon-Kulakowska et al. (2007)], protein samples were analyzed by HPLC-MS/MS, producing mean 26,621 (s.d.=1031) MS/MS spectra for SDC samples and mean 25,712 (s.d.=1374) MS/MS spectra for QIA samples. We first performed peptide spectrum matching and a SwissProt database search using SEQUEST in Proteome Discoverer. SEQUEST identified 13.3% and 16.1% of all MS/MS spectra from SDC and QIA samples, respectively, which is within the range of 10-30% reported in the literature (Houel et al., 2010). QIA samples had 1.17-fold more peptide spectrum matches (PSMs; F 1,10 =100.45, P<0.0001), 1.13-fold more peptide groups (F 1,10 =67.01, P<0.0001), 1.21-fold more protein groups (F 1,10 =59.09, P<0.0001), and 1.25fold more unique proteins (with two or more unique peptide hits; F 1,10 =24.54, P<0.001) than SDC samples ( Fig. 2A). Therefore, while total protein yields were lower with QIA, this method produced more identified peptides and proteins than SDC. We then repeated the SEQUEST search with a custom database of a translated elephant seal blubber transcriptome (Khudyakov et al., 2017 ). Numbers of PSMs and proteins were significantly different between methods (PSMs: F 1,3 =17.16, P<0.05; protein groups: F 1,3 =18.52, P<0.05; unique proteins: F 1,3 =16.33, P<0.05). Therefore, we identified hundreds of proteins predicted from the elephant seal transcriptome by MS/MS, validating our transcriptome and proteome methods and providing a workflow for proteogenomics. However, since the translated transcriptome was annotated by BLASTP using the same database, we focused subsequent functional analyses on proteins identified directly by SEQUEST SwissProt database search.
Functional annotation of proteins
To facilitate functional analyses, we repeated SEQUEST SwissProt database search with biological replicates concatenated as 'fractions' into one pooled sample for each QIA and SDC. We identified 976 and 701 proteins from pooled QIA and SDC samples, respectively (2.5-fold more unique proteins with QIA than SDC; Fig. 2B). The top 5 (by number of proteins) KEGG categories overrepresented in both datasets were focal adhesion, PI3K-Akt signaling pathway, biosynthesis of antibiotics (which includes many enzymes involved in lipid metabolism), ECM-receptor interaction, and carbon metabolism (Fig. 3A). KEGG categories enriched uniquely in the QIA dataset were ribosome, pyruvate metabolism, pentose phosphate pathway, valine, leucine and isoleucine degradation, while those unique to SDC were platelet activation and small cell lung cancer. Therefore, the QIA method recovered more proteins involved in metabolism and protein synthesis than SDC.
We identified 37 and 41 GO biological process (BP) categories enriched in the QIA and SDC datasets, respectively (Fig. 3B). The top overrepresented categories in both datasets were associated with cell-matrix interactions and protein folding. BP categories unique to the QIA dataset included translation, mRNA splicing, rRNA processing, gene expression, pentose-phosphate shunt and response to calcium ion. BP categories that were enriched only in the SDC dataset included osteoblast differentiation, epithelial cell differentiation, muscle contraction, MAPK cascade, tumor necrosis factor-mediated signaling pathway, cellular response to transforming growth factor beta stimulus, regulation of cell migration, nucleosome assembly and blood coagulation. While both methods recovered proteins involved in cell-cell and cellmatrix adhesion, QIA isolated additional proteins were involved in mRNA processing and protein synthesis, while SDC recovered more proteins involved in cell signaling and differentiation.
There were 22 and 16 GO molecular function (MF) categories enriched in the QIA and SDC datasets, respectively (Fig. 3C). Top MF categories for each method were protein and RNA binding and cadherin-mediated cell-cell adhesion. MF categories enriched only in the QIA dataset were nucleotide binding, structural constituent of ribosome, protein complex binding, heparin binding and chaperone binding. MF categories unique to the SDC dataset were oxidoreductase activity and platelet-derived growth factor binding. Therefore, while representation of molecular functions was similar in both protein datasets, some proteins with oxidoreductase activity of interest to marine mammal physiology (e.g. fatty acid synthase, peroxiredoxin, alcohol dehydrogenase) were not isolated by QIA.
We identified 50 and 48 GO cellular component (CC) categories that were enriched in the QIA and SDC datasets, respectively (Fig. 3D). Top CC categories for both methods were extracellular exosome, cytoplasm/cytosol, membrane, and extracellular space/ matrix. CC categories enriched only in the QIA dataset included perinuclear region of cytoplasm, mitochondrial matrix, ribosome, extrinsic component of membrane, spliceosomal complex and nuclear matrix. CC categories enriched only in the SDC dataset were actin cytoskeleton, protein complex, cell-cell junction, lysosomal lumen, sarcolemma, lipid particle, smooth endoplasmic reticulum and lamin filament. The large number of extracellular proteins recovered by both methods is consistent with thick basal lamina and abundant connective tissue proteins characteristic of adipose tissue (Mariman and Wang, 2010). The abundance of extracellular vesicle (EV)-related proteins recovered by both methods is interesting due to their potential role in regulation of metabolism and immunity in adipose tissue (Gao et al., 2017). However, the QIA method isolated more subcellular proteins than SDC, including those associated with the nucleus and mitochondria, suggesting that it may be more efficient at solubilizing intracellular membranes than detergent.
De novo peptide sequencing
We performed de novo peptide sequencing combined with PEAKS DB SwissProt database search (Zhang et al., 2012) to evaluate whether this approach would increase sensitivity of peptide/protein identification. PEAKS produced 4534 (s.d.=567) and 4443 (s.d.=671) de novo-only spectra for samples prepared using SDC Fig. 2. Comparison of proteins identified using two sample preparation and two data analysis workflows. (A) Numbers of peptide spectrum matches (PSMs) and peptides, protein groups, and unique proteins (with two or more unique peptide hits) identified by SEQUEST or PEAKS in samples prepared using SDC and QIA. SwissProt (2/13/2018) database was used for searches and only hits with false discovery rates (FDR) <1% were retained. Asterisks denote significant differences between sample preparation methods and software platforms: ***P<0.0001, **P<0.001. Sets of identified proteins were compared between (B) SDC and QIA methods and SEQUEST database search, (C) SEQUEST and PEAKS protein identification methods for QIA samples, and (D) between each sample preparation method and software platform. and QIA, respectively. For both sample preparation methods, PEAKS had approximately 1.10-fold more PSMs (F 1,10 =44.27, P<0.0001) and identified 1.27-fold more peptides (F 1,10 =242.24, P< 0.0001) and 2.24-fold more unique proteins (F 1,10 =173.73, P<0.0001) than SEQUEST ( Fig. 2A). However, the number of protein 'groups' identified by PEAKS was approximately 1.95-fold lower than SEQUEST (F 1,10 =572.21, P<0.0001) due to differences in protein isoform clustering approaches (Paulo, 2013). Therefore, combined de novo peptide sequencing and PEAKS DB database search had greater sensitivity than a SEQUEST database search alone, as previously reported (Zhang et al., 2012).
Lastly, we compared the sets of unique proteins identified by SEQUEST and PEAKS using the SwissProt database in the pooled QIA samples. The two software platforms identified a common set of 595 proteins; an additional 1258 were identified only by PEAKS and 381 were identified only by SEQUEST (Fig. 2C). Different protein identification approaches and database search algorithms are known to produce different sets of proteins from the same mass spectra, even with identical search parameters (Paulo, 2013;Russeth et al., 2006). Integrating results from multiple search engines can increase the number of identified proteins and validate those that are commonly identified by different algorithms, and several bioinformatics tools have been developed for this purpose (Paulo, 2013).
Overall, we identified overlapping, but distinct sets of proteins from blubber samples using two different sample preparation methods and two different MS/MS data analysis platforms (Fig. 2D). Proteins of interest to the metabolic and comparative physiology communities identified in the study include fatty acid transporters, lipid droplet proteins, and lipid metabolism enzymes (Table S3). In total, we identified 2678 proteins from the outer blubber of northern elephant seals, of which 286 were common to all four pipelines used in the study. Differences in numbers and types of proteins isolated by the QIA and SDC methods may be attributed to different membrane solubility efficienciesreagents such as Trizol® may be more efficient than detergents at removing lipids and carbohydrates to liberate proteins (Kirkland et al., 2006). However, some have suggested that tissue lysis in phenol and guanidine thiocyanate may also lead to loss of highly hydrophobic proteins, an important consideration for adipose proteomics (Butt et al., 2007;Kirkland et al., 2006). This may be improved by optimizing solubilization conditions for proteins precipitated after Trizol® or Qiazol® extraction (Kopec et al., 2017).
CONCLUSIONS
We found that blubber tissue lysis in Qiazol® increased the total number of identified proteins and enabled simultaneous isolation of high-quality RNA from the same tissue samplesignificant advantages for researchers working with small quantities of tissue and for those interested in proteogenomics. Moreover, QIA recovered more subcellular proteins, including proteins involved in metabolism and protein synthesis, than SDC. We also found that sensitivity of protein identification in a non-model organism could be significantly improved using PEAKS de novo peptide sequencing in combination with a database search. Proteins identified in this study may be of interest for functional studies of marine mammals and other species in which large adipose stores play key roles in physiology.
Materials
All chemicals were proteomics grade and purchased from VWR Life Science/Amresco (USA) or Thermo Fisher Scientific (USA), unless otherwise indicated.
Sample collection
All animal handling procedures were approved by University of the Pacific and Sonoma State University Institutional Animal Care and Use Committees and conducted under National Oceanic and Atmospheric Administration Fisheries Permit No. 19108. Four juvenile (∼0.8-year old) female northern elephant seals (M. angustirostris) were sampled at Año Nuevo State Reserve (San Mateo County, CA, USA) in October 2017. Animals were chemically immobilized as previously described (Khudyakov et al., 2017). Blubber samples were collected from the posterior flank of the animal using a sterile 6.0 mm diameter biopsy punch (Miltex, USA), blotted on sterile gauze to remove blood, and separated into two halves: an inner (closest to muscle) blubber half and an outer (closest to skin) blubber half.
Samples were placed into plastic cryogenic vials (Corning, USA), flashfrozen in liquid nitrogen, stored on dry ice, and transferred to a −80°C freezer upon return to the laboratory.
Sample preparation
Only the outer blubber half of each biopsy was used for this study (inner blubber was used for a separate study). Each outer blubber sample was weighed and minced into small pieces, which were randomly divided into two portions of approximately 100 mg each. One portion of each sample was used for protein extraction using the detergent method (mean wet mass 97.5 mg, s.d.=8.6), while the other was used for RNA and protein extraction using the Qiazol® method (mean wet mass 102.8 mg, s.d.=9.0).
Protein extraction using detergent (SDC method)
The SDC method of protein extraction was adapted from a previously published protocol (Pasing et al., 2017). Blubber was processed in two batches of ∼50 mg, which were minced with a sterile scalpel on ice and added to 500 µl SDC Lysis Buffer [1% w/v SDC, 8 M urea, 5 mM dithiothreitol (DTT) in 50 mM ammonium bicarbonate] in a Navy RINO ® bead tube (Next Advance Inc., USA). Samples were homogenized in the Bullet Blender Storm ® instrument (Next Advance Inc., USA) for two cycles of 2 min each at power 10, with 1 min of cooling on ice between cycles. Homogenates were further disrupted by sonication for three cycles, 15 s each, at 4 watts using a hand-held sonicator (VirSonic 60, Virtis, USA) and centrifuged to pellet insoluble cell debris and separate lipids. Tissue homogenates were extracted from under the top lipid layer and transferred to clean tubes. To remove any remaining lipids, four volumes of methanol and one volume of chloroform were added to homogenate aliquots, mixed, and centrifuged. The top layer containing lipids was removed, and four volumes of methanol were added to precipitate proteins. Protein pellets were air dried after centrifugation.
Protein and RNA extraction using phenol-chloroform (QIA method) The QIA method of protein extraction was adapted from the Chomczynski protocol (Chomczynski, 1993). Approximately 100 mg of blubber were minced with a sterile scalpel on ice and added to 500 µl of Qiazol® reagent (Qiagen, USA) in a Navy RINO® RNase-free bead tube (Next Advance, USA). Homogenization was conducted in the Bullet Blender Storm® (Next Advance, USA) as described above. An additional 500 µl of Qiazol® reagent was added to each tube and incubated for 5 min at room temperature with occasional vortexing. Homogenates were further disrupted using a syringe and 21-gauge needle and centrifuged to pellet insoluble cell debris and separate lipids. The homogenate was extracted from under the top lipid layer and transferred to clean tubes with 200 µl chloroform, vortexed to mix, and centrifuged at maximum speed for 15 min at 4°C to separate phases. The aqueous layer containing RNA was transferred to a clean tube and RNA purification was performed using RNeasy® Lipid Tissue Mini Kit (Qiagen, USA) (Khudyakov et al., 2017). After aqueous phase and interphase extraction, 300 µl of 100% ethanol was added to the organic phase and centrifuged to precipitate DNA. The supernatant was split into two microcentrifuge tubes and each was incubated for 10 min with 750 µl of 100% isopropanol and centrifuged again to pellet proteins. The pellets were washed twice for 20 min with 1 ml of 0.3 M guanidine hydrochloride in 95% ethanol, once with 1 ml of 100% ethanol and air dried. If not processed the same day, protein pellets were kept at −80°C in guanidine wash solution.
Protein denaturation, digestion, and desalting
Proteins isolated by SDC and QIA methods were treated identically after precipitation. Pellets were resuspended in 250 µl of Denaturing Buffer (1% w/v SDC, 8 M urea, 5 mM DTT in 50 mM ammonium bicarbonate) by homogenization with a 21-gauge needle and continuous vortexing for 1 h at room temperature. Protein samples were then incubated at 37°C for 1 h in Denaturing Buffer, followed by alkylation with 15 mM iodoacetamide in the dark at room temperature for 30 min. Alkylation was quenched by addition of DTT to 5 mM final concentration. Samples were diluted with 50 mM ammonium bicarbonate to reduce urea concentration to <2 M and protein concentration was estimated by bicinchoninic acid assay (BCA assay) as described below. In-solution trypsin digest was conducted for 14-16 h at 37°C using Trypsin Gold ® , Mass Spectrometry Grade (Promega) at 1:50 of µg enzyme to µg protein. Samples were acidified to pH<2 with trifluoroacetic acid to precipitate detergent and desalted using Pierce C18 Spin Columns (Thermo Fisher Scientific). To maximize protein retention, the flow-through after the addition of the sample to the column was passed back over the column twice. Proteins were eluted with 70% acetonitrile, diluted 1:1 with HPLC-grade water, lyophilized, and resuspended in 0.1% formic acid in LC/MS-grade water. Peptide concentration was estimated by BCA assay as described below.
BCA assay
Protein and peptide concentration was estimated using Pierce BCA Protein Assay Kit (Thermo Fisher Scientific). Samples were diluted 1:10 in 50 mM ammonium bicarbonate and each sample was used in triplicate in the BCA Assay. The mean coefficient of variation (CV) for triplicates was 1.57%. Prism7 (GraphPad, USA) was used to fit the curve (third-order polynomial fit, r 2 =0.99) and extrapolate unknown sample concentrations.
RNA quantity and quality assessment RNA yields were determined using RNA BR Assay on the Qubit 3.0 fluorometer (Life Technologies). RNA quality was evaluated by NanoDrop spectrophotometry and microcapillary gel electrophoresis (RNA 6000 Pico assay, Bioanalyzer 2100 instrument, Agilent, USA).
HPLC-MS/MS
Peptide samples were diluted to 150 ng/µl in 0.1% formic acid in LC/MSgrade water and 5 μl were loop injected by a Dionex Ultimate 3000 autosampler onto a reversed-phase trap column (Acclaim ® PepMap ® 100 C18 LC column; 75 µm i.d.×2 cm, 3 µm particle size, 100 A pore size, Thermo Fisher Scientific), and eluted onto a reversed-phase analytical column (EASY-Spray ® C18 LC column; 75 µm i.d.×15 cm, 100 A, Thermo Fisher Scientific) held at 35°C. Solvents A and B were 0.1% formic acid in water and in acetonitrile, respectively. Solvent B was used at the following concentrations: 2% for 5 min, 2-22% over 70 min, 22-38% over 25 min, 38-95% over 5 min, 95% for 5 min, return to 2% over 5 min, 2% for 25 min. Flow rates were held at 300 nl/min with each sequencing run set to 140 min. Mass spectrometry analysis was performed using Orbitrap Fusion ® Tribrid ® mass spectrometer equipped with EASY-Spray ® ion source (Thermo Fisher Scientific) operated in a data dependent acquisition (DDA) mode by Xcalibur 4.0 software (Thermo Fisher Scientific). Briefly, full MS1 scans were resolved by the orbitrap, and ions were selected for MS2 using DDA (charge state: 2-7; intensity threshold: 25,000). These precursor ions were quadrupole filtered and subsequently fragmented using stepped collision HCD at 28%±3 collision energy. MS2 product ions were resolved by the orbitrap. Instrument and data acquisition settings are presented in Table S1.
MS/MS data analysis
MS/MS data was analyzed using Proteome Discoverer v2.2 (Thermo Fisher Scientific) and PEAKS Studio v8.5 (Bioinformatics Solutions Inc., USA). Peptide spectra were searched against the entire UniProt SwissProt database (downloaded on 2/13/2018) concatenated with a common contaminant database (common Repository of Adventitious Proteins, cRAP, https://www. thegpm.org/crap/index.html) using (1) SEQUEST in Proteome Discoverer or (2) PEAKS DB after de novo sequencing. Search parameters are shown in Table S2. False discovery rate (FDR) was estimated by searching a reversed concatenated database in Proteome Discoverer and by the 'decoy-fusion' approach in PEAKS (Zhang et al., 2012). Results were filtered to retain peptides and proteins with a false discovery rate (FDR) <1% and to remove those with hits to the contaminant database. Proteins were considered 'unique' if they had two or more unique peptides that mapped to them. DAVID Bioinformatics Resources v6.8 (Huang da et al., 2009) server was used to identify KEGG and GO categories that were overrepresented in the seal blubber proteome relative to the entire human genome (P<0.05, adjusted for multiple comparisons using Benjamini correction).
Statistical analyses
All statistical analyses were conducted using JMP 13 (SAS Institute Inc., USA). Protein yields were compared by paired t-test (two-tailed), assuming unequal variances. The numbers of PSMs, peptides, proteins, and protein groups were compared using linear mixed models with method (SDC or QIA) and software (SEQUEST or PEAKS) as fixed effects and sample ID as a random effect. | 5,527.2 | 2018-10-25T00:00:00.000 | [
"Biology",
"Environmental Science"
] |
A Probabilistic Annotation Model for Crowdsourcing Coreference
The availability of large scale annotated corpora for coreference is essential to the development of the field. However, creating resources at the required scale via expert annotation would be too expensive. Crowdsourcing has been proposed as an alternative; but this approach has not been widely used for coreference. This paper addresses one crucial hurdle on the way to make this possible, by introducing a new model of annotation for aggregating crowdsourced anaphoric annotations. The model is evaluated along three dimensions: the accuracy of the inferred mention pairs, the quality of the post-hoc constructed silver chains, and the viability of using the silver chains as an alternative to the expert-annotated chains in training a state of the art coreference system. The results suggest that our model can extract from crowdsourced annotations coreference chains of comparable quality to those obtained with expert annotation.
Introduction
The task of identifying and resolving anaphoric reference to discourse entities, known in NLP as coreference resolution, has long been considered a core aspect of language interpretation (Poesio et al., 2016b), also because of its role in applications such as summarization (Baldwin and Morton, 1998;Steinberger et al., 2007), information extraction (Humphreys et al.) or question answering (Morton, 1999;Zheng, 2002).
In the 1990s the field made a paradigmatic turn towards corpus based approaches initiated by campaigns such as MUC (Grishman and Sundheim, 1995;Chinchor, 1998) and since then we have seen the development of a range of data-driven approaches, spurred by the development of ever larger and richer datasets. Nowadays, a variety of datasets exist for several languages (Poesio et al., 2016a). These include medium-scale multilingual datasets such as ONTONOTES (Pradhan et al., 2007;Weischedel et al., 2011), which led to the most recent evaluation campaigns, in particular CONLL 2012(Pradhan et al., 2012, and are used in most current research (Björkelund and Kuhn, 2014;Martschat and Strube, 2015;Clark and Manning, 2016;Lee et al., 2017). However, there are still many languages and domains for which no such resources are available, and even for English much larger corpora than ONTONOTES will eventually be required.
However, annotating data on the scale required to train state of the art systems using traditional expert annotation would be unaffordable. One alternative is to employ crowdsourcing, either via platforms like Amazon Mechanical Turk and Crowdflower, or using Games-With-A-Purpose (Poesio et al., 2017). Studies such as (Snow et al., 2008;Raykar et al., 2010) have shown that when a sufficiently large number of workers is employed, expert-level quality can be achieved, at a fraction of the cost required to create such resources using traditional methods. The one effort to create a large-scale coreference corpus entirely through crowdsourcing, the Phrase Detectives project (Poesio et al., 2013;Chamberlain, 2016), employs the Phrase Detectives game with a purpose. The Phrase Detectives corpus consists of 843 documents for a total of 1.2 million tokens and 392,741 markables; at present, 563 documents for a total of 360,000 tokens have been annotated. 1 A second coreference corpus created using crowdsourcing (in the context of a trivia game) also exists, the Quiz Bowl dataset (Guha et al., 2015). 2 However, such existing corpora are not widely used yet. One of the reasons for this is the lack of suitable aggregation methods for anaphora. Crowdsourced annotations require aggregation methods to select among the different interpretations produced by the crowd. Standard practice for crowdsourced data analysis has seen a shift in recent years from simple majority vote to much more effective aggregation methods (Smyth et al., 1994;Quoc Viet Hung et al., 2013;Sheshadri and Lease, 2013;Carpenter, 2008;Hovy et al., 2013;Passonneau and Carpenter, 2014). Probabilistic models of annotation, in particular, make it possible to characterize the accuracy of the annotators and correct for their bias (Dawid and Skene, 1979;Passonneau and Carpenter, 2014), to account for item-level effects (e.g.: difficulty) (Whitehill et al., 2009), and to employ different pooling strategies (Carpenter, 2008). However, existing models of annotation cannot be used for anaphora. Such methods assume that coders choose between a fixed set of general labels, the same labels across all annotated items. In anaphoric annotation, by contrast, coders relate markables to coreference chains which depend on the markables that are annotated in that given document (Passonneau, 2004;Artstein and Poesio, 2008) Contributions In this paper we propose a mention pair-based approach to aggregating crowdsourced anaphoric annotations. Concretely, we introduce a new model of annotation capable of inferring the most likely mention pairs from crowdannotated anaphoric relations. We then use these pairs to build the most likely coreference chains. This approach to building chains is evaluated on both crowdsourced and synthetic (via simulation) coreference datasets. The evaluations include assessing the accuracy of the inferred mention pairs; the quality of the chains; and the viability of using these chains derived from mention pairs as an alternative to gold chains when training a state of the art coreference system. We conclude by also demonstrating the quality of the proposed model 2 Another corpus creation project using crowdsourcing (and also games) for anaphoric annotation is the Groningen Meaning Bank (Bos et al., 2017). However, in the GMB crowdsourcing is not used to generate interpretations: players correct automatically annotated interpretations rather than providing the annotations themselves. Another crucial difference is that interpretations are not aggregated in the sense discussed below; rather, an expert adjudicates between the interpretations produced by players. in a standard annotation task. The implementation is available as supplementary material.
A Mention-Pair Model of Annotation
Traditional models of annotation (Dawid and Skene, 1979;Smyth et al., 1994;Raykar et al., 2010;Hovy et al., 2013) are specified assuming the annotations are chosen among a general set of classes that is consistent across the annotated items. This is the case in a type of annotation closely related to anaphoric annotation, information status annotation (Nissim et al., 2004;Riester et al., 2010). In this type of annotation, an annotator marks a mention as either discourse old (DO)referring to an existing entity (coreference chain) -or as discourse-new (DN) -introducing a new coreference chain, but without specifying which coreference chain the mention belongs to, if any. We will refer below to categories such as DN and DO as (general) classes.
Traditional models of annotation can model this type of annotation, but not the task of anaphoric annotation proper. In standard annotation schemes for anaphora/coreference (Poesio et al., 2016a) the annotator may mark a mention as referring to a discourse new entity as above; but in case the mention is identified as discourse-old, this entity, or coreference chain-the set of coreferring mentions-is also specified. The available coreference chains differ from document to document.
Our proposal for a probabilistic model of this type of annotation is based on one of the most widely used models of coreference resolution: the mention pair model. In the mention pair model, the task of linking the mention to a coreference chain/entity is split in two parts: classifying mention pairs as coreferring or not, and subsequent clustering (Soon et al., 2001;Hoste, 2016). The model we propose addresses the first part.
More formally, the crowdsourced data to be modeled consists of I mentions (indexed by i) annotated by a total of J coders (indexed by j). Each mention i has N i annotations (indexed by n), for a total of M i distinct labels (indexed by m). Each label m of mention i belongs to a class z i,m . The label of a mention could be the ID of the antecedent, in case that mention is annotated as belonging to the discourse old (general) class; or could be discourse new or another general class (e.g.: property, non referring). In these latter cases, the labels coincide with the classes they belong to.
An important difficulty we had to address is label sparsity. The solution we propose is to transform the mention-level annotations into a series of binary decisions with respect to each candidate label. In the extended literature this is often referred to as the binary relevance method (Tsoumakas and Katakis, 2007;Madjarov et al., 2012). We then model these (label-level) decisions as the result of the sensitivity (the true positive rate) and specificity (the true negative rate) of the annotators which we assume are class dependent. This latter assumption allows inferring different levels of annotator ability for each class (e.g.: capturing that DO labels are generally harder compared to DN).
The graphical model of our Mention Pair Annotations model (MPA) is presented in Figure 1, while the generative process is given below: • For every class h ∈ {1, 2, ..., K}: -Draw class specific true label likelihood π h ∼ Beta(a, b) • For every annotator j ∈ {1, 2, ..., J}: -For every class h ∈ {1, 2, ..., K}: • For every mention i ∈ {1, 2, ..., I}: -For every candidate label m ∈ {1, 2, ..., M i }: The model addresses the first part of the mention pair framework: the posterior of the true label indicators is used to link each mention with the most likely label, obtaining the mention pairs. The coreference chains are then built by following the link structure from the inferred pairs.
Note that for a traditional annotation task with no distinction between generic classes and specific labels the MPA model is equivalent to training K binary Bayesian versions of the Dawid 3 Notation: jj[i,m,n] returns the index of the annotator who made the n-th decision on the m-th label of mention i. and Skene (1979) model (one for each general class) on data processed using the binary relevance method. Note also that whereas traditional models of annotation assume one true class per annotated item, an implicit benefit of our approach is allowing for potentially multiple true classes, which can be useful to detect ambiguity (Poesio and Artstein, 2005), but we don't exploit that in this work.
Parameter Estimation
We infer the parameters of the proposed model using Variational Inference (VI). Unlike Markov Chain Monte Carlo (MCMC) approaches (e.g.: Gibbs Sampling, Hamiltonian Monte Carlo), VI is deterministic, fast, and benefits from a clear convergence criterion (Blei et al., 2017).
Specifically we approximate the intractable posterior p(θ|D) with a variational distribution q(θ) such that the Kullback-Leibler (KL) divergence between the two distributions is minimized. It can be shown this minimization is equivalent to maximizing the evidence lower bound (ELBO) below: π, α, β, c, y|a, b, d, e, t, u, z)] We need a variational distribution q that is tractable under expectations. Following common practice (Blei et al., 2003;Hoffman et al., 2013;Blei et al., 2017), we choose q to be in the mean field variational family where each hidden variable is independent and governed by its own parameter. Elegant solutions have been derived for models whose complete conditionals are in the exponential family (Blei and Jordan, 2006;Hoffman et al., 2013). Concretely, we used the fact that the natural parameters of the variational distributions are equal to the expected value of the natural parameters of the corresponding complete conditionals. The derivations are standard in the VI literature (see, for example, Hoffman et al., 2013). (To save space, we only provide here the update formulas of the variational parameters; supplementary details are in the Appendix.) Equations (2) and (3) give the variational update formulas for the class-level true label likelihood. We have q(π h |λ h , η h ) = Beta(λ h , η h ), where: In Equation (4) and (5) we list the variational update formulas for the class-level annotator sensitivity. We have In Equations (6) and (7) we list the variational update formulas for the class-level annotator specificity. We have q(β j,h |θ j,h , j,h ) = Beta(θ j,h , j,h ), where: In Equations (8) and (9) we list the variational update formulas for the true label indicator. We Finally, for the above formulas, we used the fact that E q [I(c i,m = 1)] = φ i,m . The other expectations can be easily calculated noting that for a distribution part of the exponential family, the first derivative of the log normalizer is equal to the expected value of the sufficient statistics (Blei et al., 2003).
is the digamma function. Similar observations apply to the α and β related expectations.
Evaluation
We carried out a series of evaluations of increasing complexity of our MPA model. We first assess the accuracy of the inferred mention pairs. Second, we cluster the pairs into appropriate coreference chains and evaluate the quality of these chains. Third, we assess the viability of using silver chains as an alternative to the gold chains when training a state of the art coreference system. Finally, we conclude the evaluation with a performance check in a standard annotation task.
Datasets
The largest coreference dataset with crowdsourced annotations is the Phrase Detectives corpus. A subset of this corpus is the Phrase Detectives 1.0 dataset , which also includes gold annotations and can therefore be used to evaluate the accuracy of MPA at mention-pair and coreference chain inference, but is too small to train a state-of-the-art coreference system. To carry out this second type of evaluation we used the approach, common in the crowdsourcing literature (Carpenter, 2008;Raykar et al., 2010;Hovy et al., 2013;Felt et al., 2014), of generating simulated datasets by corrupting the gold standard of an existing corpus. For this purpose, we use the CONLL-2012dataset (Pradhan et al., 2012, at present the standard dataset for coreference resolution.
Crowdsourced Data
The Phrase Detectives (PD) 1.0 dataset has been annotated using the Phrase Detectives game with a purpose. 4 The annotation scheme for PD is based on that for the ARRAU corpus (Poesio et al., 2018). Players have to label predefined 5 markables with one of the following categories: nonreferring (e.g., for expletives), discourse-new, discourse-old (in which case an antecedent is also marked, the most recent mention belonging to the antecedent's coreference chain), or property (for appositions and copular structures). The PD 1.0 dataset is the portion of the corpus that contains, in addition to the annotations by the players, a gold label for each markable. The coreference chains are obtained using a simple clustering of the mention pairs. An important limitation of this corpus is its small size (around 6000 markables from 45 documents), making it unfit for the training and evaluation of state of the art supervised systems.
Synthetic Data
The CONLL-2012 dataset specifies gold chains, not mention pairs. So we need first to extract appropriate mention pairs from these chains. To do this, for each mention we select as gold label the closest mention from its gold chain (or discourse new if the mention is the first in its chain). The simulations are then generated by extracting from each gold label a number of 'crowdsourced labels' produced by (simulated) annotators with varying degrees of ability. We considered a range of simulated scenarios, all sharing the following settings: • 10 distinct annotators per mention and 20 distinct mentions per annotator. The annotators receive random mentions to annotate. 6 • Each annotator is assigned randomly a profile. The profiles indicate the sensitivity of the annotators with respect to discourse old and new. For example, the (DO 0.8, DN 0.9) profile indicates that, given a mention whose true class is DO, the annotator has 0.8 probability of getting it right; and of 0.9 for DN. We considered both profiles reflecting the actual profiles of players in Phrase Detectives (Chamberlain, 2016) and synthetic profiles.
• 5 choices for the annotators to choose from for each mention: the correct label, the DN PD The range of options considered in the simulation is specified by two aspects: the sensitivity from the annotator profiles and the distribution of the errors they make. We use the following two profile types: The profiles roughly correspond to two experts and three novices whose class sensitivities are relatively close -with extra mass associated with DN because this class is generally easier compared to DO.
• Phrase Detectives inspired profiles: from the PD annotators who annotated more than 10 DO and 10 DN mentions (thresholds set to have a minimum confidence) we extracted a total of 89 profiles. This gave us much more interesting sensitivity pairs compared to the ones from the synthetic profiles, i.e., contrasting class abilities -see Figure 2.
We also considered a range of ways in which annotators may make mistakes: • Distribute the errors uniformly random given the remaining mass (1 -sensitivity) • Distribute the errors in a sparse manner, i.e., assume that some errors will be more likely than others. This can be achieved by drawing randomly from a 4-dimensional (4 = number of errors) uniform Dirichlet for each mention. The annotator probabilities over the 5 choices will then consist of their sensitivity, and the error distribution normalized with respect to the remaining mass.
The settings just discussed lead to 4 simulations summarized in Table 1.
Evaluation 1: Mention Pair Accuracy
We use MPA to link each mention with the most likely label based on the posterior of the true label indicators. We then assess the accuracy of the inferred mention pairs against the gold standard, i.e., the agreement with the gold mention pairs. In this task the proposed model is compared against a majority vote baseline where each mention is paired with the most voted label. 7 The evaluation is conducted on the crowdsourced annotated PD 1.0 dataset and on simulated data generated from the CONLL-2012 test set. The results, summarized in Table 2, indicate the mention pairs inferred by our model (MPA) obtain a much better level of agreement with the gold mention pairs, compared with the output of the majority vote (MV) baseline. MV implicitly assumes equal expertise among the annotators, which has repeatedly been shown to be false in annotation practice (Poesio and Artstein, 2005;Passonneau and Carpenter, 2014;Plank et al., 2014).
Evaluation 2: Silver Chain Quality
After the mention pairs have been inferred using MPA, producing the coreference chains -we will henceforth refer to the coreference chains thus ob-7 Throughout the paper we report the best majority vote result after 10 random rounds of splitting ties. tained as silver coreference chains 8 -is a straightforward clustering task: we simply follow the link structure from the pairs. In this Section we assess the quality of the silver chains using standard coreference metrics -in particular, the Extended Scorer introduced in (Poesio et al., 2018) which extends the official CONLL scorer to include in the evaluation system-predicted singletons and non referring expressions, both of which are annotated in Phrase Detectives; when singletons and nonreferring expressions are not considered, the Extended Scorer is identical to the official scorer. As in the previous experiment, the evaluation is conducted on the crowdsourced annotated PD 1.0 dataset and on simulated data generated from the CONLL-2012 test set. We compare silver chains produced using our MPA model, using MV, and using the Stanford deterministic coreference system (Stanford) (Lee et al., 2011). To run the latter on PD 1.0, we used the default annotators of the CoreNLP toolkit (Manning et al., 2014) to supply the information required by the coreference sys-8 Our use of the term 'silver standard' should not be confused with the other common use of standard generated out of automatic annotations. tem and switched off the post-processing to output singleton clusters; for the CONLL-2012 data we set the dcoref.replicate.conll = true to run exactly the same method as Lee et al. (2011). On both datasets we evaluated on gold mentions. Table 3 summarizes the results on the crowdsourced annotated PD 1.0 dataset. The silver chains obtained using our MPA model are of a far better quality than those of baseline alternatives such as MV and Stanford. Note also that even the simple MV baseline built from crowdsourced annotations yields much better chains compared to a standard coreference system such as the Stanford system. This underlines the advantage of crowdsourced annotations for coreference over automatically produced annotations. In Table 4 we present the scores of MPA and MV on cases of non referring. In this case, as well, the probabilistic model substantially outperforms the MV baseline.
In Table 5 we present the results obtained on simulated data from the CONLL-2012 test set. The results follow a similar trend to those observed using actual annotations: a much better quality of the chains produced using the mention pairs inferred by our MPA Table 6: Results of a state of the art coreference system trained on silver chains obtained in different ways. Each simulated scenario is randomly generated 10 times (summary is in terms of average result and standard deviation) scenarios. Furthermore, the MV baseline achieves better chains compared to the Stanford system in 3 out of 4 simulation settings, again showcasing the potential of crowdsourced annotations.
Training on Silver Chains
In this Section we assessed the viability of using the (silver) chains extracted from crowdsourcing as an alternative to gold chains when training a state of the art coreference system. Concretely, we train the best-performing current system Lee et al. (2017) on chains produced using our MPA model, the MV baseline and the Stanford deterministic system (Lee et al., 2011) (used mainly for calibration, i.e., an alternative baseline that's not based on crowdsourced annotations). We also include the results obtained using actual gold chains. The results are in Table 6. Across all simulated scenarios, the silver chains produced by our MPA model obtain the closest performance to training on gold chains, and the best result is only 1 percentage point less than the result with gold chains. Again, the MV chains lead to better performance than those obtained using a system (Stanford).
These results, once again, indicate the utility of crowdsourced annotations for coreference tasks.
Traditional Crowdsourcing Tasks
In this Section we show that MPA is state of the art also on traditional crowdsourcing datasets, where annotations fall into general classes that are consistent across the annotated items. This evaluation was done on the datasets (WSD, RTE and TEMP) introduced by Snow et al. (2008) and widely used as benchmarks in the literature on annotation models (Hovy et al., 2013;Carpenter, 2008).
We compare the results against a majority vote baseline and two well-known state of the art models: a Bayesian version of the Dawid and Skene (1979) (DS) model and MACE (Hovy et al., 2013). We implement DS ourselves using variational inference, while for MACE, we simply report the published results. As in Hovy et al. (2013) the assessment is done in terms of accuracy against the gold standard. The results, presented in Table 7, indicate the proposed model achieves performance on par with the state of the art.
Related Work
To our knowledge, this is the first paper proposing a model of crowdsourced annotations for coreference. We did draw inspiration however from existing mention pair models of coreference and traditional models of annotation.
The so-called mention pair model is one of the early machine learning approaches to coreference resolution, made popular by Soon et al. (2001). The model is based on a two step procedure: a classification step which identifies the coreferent mention pairs, followed by a clustering step which builds the coreference chains from the aforementioned pairs. The diversity of mention pair models comes from the distinct approaches taken for each of the two steps (Hoste, 2016). Although we follow a similar two step procedure, there are also important differences. Our way of identifying the mention pairs is completely unsupervised, and relies entirely on the crowdsourced annotations. Furthermore, we pair every mention with only one label, reducing the second step of clustering mention pairs into appropriate coreference chains to a simple grouping task guided by a unique path which arises from the pairs.
All existing probabilistic models of annotation (Dawid and Skene, 1979;Smyth et al., 1994;Raykar et al., 2010;Hovy et al., 2013;Passonneau and Carpenter, 2014) assume the annotations fall into a general set of classes that is consistent across the annotated items. This is clearly not the case in a coreference resolution task, a limitation we had to address. We first transformed the annotations into a series of (per label) binary decisions, approach often referred to, in the multi-class classification literature, as the binary relevance method (Tsoumakas and Katakis, 2007;Madjarov et al., 2012). The transformation avoids modeling the sparse labels directly. We further exploited the fact that the annotations fall into a general set of classes and assumed the inter-label decisions are the result of the class-dependent ability of the annotators.
Conclusions
Crowdsourced annotations are an increasingly popular alternative to expert annotation. Even so, their viability for coreference annotation had not been explored so far. This paper is a first step to filling this gap. We introduced a mention pair-based approach to aggregating crowd- Table 7: Accuracy on standard crowdsourced data sourced anaphoric annotations and assessed the quality of the inferred pairs, of the post-hoc constructed coreference chains, and the viability of using the inferred chains as an alternative to gold chains when training a state of the art coreference system. Throughout the experiments, the model introduced was superior to baseline alternatives such as majority vote and chains obtained automatically using a coreference system, across both genuinely crowdsourced and simulated coreference datasets. Furthermore, even the annotationbased baseline achieved results consistently better than those obtained by automatic coreference resolvers, strengthening the case for using crowdsourced annotations to create coreference datasets. | 5,889.8 | 2018-01-01T00:00:00.000 | [
"Computer Science"
] |
Inference of biological networks using Bi-directional Random Forest Granger causality
The standard ordinary least squares based Granger causality is one of the widely used methods for detecting causal interactions between time series data. However, recent developments in technology limit the utilization of some existing implementations due to the availability of high dimensional data. In this paper, we are proposing a technique called Bi-directional Random Forest Granger causality. This technique uses the random forest regularization together with the idea of reusing the time series data by reversing the time stamp to extract more causal information. We have demonstrated the effectiveness of our proposed method by applying it to simulated data and then applied it to two real biological datasets, i.e., fMRI and HeLa cell. fMRI data was used to map brain network involved in deductive reasoning while HeLa cell dataset was used to map gene network involved in cancer.
Recently, two viable options were discussed by Furqan and Siyal (2015) and Cheng et al. (2014). Furqan and Siyal (2015) proposed to use Random Forest as a regularization technique for evaluating Granger causality whereas Cheng et al. (2014) proposed an LASSO-based method to reuse the time series data by reversing the time stamp of the time series. This concept of time reversal is also discussed and used by other researchers including Haufe et al. (2012), Hu et al. (2015) and others.
In this paper, we are proposing an improved method based on a combination of Random Forest Granger causality and re-utilization of time series data. We are calling it Bi-directional Random Forest Granger causality. This proposed method has increased precision and efficiency compared to existing LASSO-based method proposed by Cheng et al. (2014). In order to provide the proof of improvements of our method, we applied these methods to simulated data before mapping two different real biological networks i.e., gene and brain network.
Random Forest Granger causality
Random Forest is a decision tree based learning algorithm that was initially proposed by Breiman (2001) as a classification technique. However, later Liaw and Wiener (2002) suggested that Random Forest can also be used as regularization technique. This proposition of Liaw and Wiener (2002) to use Random Forest as a regularization technique was discussed and applied by Furqan and Siyal (2015) for evaluating coefficients of vector autoregressive model. They have performed Rigorous experimentations to prove its effectiveness. Its implementation follows the ray diagram shown in Fig. 1. Cheng et al. (2014) proposed Naïve Forward Backward LASSO Granger causality which can handle the shortage of data by reusing the time series data after reversing the time stamp of data. They called this method Naïve Forward Backward LASSO Granger causality. In explaining their proposed method, they use the assumption that the original time series validates all necessary conditions to perform Granger casualty analysis as studied in Bahadori and Liu (2013) and Eichler (2011) Furqan and Siyal (2015) time series. Once all the conditions are validated, they have proposed to use the pseudo code discussed below that uses LASSO-Based Granger causality analysis algorithm that is available at .
Bi-direction Random Forest Granger causality
Based on the findings of Naïve Forward Backward LASSO Granger Causality and Random Forest Granger causality, we are proposing to use Random Forest Granger causality together with the concept of re-utilization of time series data by reversing the data time stamps in order to maximize advantages in terms of precision, false discovery rate, recall, and F1-score. The pseudo code for evaluating Bi-directional Random Forest Granger causality is as follow:
Experimental details
We have implemented the basic Random Forest method on MATLAB with the help of R package (Breiman 2001). Later, we merged the implemented code with Granger causality analysis (GCCA) toolbox (Seth 2010) for evaluating Granger causality that uses BSMART toolbox (Cui et al. 2008). Whereas, we have used Akaike Information Criterion (AIC) as discussed by Akaike (1974) for VAR model order selection. After the implementation of proposed method, we have compared our method with Cheng et al. (2014), LASSO-based method. Cheng's method, using four measures: precision, false discovery rate, recall, and F1-score. These measures were evaluated against ground truth network shown in Fig. 5 using the following mathematical equations:
Simulated network
In order to remain unbiased in our comparative study, we utilized a simulated network dataset that has been previously used by researchers like Furqan and Siyal (2015), Schelter et al. (2006), and more. The simulated data set simulates five variable scenarios. Its ground truth network is shown in Fig. 2, and its network can be modeled using following mathematical equations: where ɛ 1 (t), ɛ 2 (t), ɛ 3 (t), ɛ 4 (t), and ɛ 5 (t) are independent and identically distributed white
Real fMRI dataset
In this paper, we have utilized StarPlus data set which was collected to study the working of the brain related to human deductive reasoning. This StarPlus dataset was collected by Keller et al. (2001) and can be freely accessed from Mitchell and Wang (2001).
In this dataset, they had studied 13 normal subjects using 40 trials on each subject. Each trial consists of two major egments. In one segment of the trial, the subject was presented with a visual stimulus in the form of Image for 4 s followed by a 4-s blank screen. Then, in next segment, another visual stimulus was presented for another 4-s in the form of a sentence wich may or may not be related to the image. This visual stimulus was followed by 4-s blank screen. After both stimuli, the subject was asked to decide the presence of a relation between image and sentence. Moreover, each subject was allowed to rest for 15-s before the start of next trial.
Precision =
True positive edges True positive edges + False positive edges .
Recall =
True positive edges True positive edges + False negative edges F 1-Score = 2 × True positive edges 2 × True positive edges + False positive edges + False negative edges In order to introduce randomness in the experiment, 40 trials were divided into two parts of 20 trials each. In 20 trials, subjects were shown image first and then the sentence whereas for remaining 20 trials, they reversed the order of image and sentence. Further information related to experiment settings, sentences, and picture, are explicitly not discussed here and can be referred to Keller et al. (2001).
While performing these trials, T2-weighted fMRI images were collected using 3T Signa scanner at an interval of 500 ms, and with TE = 18 ms and flip angle of 50°. These settings yield images that have approximately 5000 voxels per subjects in 8 oblique axial slices in two different non-contiguous four-slice volumes. The first volume set captures prefrontal areas and superior parietal regions, while, another volume set covers posterior temporal, inferior frontal and occipital areas.
After acquiring T2-weighted fMRI images for each subject, images were pre-processed using FIASCO program (Eddy et al. 1999). This pre-processing helps in reducing the artifacts that arise during image acquisition process due to signal drift, head motion, and others.
After pre-processing of images, 25 anatomical regions of interest were selected that includes left dorsolateral prefrontal cortex (LDLPFC) and right dorsolateral prefrontal cortex (RDLPFC), calcarine sulcus (
Real Hela dataset
The HeLa human cancer cell line dataset used in our study was compiled by Whitfield et al. (2002) by performing series of experiments using DNA microarray technique. These experimental results are freely available (Whitfield et al. 2000).
As the observational points are not homogeneously sampled, the data was first interpolated by using cubic smoothing splines (Green and Silverman 1994) as recommended by Hlavácková-Schindler and Bouzari (2013) and Ogutu et al. (2012) before using in our study.
Simulated dataset
Based on the results of simulated studies shown in Fig. 3, we found that LASSO-based Forward Backward Granger causality on average yields approximately 25 % precision, 75 % false discovery rate, 67 % recall and 37 % F1 score. Whereas using the same set of data, our proposed method yields 28 % precision, 70 % false discovery rate, 87 % recall, and 40 % F1 score.
These findings suggest that our proposed method has outperformed the existing method in all measures, with a significant improvement in recall. Our proposed method shows 20 % improvement in recall compared to existing LASSO-based method.
During this study, we have observed that the proposed method is less prone to outliers compared to the LASSO-based method. This ability of insensitivity of outlier is achieved due to inherent advantage of regularized tree methods. We have also observed that the proposed method is highly dependent on selecting the right number of features and number of trees. In this study, we have used the setting of 10 features and 500 trees. However, further studies are required to devise some ideal relationship between both number features and number of trees.
HeLa cell dataset
Following the findings of simulated data set studies, we have applied the proposed method to real HeLa cell dataset. The resultant gene network that is involved in cancers is shown in Fig. 4 where the green arrow shows a uni-directional link between two nodes.
As there is no way to verify the resultant network, we have used Biological General Repository for Interaction Datasets BIOGRID database (Chatr-aryamontri et al. 2014) to look for genes interactions that were already reported. The BIOGRID is a public database that archives and disseminates genetic and protein interaction data from model organisms and humans. Given the above network map, we were able of find 6 out 16 interactions that yield around 37 % precision and 63 % false discovery rate. These statistics are in line with the results of the simulated dataset where BRFGC produces 28 % precision and 63 % false discovery rate.
StarPlus fMRI dataset
For discussing results of real StarPlus dataset shown in Fig. 5, let's first overview the functions of the pre-selected regions studied in this paper. The first region under Fig. 3 Results of five variable simulated datasets consideration is calcarine sulcus (CALC). CALC consist of calcarine cortex that maps the point-to-point representation from the retina to the cortex as discussed by Meadows (2011). The next region under consideration is left intraparietal sulcus (LIPS). This region of the brain is associated with the processing of light contrast elements seen by eyes without analyzing the relationship between those elements (Smith et al. 2014).
Other regions of interest are left opercularis (LOPER) and left triangularis (LTRIA) which are also called Brodmann Area 44 and Brodmann Area 45 (Nishitani et al. 2005), and together they constitute Broca's region. The Broca's region is associated with the processing of words, pseudo-words, and non-words during different parts of reading and their interaction as discussed in Heim et al. (2005).
Left dorsolateral prefrontal cortex (LDLPFC) is associated with manipulation of auditory and spatial information in working memory (Barbey et al. 2013) whereas left inferior parietal lobule (LIPL) is necessary for comparison (Chochon et al. 1999), memory related to motor processes (e.g., movement of hand), mechanical and technical reasoning associated with the use of objects (van Elk 2014) and more. Whereas, the remaining region under consideration is left Temporal Lobe (LT) which is mainly associated with the primary organization of sensory inputs (Read 1981).
Based on the functional knowledge of regions of interests, our resulted network in Fig. 3 shows that the connection between CALC with LIPS seems to transfer visual information (picture or sentence displayed on screen), the bi-direction link between LOPER and LIPS signifies the feed-backed link for recognizing the objects and words. The connection between Brodmann area 44 and 45 shows the movement of information from area 44 to area 45 for further processing of information.
The other links such as the links from Brodmann area 45 represents the transfer of information to and from LDLPFC, LIPL and LT for further processing to evaluate the meaning, relation and deduction of the task performed. The remaining bidirectional link between LIPL ↔ LDLPFC and LT ↔ LDLPFC exchange information related to the movement to finger for registering the answer to the task.
Conclusion
In this paper, we have proposed an improved method called Bi-directional Random Forest Granger causality. It takes the advantage of Random Forest regularization to handle dimensionality issues and at the same time using reversing time stamping property it limits the data shortage problem. Using simulated dataset we have shown the effectiveness of our proposed method and later, we have applied the proposed approach to real StarPlus fMRI data set to study the network involved in human deductive reasoning and to real HeLa cell dataset to map gene network that is involved in cancer. In future, this method can be used in other areas such as econometrics, and social networking. | 2,968 | 2016-04-26T00:00:00.000 | [
"Biology",
"Computer Science"
] |
Adaptive p -value weighting with power optimality
: Weighting the p -values is a well-established strategy that improves the power of multiple testing procedures while dealing with heterogeneous data. However, how to achieve this task in an optimal way is rarely considered in the literature. This paper contributes to fill the gap in the case of group-structured null hypotheses, by introducing a new class of procedures named ADDOW (for Adaptive Data Driven Optimal Weighting) that adapts both to the alternative distribution and to the proportion of true null hypotheses. We prove the asymptotical FDR control and power optimality among all weighted procedures of ADDOW, which shows that it dominates all existing procedures in that framework. Some numerical experiments show that the proposed method preserves its optimal properties in the finite sample setting when the number of tests is moderately large.
Introduction
Recent high-throughput technologies bring to the statistical community new type of data being increasingly large, heterogeneous and complex.Addressing significance in such context is particularly challenging because of the number of questions that could naturally come up.A popular statistical method is to adjust for multiplicity by controlling the False Discovery Rate (FDR), which is defined as the expected proportion of errors among the items declared as significant.
Once the amount of possible false discoveries is controlled, the question of increasing the power, that is the amount of true discoveries, arises naturally.In the literature, it is well-known that the power can be increased by clustering the null hypotheses into homogeneous groups.The latter can be derived in several ways: • sample size: a first example is the well-studied data set of the Adequate Yearly Progress (AYP) study (Rogosa, 2005), which compares the results in mathematics tests between socioeconomically advantaged and disadvantaged students in Californian high school.As studied by Cai and Sun (2009), ignoring the sizes of the schools tends to favor large schools among the detections, simply because large schools have more students and not because the effect is stronger.By grouping the schools in small, medium, and large schools, more rejections are allowed among the small schools, which increases the overall detection capability.This phenomenon also appears in more large-scale studies, as in GWAS (Genome-Wide Association Studies) by grouping hypotheses according to allelic frequencies, (Sun et al., 2006) or in microarrays experiments by grouping the genes according to the DNA copy number status (Roquain and van de Wiel, 2009).Common practice is generally used to build the groups from this type of covariate.• spatial structure: some data sets naturally involve a spatial (or temporal) structure into groups.A typical example is neuroimaging: in Schwartzman, Dougherty and Taylor (2005), a study compares diffusion tensor imaging brain scans on 15443 voxels of 6 normal and 6 dyslexic children.By estimating the densities under the null of the voxels of the front and back halves of the brain, some authors highlight a noteworthy difference which suggests that analysing the data by making two groups of hypotheses seems more appropriate, see Efron (2008) and Cai and Sun (2009).• hierarchical relation: groups can be derived from previous knowledge on hierarchical structure, like pathways for genetic studies, based for example on known ontologies (see e.g.The Gene Ontology Consortium (2000)).Similarly, in clinical trials, the tests are usually grouped in primary and secondary endpoints, see Dmitrienko, Offen and Westfall (2003).
In these examples, while ignoring the group structure can lead to overly conservative procedures, this knowledge can easily be incorporated by using weights.This method can be traced back to Holm (1979) who presented a sequentially rejective Bonferroni procedure that controls the Family-Wise Error Rate (FWER) and added weights to the p-values.Weights can also be added to the type-I error criterion instead of the p-values, as presented in Benjamini and Hochberg (1997) with the so-called weighted FDR.Blanchard and Roquain (2008) generalized the two approaches by weighting the p-values and the criterion, with a finite positive measure to weigh the criterion (see also Ramdas et al. (2017) for recent further generalizations).Genovese, Roeder and Wasserman (2006) introduced the p-value weighted BH procedure (WBH) which has been extensively used imsart-generic ver.2014/10/16 file: Durand2018v2.texdate: March 1, 2022 afterwards with different choices for the weights.Roeder et al. (2006); Roeder and Wasserman (2009) have built the weights upon genomic linkage, to favor regions of the genome with strong linkage.Hu, Zhao and Zhou (2010) calibrated the weights by estimating the proportion of true nulls inside each group (procedure named HZZ here).Zhao and Zhang (2014) went one step further by improving HZZ and BH with weights that maximize the number of rejections at a threshold computed from HZZ and BH.They proposed two procedures Pro1 and Pro2 shown to control the FDR asymptotically and to have a better power than BH and HZZ.
However, the problem of finding optimal weights (in the sense of achieving maximal averaged number of rejected false nulls) has been only scarcely considered in the literature.For FWER control and Gaussian test statistics, Wasserman and Roeder (2006) designed oracle and datadriven optimal weights, while Dobriban et al. (2015) considered a Gaussian prior on the signal.For FDR control, Roquain and van de Wiel (2009) and Habiger (2014) designed oracle optimal weights by using the knowledge of the distribution under the alternative of the hypotheses.Unfortunately, this knowledge is not reachable in practice.This leads to the natural idea of estimating the oracle optimal weights by maximizing the number of rejections.This idea has been followed by Ignatiadis et al. (2016) with a procedure called IHW.While they proved that IHW controls asymptotically the FDR, its power properties have not been considered.In particular, it is unclear whether maximizing the overall number of rejections is appropriate in order to maximize power.Other recent works (Li and Barber, 2016;Ignatiadis and Huber, 2017;Lei and Fithian, 2018) suggest weighting methods (with additional steps or different threshold computing rules) but they don't address the power question theoretically either.
In this paper, we present a general solution to the problem of optimal data-driven weighting of BH procedure in the case of grouped null hypotheses.The new class of procedures is called AD-DOW (for Adaptive Data-Driven Optimal Weighting).It relies on the computation of weights that maximize the number of detections at any rejection threshold, combined with the application of a step-up procedure with those weights.This is similar to IHW, however, by taking a larger weight space thanks to the use of estimators of true null proportion in each group, we allow for larger weights, hence more detections.With mild assumptions, we show that ADDOW asymptotically controls the FDR and has optimal power among all weighted step-up procedures.Interestingly, our study shows that the heterogeneity with respect to the proportion of true nulls should be taken into account in order to attain optimality.This fact seems to have been ignored so far: for instance we show that IHW has optimality properties when the true nulls are evenly distributed across groups but we also show that its performance can quickly deteriorate otherwise with a numerical counterexample.
In Section 2, we present the mathematical model and assumptions.In Section 3, we define what is a weighting step-up procedure and discuss some procedures of the literature.In Section 4, we introduce ADDOW.Section 5 provides our main theoretical results.Our numerical simulations are presented in Section 6, while the overfitting problem is discussed in Section 7 with the introduction of a variant of ADDOW.We conclude in Section 8 with a discussion.The proofs of the two main theorems are given in Section 9 and more technical results are deferred to appendix.Let us underline that an effort has been made to make the proofs as short and concise as possible, while keeping them as clear as possible.
In all the paper, the probabilistic space is denoted (Ω, A, P).The notations a.s.
−→ and P −→ stand for the convergence almost surely and in probability.
Model
We consider the following stylized grouped p-value modeling: let G ≥ 2 be the number of groups.Let us emphasize that G is kept fixed throughout the paper.Because our study will be asymptotic in the number of tests m, for each m we assume that we test m g hypotheses in group g ∈ {1, . . ., G}, where the m g are non-decreasing integer sequences depending on m (the dependence imsart-generic ver.2014/10/16 file: Durand2018v2.texdate: March 1, 2022 is not written for conciseness) and such that G g=1 m g = m.In each group g ∈ {1, . . ., G}, let H g,1 , . . ., H g,mg be some binary variables corresponding to the null hypotheses to be tested in this group, with H g,i = 0 if it is true and H g,i = 1 otherwise.Consider in addition p g,1 , . . ., p g,mg some random variables in [0, 1] where each p g,i corresponds to the p-value testing H g,i .Note also m g,1 = mg i=1 H g,i the number of false nulls and m g,0 = m g − m g,1 the number of true nulls in group g.
We make the following marginal distributional assumptions for p g,i .
Assumption 2.2.If H g,i = 1, p g,i follows a common distribution corresponding to c.d.f.F g , which is strictly concave on [0, 1].
In particular, note that the p-values are assumed to have the same alternative distribution within each group.Note that the concavity assumption is mild (and implies continuity on R as proven in Lemma A.1 for completeness).Furthermore, by concavity, x → Fg(x)−Fg(0) x−0 has a right limit in 0 that we denote by f g (0 x−1 has a left limit in 1 that we denote by Assumption 2.3.There exists π g > 0 and π g,0 > 0 such that for all g, m g /m → π g and m g,0 /m g → π g,0 when m → ∞.Additionally, for each g, π g,1 = 1 − π g,0 > 0. The above assumption means that, asymptotically, no group, and no proportion of signal or sparsity, is vanishing.We denote π 0 = g π g π g,0 the mean of the π g,0 's and denote the particular case where the nulls are evenly distributed in each group by (ED): (ED) Let us finally specify assumptions on the joint distribution of the p-values.
Assumption 2.4.The p-values are weakly dependent within each group: This assumption is mild and classical, see Storey, Taylor and Siegmund (2004).Note that weak dependence is trivially achieved if the p-values are independent, and that no assumption on the p-value dependence accross groups is made.Finally note that there is a hidden dependence in m in the joint distribution of the p-values (p g,i ) 1≤g≤G 1≤i≤mg but that does not impact the remaining of the paper as long as (2.1) and (2.2) are satisfied.
π g,0 estimation
Assumption 2.5.For each g, we have at hand an (over-)estimator πg,0 ∈ (0, 1] of m g,0 /m g such that πg,0 P −→ πg,0 for some πg,0 ≥ π g,0 . Let also π0 = g π g πg,0 .In the model of Section 2.1, this assumption can be fulfilled by using the estimators introduced in Storey, Taylor and Siegmund (2004): for a given parameter λ ∈ (0, 1) let arbitrary (the 1 m is here just to ensure πg,0 (λ) > 0).It is easy to deduce from (2.1) and (2.2) that 1 mg mg i=1 1 {pg,i≤λ} P −→ π g,0 λ + π g,1 F g (λ), which provides our condition: πg,0 (λ) While (π g,0 ) g is let arbitrary in our setting, some particular cases will be of interest in the sequel.First is the Evenly Estimation case (EE) one where πg,0 = π0 , 1 ≤ g ≤ G. (EE) In that case, our estimators all share the same limit, and doing so they do not take in account the heterogeneity with respect to the proportion of true nulls.Case (EE) is relevant when the proportion of true nulls is homogeneous across groups, that is, when (ED) holds.A particular subcase of (EE) is the Non Estimation case (NE) where: Case (NE) is basically the case where no estimation is intended, and the estimators are simply taken equal to 1. Let us also introduce the Consistent Estimation case (CE) for which the estimators πg,0 are assumed to be all consistent: πg,0 = π g,0 , 1 ≤ g ≤ G. (CE) While this corresponds to a favorable situation, this assumption can be met in classical situations, where f g (1 − ) = 0 and λ = λ m tends to 1 slowly enough in definition (2.3), see Lemma A.2 in Section A. The condition f g (1 − ) = 0 is called "purity" in the literature.It has been introduced in Genovese and Wasserman (2004) and then deeply studied, along with the convergence of Storey estimators, in Neuvial (2013).
Finally, the main case of interest is the Multiplicative Estimation case (ME) defined as the following: ∃C ≥ 1, πg,0 = Cπ g,0 , 1 ≤ g ≤ G. (ME) Note that the constant C above cannot depend on g.Interestingly, the (ME) case covers the (CE) case (in this respect, C = 1) and also the case where (ED) and (EE) both hold (in this respect, C = π0 π0 ).So the (ME) case can be viewed as a generalization of previous cases.
Criticality
Depending on the choice of α, multiple testing procedures may make no rejection at all when m tends to ∞.This case is not interesting and we should focus on the other case.To this end, Chi (2007) introduced the notion of criticality: they defined some critical alpha level, denoted α * , for which BH procedure has no asymptotic power if α < α * .Neuvial (2013) generalized this notion for any multiple testing procedure (see Section 2.5 therein) and also established a link between criticality and purity.
In Section A, Definition A.1, we define α * in our heterogeneous setting and will focus in our results on the supercritical case.
Assumption 2.6.The target level α lies in (α * , 1).Lemma A.3 states that α * < 1 so such an α always exists.While the formal definition of α * is reported to the appendix for the sake of clarity, let us emphasize that it depends on the parameters of the model, that are (F g ) g , (π g ) g and (π g,0 ) g , and on the parameters of the chosen estimators, that are (π g,0 ) g .
Leading example
While our framework allows a general choice for F g , a canonical example that we have in mind is the Gaussian one-sided framework where the p-values are derived from Gaussian test statistics.
In this case, Assumption 2.1 is fulfilled, and hence F g is strictly concave and Assumption 2.2 is also fulfilled.Furthermore we easily check that f g (0 + ) = ∞, so α * = 0 and f g (1 − ) = 0 which means that this framework is supercritical (α * = 0, see Definition A.1) with purity and then can achieve consistent estimation (CE) with additional independence assumptions.Two particular subcases of interest arise when Σ has a particular form and can be written as , where Σ (g) is a square matrix of size m g .The first subcase is when Σ (g) is the identity matrix.In this case, the p-values are all independent and Assumption 2.4 is fulfilled by the law of strong numbers.The second subcase is when Σ (g) is a Toeplitz matrix with Σ (g) j,k = 1 |j−k|+1 .In this case, Assumption 2.4 is also fulfilled (see e.g.Delattre and Roquain, 2016, Proposition 2.1, Equation (LLN-dep) and Theorem 3.1).
Criterion
The set of indices corresponding to true nulls is denoted by H 0 , that is (g, i) ∈ H 0 if and only if H g,i = 0, and we also denote H 1 = H 0 c .In this paper, we define a multiple testing procedure R as a set of indices that are rejected: p g,i is rejected if and only if (g, i) ∈ R. The False Discovery Proportion (FDP) of R, denoted by FDP(R), is defined as the number of false discoveries divided by the number of rejections if there are any, and 0 otherwise: We denote FDR(R) = E [FDP(R)] the FDR of R. Its power, denoted Pow(R), is defined as the mean number of true positives divided by m: Note that our power definition is slightly different than the usual one for which the number of true discoveries is divided by m 1 = g m g,1 instead of m.This simplifies our expressions (see Section 9.1) and does not have any repercussion because the two definitions differ only by a multiplicative factor converging to 1 − π 0 ∈ (0, 1) when m → ∞.
Finally, let us emphasize that the power is the (rescaled) number of good rejections, that is, the number of rejected hypotheses that are false.The power is different from the number of total rejections, this distinction is fundamental and will be discussed all along this paper (like, for example, when discussing Heuristic 3.1, or in the simulations of Section 6.4).
Weighting the BH procedure
Say we want to control the FDR at level α.Assume that the p-values are arranged in increasing order p (1) ≤ . . .≤ p (m) with p (0) = 0, the classic BH procedure consists in rejecting all p g,i ≤ α k m where k = max k ≥ 0 : p (k) ≤ α k m .Take a nondecreasing function h defined on [0, 1] such that h(0) = 0 and h(1) ≤ 1, we denote I(h) = sup {u ∈ [0, 1] : h(u) ≥ u} .Some properties of the functional I(•) are gathered in Lemma A.4, in particular h (I(h)) = I(h).We now reformulate BH with the use of I(•), because it is more convenient when dealing with asymptotics.Doing so, we follow the formalism notably used in Roquain and van de Wiel (2009) and Neuvial (2013).Define the empirical function . This is a particular case of Lemma A.5.Note that G(u) is simply the number of p-values that are less than or equal to αu, divided by m.
The graphical representation of the two points of view for BH is depicted in Figure 1 with m = 10.The p-values are plotted on the right part of the figure along with the function k → αk/m and we see that the last p-value under the line is the sixth one.On the left, the function G corresponding to these p-values is displayed alongside the identity function, with the last crossing point being located between the sixth and seventh jumps, thus I( G) = 6/m and 6 p-values are rejected.
The weighted BH (WBH) with weight vector w ∈ R G + is defined by computing and rejecting all p g,i ≤ αI (G w ) w g .We denote it WBH(w).Note that w is authorized to be random, hence it can be computed from the p-values.In particular, BH = WBH(1) where 1 = (1, . . ., 1) ∈ R G + .Following Roquain and van de Wiel (2009), to deal with optimal weighting, we need to further generalize WBH into a multi-weighted BH (MWBH) procedure by introducing a weight function W : [0, 1] → R G + , which can be random, such that the following function: of the class of WBH procedures because for any weight vector w, w can be seen as a constant weight function u → w and G w is nondecreasing.Note that, there is a simple way to compute ûW .For each r between 1 and m denote the W (r/m)-weighted p-values p
[r]
g,i = p g,i /W g (r/m) (with the convention p g,i /0 = ∞), order them p (r) ≤ α r m (this is Lemma A.5).As in previous works (see e.g.Genovese, Roeder andWasserman, 2006 or Zhao andZhang, 2014), in order to achieve a valid FDR control, these procedures should be used with weights that satisfy some specific constraints.The following weight spaces will be used in the following of the paper: Note that K may appear unusual because it depends on the estimators πg,0 , however it is completely known and usable in practice.Some intuition about the choice of K is given in next section.Note also that K = KNE in the (NE) case.Finally, for a weight function W and a rejection threshold u ∈ [0, 1], we denote by R u,W the double indexed procedure rejecting the p-values less than or equal to αuW g is the number of rejections of R u,W , divided by m) and that MWBH(W ) can also be written as R ûW ,W .
Choosing the weights
Take W and u, and let P (m) W (u) = Pow (R u,W ).We have Note that these relations are valid only if W and u are deterministic.In particular, they are not valid when used a posteriori with a data-driven weighting and u = ûW .
In Roquain and van de Wiel (2009), the authors define the oracle optimal weight function W * or as: Note that they defined W * or only in case (NE), but their definition easily extends to the general case as above, by replacing KNE by K.They proved the existence and uniqueness of W * or when both (ED) and (NE) hold and that, asymptotically, MWBH(W * or ) controls the FDR at level π 0 α and has a better power than every MWBH(w (m) ) for w (m) ∈ KNE some deterministic weight vectors satisfying a convergence criterion.
However, computing W * or requires the knowledge of the F g , not available in practice, so the idea is to estimate W * or with a data driven weight function W * and then apply MWBH with this random weight function.For this, consider the functional defined by, for any (deterministic) weight function W and u ∈ [0, 1]: where W (u) is the mean ratio of rejections for the procedure rejecting each p g,i ≤ αuW g (u).P W (u) is the rescaled mean of the number of true positives (i.e. the power) of this procedure while H and let ûM = max(û ŵ(1) , û ŵ(2) ).In the second stage, they maximize G w (û M ) over K, which gives rise to the weight vector W * (û M ) according to our notation.Then they define their procedures as the following: and Pro 2 = WBH W * (û M ) .
Pro 2 comes from an additional step-up step compared to Pro 1, hence its rejection threshold, û W * (û M ) , is larger than ûM and allows for more detections.The caveat of this approach is that the initial thresholding, that is the definition of ûM , seems somewhat arbitrary, which will result in sub-optimal procedures, see Corollary 5.3.As a side remark, ŵ(1) and ŵ(2) are involved in other procedures of the literature.The HZZ procedure of Hu, Zhao and Zhou ( 2010) is WBH( ŵ(2) ), and WBH( ŵ(1) ) is the classical Adaptive BH procedure (see e.g.Lemma 2 of Storey, Taylor and Siegmund ( 2004)) denoted here as ABH.Ignatiadis et al. (2016) actually used Heuristic 3.1 with multi-weighting (while their formulation differs from ours) which consists in maximizing G w (u) in w for each u.However, their choice of the weight space is only suitable for the case (NE) and can make Heuristic 3.1 break down, because in general H (m) W (u) can still depend on w, see remark 3.1 below.In the next section, we take the best of the two approaches to attain power optimality with data-driven weighting.Let us already mention that the crucial point is Lemma B.3, that fully justifies Heuristic 3.1, but only in case (ME).When (ME) does not hold, we must take care that Heuristic 3.1 can fail for the same reason that it can fail with IHW.Thereby, in general, more detections do not necessarily imply more power.
Remark 3.1.In particular, we can compute numerical counterexamples where BH has larger asymptotic power than IHW.For example, if we break (ED) by taking a small π 1,0 (almost pure signal) and a large π 2,0 (sparse signal), along with a small group and a large one (π 1 much smaller than π 2 ) and strong signal in both groups, we can achieve a larger power with BH than with IHW.Our interpretation is that, in that case, IHW slightly favors group 2 because of its size, whereas the oracle optimal favors group 1 thanks to the knowledge of the true parameters.BH, by weighting uniformly, does not favor any group, which allows its power to end up between the power of the oracle and the power of IHW.This example is studied in Section 6.4 and illustrated in Figures 8 and 7.
Recent weighting methods
Besides IHW, there are several recent methods putting weights on p-values.We briefly discuss three of them.The first is a variation of IHW by the same authors, IHWc (Ignatiadis and Huber, 2017), where the letter 'c' stands for 'censoring'.The method bring two innovations to IHW.First, the use of cross-weighting thanks to a subdivision of the hypotheses into folds: the weights of the p-values of a fold are computed by only using the p-values of the other folds.This approach reduces overfitting since, during the step-up procedure, the information brought by a given p-value is used only once instead of twice.The second innovation is the censoring, where a threshold τ is fixed and only p-values larger than τ are used to compute the weights, while only p-values lesser than τ can be rejected during the step-up.Together, these innovations allow IHWc to control the FDR in finite sample at level α if the p-values associated to true nulls are independent.However, using only large p-values to compute the weights seems somehow counterintuitive: large p-values are likely to be associated to true nulls and to be uniform, so they won't allow the weights to properly discriminate the groups and to increase the power compared to BH.We will verify this intuition in Section 6.3.Finally, it is worth noting that IHWc allows for a kind of π g,0 estimation à la Storey, with a variant called IHWc-Storey.
The censoring idea originates from the Structure Adaptive BH Algorithm (SABHA, Li and Barber, 2016), which has a group structured version with an FDR bounded by αC for a known imsart-generic ver.2014/10/16 file: Durand2018v2.texdate: March 1, 2022 constant C > 1 when the p-values are independent.Hence, applying the group structured SABHA at level α/C gives FDR control at level α, but using a target level < α can induce conservatism, especially since computing the weights only with the large p-values involve the same risks that we highlighted when discussing of IHWc.
Lastly, AdaPT (Lei and Fithian, 2018) introduces threshold surfaces s t (x) that can be considered as weights and adapted to group setting.AdaPT is not a WBH procedure, its whole philosophy is totally different and relies on symmetry properties of the true null distribution of the p-values by using an estimator of the FDP, different than the one implicitly used in BH-like methods, which also relies on symmetry and allow to mask p-values during the procedure (see also Barber et al., 2015 andArias-Castro et al., 2017 for more details on this pioneering paradigm).We won't further consider AdaPT because of its fundamental differences with WBH procedures and because we are mainly interested by optimality among said WBH procedures.
New procedure: ADDOW
We exploit Heuristic 3.1 and propose to estimate the oracle optimal weights W * or by maximizing in w ∈ K the empirical counterpart to G (m) w (u), that is G w (u).Definition 4.1.We call an adaptive data-driven optimal weight function a random function Such maximum is guaranteed to exist because G w (u), w ∈ K is a finite set.Indeed, it is a subset of k m , k ∈ 0, m .However, for a given u, W * (u) may not be uniquely defined, hence there is no unique optimal weight function W * in general.So, in all the following, we fix a certain W * , and our results do not depend on the choice of W * .An important fact is that G W * is nondecreasing (see Lemma A.6) so û W * exists and the corresponding MWBH procedure is well-defined: Definition 4.2.The ADDOW procedure is the MWBH procedure using W * as the weight function, that is, ADDOW = MWBH W * .
One shall note that ADDOW is in fact a class of procedures depending on the estimators πg,0 through K. Its rationale is similar to IHW in that we intend to maximize the number of rejections, but incorporating the estimators πg,0 allows for larger weights and more detections.Finally, note that, in the (NE) case, ADDOW reduces to IHW.Remark 4.1.It turns out that ADDOW is equal to a certain WBH procedure.It comes from part 2 of the proof of Theorem 5.2 and Remark 9.2.Moreover, to every MWBH procedure, corresponds a WBH procedure with power higher or equal.This fact does not limit the interest of the MWBH class, because computing the dominating WBH procedure of a given MWBH( W ) procedure requires the knowledge of the step-up threshold û W which is known by actually computing MWBH( W ).
Main results
Now we present the two main theorems of this paper.The two are asymptotical and justify the use of ADDOW when m is large.The first is the control of the FDR at level at most α.The second shows that ADDOW has maximum power over all MWBH procedures in the (ME) case.The two are proven in Section 9.
Relation to IHW
Recall that IHW reduces ADDOW in the (NE) case, that (NE) is a subcase of (EE), and that when both (EE) and (ED) hold then (ME) is achieved.Hence, as a byproduct, we deduce from Theorems 5.1 and 5.2 the following result on IHW.2016) (with slightly stronger assumption on the smoothness of the F g s), the FDR controlling result of Corollary 5.1 gives a slightly sharper bound (π 0 α instead of α) in (ED) case.
The power optimality stated in Corollary 5.1 is new and was not shown in Ignatiadis et al. (2016).It thus supports the fact that IHW should be used under the assumption (ED) and when π 0 is close to 1 or not estimated.
Comparison to other existing procedures
For any estimators πg,0 ∈ [0, 1], any weighting satisfying g mg m w g ≤ 1 also satisfies g mg m πg,0 w g ≤ 1, that is KNE ⊂ K. Hence, any MWBH procedure estimating The next corollary simply states that ADDOW outperforms many procedures of the "weighting with π 0 adaptation" literature.
The results for Pro2, HZZ and ABH follow directly from Theorem 5.2 because these are MWBH procedures.The proof for Pro1 (which is not of the MWBH type) can be found in Section D.
Numerical experiments
6.1.Simulation setting FDR analysis and power analysis from Sections 6.2 and 6.3 are conducted using simulations which setting we describe here.Section 6.4 presents a counter-example using its own setting.
Our experiments have been performed by using the four following scenarios Each simulation of each scenario is replicated 1000 times.
• Scenario 4: µ 1 = μ and µ 2 = 0.01 and the dependence follows the Toeplitz pattern described in the end of Section 2.4.
In each scenario, three groups of procedures are compared.The difference between the three groups lies in the way π 0 is estimated.Group 1 corresponds to the (NE) case: πg,0 = 1.Group 2 corresponds to the (CE) case, with an oracle estimator: πg,0 = π g,0 .Groups 3 use the Storey estimator πg,0 (1/2) defined in Equation (2.3).We choose λ = 1/2 as it is a standard value (see e.g.Storey, 2002).The compared procedures are the following: • ABH as defined in section 3.2 (which is BH in Group 1), • HZZ as defined in section 3.2 (except in Group 1 where it is not defined), • Pro2 as defined in section 3.2 (for Group 1, we only use the BH threshold), • ADDOW (which is equal to IHW in Group 1), • An oracle ADDOW wich is the MWBH procedure using the oracle weights W * or given by equation 3.5 (only in Groups 1 and 2), • IHWc (only in Groups 1 and 3).The version of IHWc used in Group 3 is IHWc-Storey.
For IHWc, the censoring level chosen is the default of the IHW R package, that is α.
In the following, only plots of scenarios 1 and 3 are shown, as the situation with Toeplitz dependence is found to be similar to the independent case, up to a slight increase of the FDR of most of the procedures.
FDR control
The FDR of all above procedures are compared in Figure 2 and Figure 3. 1.5 1.75 2 2.25 2.5 2.75 3 q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q FDR plot 1.5 1.75 2 2.25 2.5 2.75 3 q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q FDR plot In scenario 1, we can distinguish two different regimes depending on the signal strength.For µ ≥ 1 the signal strength is not weak in both groups (from µ 1 = μ and µ 2 = 2μ) and the FDR is controlled at level α for all procedures of Groups 2 & 3 except ADDOW and Pro2, the two procedures using the data-driven weights, that is W * .In particular, Oracle ADDOW in Group 2 controls the FDR at level α.As the data driven weights converge to the oracle weights (see Lemma C.4), we get an illustration of Theorem 5.1 in the (CE) case.The situation is similar for Group 1 and level π 0 α, except for Oracle ADDOW which controls the FDR only for µ ≥ 2.
The situation get more confused when the signal is weak (µ < 1).The FDR of ADDOW (in each group) is largely inflated.The FDR control at level α also fails sometimes for Oracle ADDOW, Pro2, ABH and HZZ (only in Group 2).
In scenario 3, one group has always weak signal.The FDR inflation of ADDOW (in each group) and Group 2 is worse for small μ, whereas, for large μ, the situation is similar to scenario 1, up to one exception: the FDR of ABH and IHWc in Group 3 does not reach α as it did in scenario 1, which suggests some sort of conservatism.
In both scenarios, procedures of Group 2 have a larger FDR than their equivalent in Group 3, which in turn have larger FDR than in Group 1.
As a side note, in both scenarios, and both Groups 1 and 3, the FDR plots of IHWc and ABH are nearly indistinguishable.
In both settings regarding μ (large or small), procedures based on W * suffer from some sort of overfitting causing a loss of FDR control.This is discussed in Section 7 with an attempt to stabilize the weights.Let us underline that this does not contradict Theorem 5.1 because a small µ g might imply a smaller convergence rate while m stays < 10 4 in our setting.
Power analysis
Now that the FDR control has been studied, let us compare the procedures in terms of power.First, to better emphasize the benefit of adaptation, the power is rescaled in the following way: we define the normalized difference of power with respect to BH, or DiffPow, by for any procedure R.
Figures 4 and 5 display the power of all the procedures defined in Section 6.1.Figures 6a and 6b display only a subset of them in Scenario 1, for clarity.We can make several observations: • In both scenarios, procedures of Group 2 are more powerful than their equivalent in Group 3, which are better than in Group 1 (up to one exception, see next point), see e.g. Figure 6a.
In particular, the difference between Group 2 and Group 1 is huge.This illustrates the importance of incorporating the knowledge of π 0 to improve power.• In scenario 2, HZZ is largely better in Group 3 than in Group 2. Our interpretation is that the signal is so weak in the second group of p-values that the estimator π2,0 (1/2) is close to one, while π1,0 (1/2) stays close to π 1,0 .Hence ŵ(2) 1 in Group 3 is larger than ŵ(2) 1 in Group 2 which allows for more good discoveries.The drawback of having ŵ(2) 2 in Group 2 is not a real one since the signal is so small that it is impossible to detect no matter the weight.Recall that ŵ(2) is defined in Section 3.2.
• In every Group (that is for any choice of πg,0 ), and for both scenarios, ADDOW achieves the best power (see e.g. Figure 6b), which supports Theorem 5.2.Additionnaly, maybe surprisingly, Pro2 behaves quite well, with a power close to the one of ADDOW (sometimes larger than Oracle ADDOW) and despite its theoretical sub-optimality.• Inside Group 2 or Group 3, and for both scenarios, comparing ABH and HZZ to ADDOW and Pro2 shows the benefit of adding the F g adaptation to the π 0 adaptation: the ADDOW and Pro2 have better power than ABH and HZZ for all signals (see e.g. Figure 6b).In scenario 1, for Groups 2 and 3, we can see a zone of moderate signal (around μ = 1.5)where q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q Difference of power w.r.q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q Difference of power w.r.q q q q q q q q q q q q Difference of power w.r. the two categories of procedures are close.That is the same zone where HZZ becomes better than ABH.We deduce that in that zone the optimal weighting is the same as the uniform ŵ(1) weighting of ABH.• The comparison of the DiffPow between, on the one hand, IHW and, on the other hand, ABH or HZZ from Group 2, in Figure 4, shows the difference between adapting only to the F g 's versus adapting only to π 0 .No method is generally better than the other: as we see in the plot, it depends on the signal strength.We also see that neither ABH nor HZZ is better than the other.• In scenario 1, for all signals, methods of Group 3 are close to their equivalent of Group 2, which indicates that using λ = 1/2 gives a good estimate of π g,0 in practice (see e.g. Figure 6a).Furthermore, the larger the signal is, the more methods of Group 3 get closer to Group 2. • In both scenarios, once again IHWc and ABH are nearly indistinguishable, which confirms the intuition given in Section 3.3 that IHWc performs badly in terms of power due to using only large p-values to compute the weights.See in particular how the power of IHW is larger than the power of IHWc (and even than the power of IHWc-Storey) in Figure 4.
Importance of (ME) for optimality results
We provide here a setting and a simulation where Corollary 5.1 fails because (ED) does not hold, to illustrate the importance of (ME) in Theorem 5.2 and in Theorem 5.1 (to get (5.2)).The setting is chosen according to what we sketched in Remark 3.1 and is the following.
We consider again the one-sided Gaussian framework described in Section 2.4 for G = 2 groups and independent p-values.The parameters are the same as in Section 6.1 and each simulation of each scenario is replicated 1000 times.We choose a large value for α (α = 0.7) which is unlikely to appear in practice but allows us to get our counterexample.We set m 1 = 1000 and m 2 = 9000, m 1,0 /m 1 = 0.05 and m 2,0 /m 2 = 0.85.So group 1 is small and has a lot of signal, while group 2 is large but has not much signal.The signal strength is given by µ 1 = 2 and µ 2 = μ, and μ ∈ {1.7, 1.8, 1.9, 2, 2.1, 2.2, 2.3}, so the signal is strong and almost equal in both groups.
FDR plot
Fig 7 : FDR of ADDOW and BH against μ in the simulation of Section 6.4.The two solid lines are the α and π 0 α levels, the FDR of BH is confounded with the π 0 α level.ADDOW in the (NE) case is given by the black triangles and ADDOW in the (CE) case is given by the red triangles.
a DiffPow plot in Figure 8.In Figure 7, the FDR of BH is π 0 α as expected, and we see that the FDR of IHW is above that level, hence Equation (5.3) is violated.On a side note, we see that, thanks to a large m (10 4 ) and a rather strong signal, ADDOW in (CE) does not overfit and we get an illustration of Equation (5.2) with C = 1.
Figure 8 is rather unequivocal and shows that our parameter choice implies that IHW has a power smaller than BH (ADDOW in (CE) case stays better as expected), hence Equation (5.4) is violated.Let us recall our interpretation proposed in Remark 3.1: IHW favors the large and sparse second group of hypotheses whereas the optimal power is achieved by favoring the small first group of hypotheses which contains almost only signal.As a WBH procedure with weights (1,1), BH does not favor any group.Figure 8 demonstrates the limitation of Heuristic 3.1 by providing a direct counterexample, and underlines the necessity of estimating the π g,0 when nothing lets us think that (ED) may be met.
Overfitting phenomena
Since ADDOW uses the data both through the p-values and the weights, it suffers from an overfitting phenomena where the FDR in finite samples is above the target level α, as we saw in Section 6.2.In our setting, if the signal is strong enough, this drawback is proved to vanish when m is large enough, see the simulations and Theorem 5.1.However, the latter is not true for weak signal: if the data are close to be random noise, making the weight optimization leads ADDOW to assign its weighting budget at random, and giving large weights to the wrong groups increases the FDP.As said before, our intuition is that the overfitting is at least partly due to using each p-value twice in the step-up procedure of ADDOW: in the expression 1 {pg,i≤αu W * g (u)} , p g,i appears in both sides of the inequality because it is used to compute W * g (u).Following this, we propose a variation of ADDOW that uses the same cross-weighting trick as IHWc.
The crADDOW variant
The main idea is to split the p-values into F folds, where F is some fixed integer ≥ 2, and to use only p-values of the remaining F − 1 folds to compute the weights assigned to the p-values of a given fold.The resulting procedure can be seen as a WBH procedure using F × G groups.
Formally, for each m we have a random function , which simply means that the p-values of each group g are evenly distributed between the F folds.Some dependence assumptions are required: Assumption 7.1.The σ-algebra generated by (F m ) m and the σ-algebra generated by (p g,i ) (g,i) m are independent.
Assumption 7.2.Conditionally to (F m ) m , we have weak dependence (as in Assumption 2.4) inside each fold.
For each fold f ∈ {1, . . ., F }, we compute ADDOW −f , that is ADDOW but using only p-values for the folds in {1, . . ., F } \ {f }.This is done by constructing the empirical function and then maximizing it in w ∈ K−f for each u ∈ [0, 1], where: While this expression seems complicated, note that if F divides each m g , then |{1 ≤ i ≤ m g : The maximization provides a weight function W * −f and the MWBH procedure provides a step-up threshold û Our ADDOW variant, named crADDOW for cross-ADDOW, is the WBH procedure which assigns the weight w * g,f to all p-values p g,i such that F m (g, i) = f .Now, in 1 {pg,i≤αuw * g,f } , p g,i is only used once.While we don't have a finite-sample result about crADDOW, we expect it to have a lesser FDR than ADDOW, especially for weak signal.We expect crADDOW to act like a stabilization of ADDOW and to not lose the good performances of ADDOW when the signal is not weak.Those intuitions are verified in the simulations of Section 7.3.Still, crADDOW has the nice property of being asymptotically equivalent to ADDOW.
Theorem 7.1.Let us assume that Assumptions 2.1 to 2.6, 7.1 and 7.2 are fulfilled.Assume also that α ≤ π0 .We have This Theorem is proved in Section E.
Simulations with crADDOW
The simulations presented here are the same as the simulations depicted in Section 6.1, with the addition of crADDOW in each Group.
From the FDR plots, we see that the FDR is hugely deflated and is now controlled at level α for weak μ in each scenario, while for large μ we are still slightly above the target level but with a small improvement over ADDOW.In scenario 1 there is a small window between large and small μ, around μ = 0.75, where crADDOW in Group 2 overfits more than for really large μ, but even there we see a large improvement over ADDOW.
As for the power, we see that crADDOW is less powerful than ADDOW, as expected since we reject less hypotheses, but we see that in most Groups and scenarios the loss of power is almost negligible and crADDOW remains even as powerful as Oracle ADDOW (with the exception of Group 1 in scenario 1).The difference of power between crADDOW and Pro2 is even smaller and crADDOW is better in most configurations, with the exception of Groups 2 and 3 around μ = 1.5, which is the zone that we identified in Section 6.3 as the zone where the optimal weights are given by the uniform ŵ(1) weighting of ABH.
The simulations hence confirm our intuitions about the stabilization properties of crADDOW especially for weak signal where ADDOW was totally unreliable.Studying the finite sample properties of crADDOW, especially its FDR, is an interesting direction for future works.
Concluding remarks
In this paper we presented a new class of data-driven step-up procedures, ADDOW, that generalizes IHW by incorporating π g,0 estimators in each group.We showed that while this procedure q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q FDR plot Fig 9 : FDR against μ in scenario 1. Same legend as in Figure 9, with the addition of crADDOW (yellow lines).
The color of the points (black, red, green) indicates the Group (respectively, 1, 2 and 3).1.5 1.75 2 2.25 2.5 2.75 3 q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q FDR plot Fig 10 : FDR against μ in scenario 3. Same legend as in Figure 9. q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q Difference of power w.r.t.BH 1.5 1.75 2 2.25 2.5 2.75 3 q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q Difference of power w.r.t.BH asymptotically controls the FDR at the targeted level, it has the best power among all MWBH procedures when the π 0 estimation can be made consistently.In particular it dominates all the existing procedures of the weighting literature and solves the p-values weighting issue in a groupstructured multiple testing problem.As a by-product, our work established the optimality of IHW in the case of homogeneous π 0 structure.Finally we proposed a stabilization variant designed to deal with the case where only few discoveries can be made (very small signal strength or sparsity).Some numerical simulations illustrated that our properties are also valid in a finite sample framework, provided that the number of tests and the signal strength are large enough.We also introduced crADDOW, a variant of ADDOW that uses cross-weighting to reduce the overfitting while having the exact same asymptotic properties.
Assumptions Our assumptions are rather mild: basically we only added the concavity of the F g to the assumptions of Ignatiadis et al. (2016).Notably we dropped the other regularity assumptions on F g that were made in Roquain and van de Wiel ( 2009) while keeping all the useful properties on W * in the (NE) case.Note that the criticality assumption is often made in the literature, see Ignatiadis et al. (2016) (assumption 5 of the supplementary material), Zhao and Zhang (2014) (assumption A.1), or the assumption of Theorem 4 in Hu, Zhao and Zhou (2010).Finally, the weak dependence assumption is extensively used in our paper.An interesting direction could be to extend our result to some strong dependent cases, for instance by assuming the PRDS (positive regression dependence), as some previous work already studied properties of MWBH procedures under that assumption, see Roquain and Van De Wiel (2008).
Computational aspects
The actual maximization problem of ADDOW is difficult, it involves a mixed integer linear programming that may take a long time to resolve.Some regularization variant may be needed for applications.To this end, we can think to use the least concave majorant (LCM) instead of the empirical c.d.f. in equation (3.1) (as proposed in modification (E1) of IHW in Ignatiadis et al., 2016).As we show in Section 9, ADDOW can be extended to that case (see especially Section 9.1) and our results are still valid for this new regularized version of ADDOW.
Toward nonasymptotic results
Interesting direction for future research can be to investigate the convergence rate in our asymptotic results.One possible direction can be to use the work of Neuvial (2008).However, it would require to compute the Hadamard derivative of the functional involved in our analysis, which might be very challenging.Finally, another interesting future work could be to develop other versions of ADDOW that ensure finite sample FDR control property: this certainly requires to use a different optimization process, which will make the power optimality difficult to maintain.A possible such variation is crADDOW, whose FDR in finite sample has yet to be investigated.
Further generalization
Define, for any u and W , 1 {pg,i≤αuWg(u),Hg,i=0} , and For the sake of generality D g is not the only estimator of D g (defined in equation (B.1)) that we will use to prove our results (for example, we can use the LCM of D g , denoted LCM( D g ), see Section 8).So let us increase slightly the scope of the MWBH class by defining G W (u) = If W is such that G W is nondecreasing, we then define the generalized MWBH as If ( D g ) g is such that we can define, for all u ∈ [0, 1], we define the generalized ADDOW by the latter being well defined because G W * is nondecreasing (by a proof similar to the one of Lemma A.6).Note that for any continuous D g , such as LCM( D g ) or D g itself, the arg max in (9.1) is non empty and GADDOW can then be defined.
What we show below are more general theorems, valid for any GADDOW ( D g ) g .Our proofs combined several technical lemmas deferred to Sections B and C, which are based on the previous work of Roquain and van de Wiel (2009); Hu, Zhao and Zhou (2010); Zhao and Zhang (2014).
Remark 9.1.GADDOW ( D g ) g when D g = LCM( D g ) and πg,0 = 1 is exactly the same as IHW with modification (E1) defined in the supplementary material of Ignatiadis et al. (2016).In our notation, the latter is WBH W * ũ W * , which is the same as GADDOW ( D g ) g because ũ W * = ũ W * (ũ W * ) (same proof as in Remark 9.2).
Proof of Theorem 5.2
First, in any case, For the rest of the proof, we assume we are in case (ME), which implies by Lemma B.3 that W * (u) ∈ arg max w∈K ∞ P ∞ w (u) for all u, and that P ∞ W * is nondecreasing.We also split the proof in two parts.For the first part we assume that for all m, W is a weight vector ŵ ∈ K therefore not depending on u.In the second part we will conclude with a general sequence of weight functions.
Part 1 W = ŵ ∈ K for all m.Let = lim sup Pow (MWBH ( ŵ)).Up to extracting a subsequence, we can assume that = lim E P ŵ(û ŵ) , πg,0 a.s.−→ πg,0 for all g, and that the convergences of Lemma C.1 are almost surely.. Define the event Now consider that Ω occurs and fix a realization of it, the following of this part 1 is deterministic.Let = lim sup P ŵ(û ŵ).The sequences m mg πg,0 are converging and then bounded, hence the sequence ( ŵ) is also bounded.By compacity, once again up to extracting a subsequence, we can assume that = lim P ŵ(û ŵ) and that ŵ converges to a given w cv .By taking m → ∞ in the relation Part 2 Now consider the case where W is a weight function u → W (u). Observe that so by definition of I(•), û W ≤ û W (û W ) , and then As a consequence, Pow MWBH W ≤ Pow WBH W (û W ) .Finally, apply part 1 to the weight vector sequence W (û W ) to conclude.
Remark 9.2.We just showed that for every MWBH procedure, there is a corresponding WBH procedure with better power.In particular, by defining û = u W * the ADDOW threshold, we showed that û ≤ û W * (û) .But G W * ≥ G ŵ and then û ≥ u ŵ for any ŵ.Hence û = û W * (û) and ADDOW is equal to the WBH procedure associated to the weight vector W * (û).
Remark 9.3.We actually proved a stronger result, as we can replace the statement W : πalt g,0 w g ≤ 1 and the πalt g,0 are such that πalt g,0 P −→ πalt g,0 for some πalt g,0 ≥ πg,0 .That is, the weight space W belongs to does not have to be the same weight space where we apply ADDOW, as long as it uses over-estimators of the limits of the over-estimators used in K.
Appendix A: Lemmas and proofs of Section 2 Lemma A.1.For all g, F g is continuous.
Proof.F g is concave so it is continuous over R \ {0, 1}.F g is continuous in 0 because it is càdlàg.F g is continuous in 1 by concavity and monotonicity.
Lemma A.2. Take a real valued sequence (λ m ) with λ m ∈ (0, 1), converging to 1, such that 0 for all g and the p-values inside each group are mutually independent, then ∀g ∈ {1, . . ., G}, πg,0 (λ m ) The two suprema of the last display, when multiplied by √ m, converge in distribution (by Kolmogorov-Smirnov's theorem).So when divided by 1 − λ m they converge to 0 in distribution and then in probability (because Definition A.1.The critical alpha value is , where Proof.We only need to show that for one w ∈ K ∞ , we have g π g w g π g,0 + π g,1 f g (0 + ) > 1.
Let us show that this is true for every w ∈ K ∞ such that g π g πg,0 w g = 1, e.g. the w defined by w g = 1 πg,0 for all g.We use the fact that f g (0 + ) > Fg(1)−Fg(0) 1−0 = 1 by the strict concavity of F g .Then π g,0 + π g,1 f g (0 + ) > 1 and Recall that I(•) is defined as I(h) = sup {u ∈ [0, 1] : h(u) ≥ u} on the function space: which has the natural order F is also normed with the sup norm • .
Lemma A.4.For all h ∈ F, I(h) is a maximum and h (I(h)) = I(h).Moreover, I(•), seen as a map on F, is nondecreasing and continous on each continuous h 0 ∈ F such that either u → h 0 (u)/u is decreasing over (0, 1], or I(h 0 ) = 0.
Proof.I(h) is a maximum because there exists n → 0 such that
So h (I(h)) ≥ I(h). Then h (h (I(h))
) ≥ h (I(h)) thus h (I(h)) ≤ I(h) by the definition of I(h) as a supremum.
If u + > 1 then obviously It is a maximum by continuity over a compact and is such that s γ < 0, because s γ ≥ 0 would contradict the maximality of I(h 0 ).Then, for all u ∈ [u + , 1], and then sup Hence, as soon as We can then write the following: as soon as h − h 0 ≤ 1 2 (h 0 (u − ) − u − ).This implies I(h) > u − .Taking completes the proof.g,i = p g,i /W g (r/m) (with the convention p g,i /0 = ∞), order them p (1) , . . ., p [r] (r) ≤ α r m .Then r/m ≤ ûW by definition of ûW .Second, we know that ûW can be written as κ/m because ûW = G W (û W ), so we want to show that κ ≤ r which is implied by r, p Furthermore,
Appendix B: Asymptotical weighting
Define, for a weight function W : [0, 1] → R G + , possibly random, and W is nondecreasing, we also define It is the asymptotic version of K. We now define oracle optimal weights over K ∞ for G ∞ • (u) and P ∞ • (u), for all u > 0.
, it is a singleton.In this case, its only element w * belongs to [0, 1 αu ] G and satisfies g π g πg,0 w (u) ≤ 1 with equality if and only if αu ≥ π0 .The same statements are true for P ∞ • , except that the upper bound of max w∈K ∞ P ∞ w (u), which is achieved if and only if αu ≥ π0 , is not For the rest of the proof u is greater than 0.
First we show that any w (u) such that αuw * g1 > 1 and αuw * g2 < 1 for some g 1 , g 2 ≤ G. Now then we define w such that wg = w * g for all g ∈ {g 1 , g 2 }, wg1 = 1 αu and So w belongs to K ∞ and satisfies π g D g (αuw * g ) + π g1 + π g2 D g2 (αu wg2 ) because D g is increasing over [0, 1] and then constant equal to 1.This contradicts the definition of w * so is impossible.
Next we distinct three cases.
We showed that the only w * ∈ arg max Furthermore g π g πg,0 w * g = 1 : if not there would exist a w with wg1 > w * g1 (for the same g 1 as in previous sentence) and wg = w * g for all g = g 1 such that w • , by replacing D g by π g,1 F g .
From now on, W * (u) denotes an element of arg max w∈K ∞ G ∞ w (u) (just like we write W * (u) as an element of arg max w∈ K G w (u)), our results will not depend on the chosen element of the argmax.Next Lemma gives some properties on the function W * is nondecreasing by exactly the same argument as in the proof of Lemma A.6.The result can be strengthened thanks to Lemma B.1, by writing, for u Because the expression above is continuous of the w g , they can always be chosen nonzero.We have α ∧1] with a < b and λ ∈ (0, 1), by Lemma B.1, we have that αaW * g (a), αbW * g (b) ≤ 1 and then, for all g: .We have w ∈ K ∞ and then for all g: , the inequality being strict for g 1 .Finally by summing: 1.The fact that u * = 1 ⇐⇒ α ≥ π0 follows directly from the previous statements and Lemma B.1.The decreasingness of u → G ∞ W * (u)/u is straightforward from strict concavity properties because it is the slope of the line between the origin and the graph of G ∞ W * at abscissa u > 0. Previous statements imply that G ∞ W * is continuous at least over (0, π0 α ∧ 1) and, if α ≥ π0 , over → 0 when u → 0 which gives the continuity in 0. As in the proof of Lemma A.1, the continuity in π0 α ∧ 1 is given by the combination of concavity and nondecreasingness.Remark B.1.The case α ≥ π0 is rarely met in practice because α is chosen small and the signal is assumed to be sparse (so π0 is large) but it is kept to cover all situations.It confirms the intuitive idea that in this situation the best strategy is to reject all hypotheses because then the FDP is equal to π 0 ≤ π0 ≤ α.
Remark B.2.For a weight vector w Figure 13 illustrates all the properties stated in Lemma B.2, with the two cases α ≥ π0 and α < π0 .
The next Lemma justifies the intuitive idea that maximizing the rejections and the power is the same thing (as exposed in Section 3.2), but only under (ME).
u) and arg max w∈K ∞ P ∞ w (u) are both equal to the set of weights w ∈ K ∞ such that αuw g ≥ 1 for all g.Now if αu ≤ π0 , both arg max are singletons.Take w * the only element of arg max w∈K ∞ P ∞ w (u).Recall that there exists C ≥ 1 such that, for all 1 ≤ g ≤ G, πg,0 = Cπ g,0 , and write, for all w ∈ K ∞ , imsart-generic ver.2014/10/16 file: Durand2018v2.texdate: March 1, 2022 because g π g πg,0 w * g = 1 and αuw * g ≤ 1 for all g, by Lemma B.1.This means that w * is also the unique element of arg max w∈K ∞ G ∞ w (u).
In particular, if αu ≤ π0 , Proof.If the statement is false, there exists some > 0 and some sequence (w
Appendix C: Convergence lemmas
Recall that • is the sup norm for the bounded functions on their definition domain: Lemma C.1.The following quantities converge to 0 in probability: From now on D g is assumed to converge uniformly to D g in probability and that W * (u) ∈ arg max w∈ K G w (u) exists for all u.
Next Lemma is the main technical one (with the longest proof).
Lemma C.2.We have the following convergence in probability: Proof.First, where the first term tends to 0 by (C.1), so we work on the second term.
The main idea is to use the maximality of G w (u) in W * (u) and the maximality of G ∞ w (u) in W * (u).The problem is that one is a maximum over K and the other is over K ∞ .The solution consists in defining small variations of W * (u) and W * (u) to place them respectively in K ∞ and K.
W
mean of the number of its false positives.Heuristic 3.1.Maximizing G (m) W (u) should be close to maximizing P (m) W (u). Indeed, consider weight functions W such that g mg,0 m W g (u) = 1 and then replace U (x) by x for all x ∈ R + (whereas U (x) = x only holds for x ≤ 1), then H (m) W (u) becomes αu g mg,0 m W g (u) = αu and it does not depend on the weights.So P (m) W (u) is the only term depending on W in (3(u) is the same.Now, we can evaluate the constraint we just put on W by estimating to the weight space K defined in equation (3.3)), and G (m) w (u) can be easily estimated by the (unbiased) estimator G w (u).As a result, maximizing the latter in w should lead to good weights, not too far from W * or (u).Zhao and Zhang (2014) followed Heuristic 3.1 by applying a two-stage approach to derive two procedures, named Pro1 and Pro2.Precisely, in the first stage they use the weight vectors ŵ(1π0) , where πg,1 = 1 − πg,0 , imsart-generic ver.2014/10/16 file: Durand2018v2.texdate: March 1, 2022 Corollary 5.1.Let us assume that Assumptions 2.1 to 2.6 are fulfilled, with the additional assumption that (ED) holds.Then lim m→∞ FDR (IHW) = π 0 α, (5.3) and for any sequence of random weight functions ( W ) m≥1 such that W : [0, 1] → KNE and G W is nondecreasing, we have lim m→∞ Pow (IHW) ≥ lim sup m→∞ Pow MWBH W . (5.4)While equation (5.1) of Theorem 5.1 covers Theorem 4 of the supplementary material of Ignatiadis et al. ( by 1 uses a weight function valued in K.This immediately yields the following corollary.Corollary 5.2.Let us assume that Assumptions 2.1 to 2.6 are fulfilled, with the additional assumption that (ME) holds.Then lim m→∞ Pow (ADDOW) ≥ lim sup m→∞ Pow (R) , for any R ∈ {BH, IHW}.
Fig 2 :
Fig 2:FDR against μ in scenario 1. Group 1 in black; Group 2 in red; Group 3 in green.The type of procedure depends on the shape: Oracle ADDOW (triangles and solid line); ADDOW (triangles and dashed line); Pro2 (disks); HZZ (diamonds) and finally BH/ABH (crosses).IHWc and IHWc-Storey are in blue, respectively with black and green points.Horizontal lines: α and π 0 α levels.See Section 6.1.
Fig 6 :
Fig 6: Details of Figure 4 where only a subset of procedures is plotted.
Fig 8 :
Fig 8:DiffPow of ADDOW against μ in the simulation of Section 6.4.ADDOW in the (NE) case is given by the black triangles and ADDOW in the (CE) case is given by the red triangles.
strictly concave over [0, u ] and constant equal to M on [u , 1], hence u → G ∞ w (u)/u is decreasing.So whether w = 0 or not, I(•) is continuous in G ∞ w by Lemma A.4. Remark B.3.The proof of the strict concavity of G ∞ W * can easily be adapted to show the (non necessary strict) concavity of G W * when D g = LCM D g .
Finally the properties on P ∞ W * are obtained by the same proof as Lemma B.2.The next lemma is only a deterministic tool used in the proof of Lemma C.4.Define the distance d of a weight vector w to a subset S of R G + by d(w, S) = inf w∈S max g |w g − wg |.Let M u = arg max w∈K ∞ G ∞ w (u) to lighten notations.Lemma B.4.Take some u ∈ (0, 1].Then we have: and D g − D g , for all g ∈ {1, . . ., G}.Furthermore, for any( D g ) g such that D g − D g P −→ 0, sup w∈R G + G w − G ∞ w P −→ 0. (C.1)imsart-generic ver.2014/10/16 file: Durand2018v2.texdate: March 1, 2022Proof.By using the same proof as the one of the Glivenko-Cantelli theorem (which adapts trivially to convergence in probability instead of almost surely), we get from (2.1) and (2.2) that, for all g, ,i≤•,Hg,i=1} − π g,1 F g P −→ 0. So by summing, D g − D g P g + π g D g − D g P −→ 0.
The BH procedure applied to a set of 10 p-values.Right plot: the p-values and the function k → αk/m. | 16,457 | 2017-10-03T00:00:00.000 | [
"Mathematics"
] |
Genetic transformation of the dinoflagellate chloroplast
Coral reefs are some of the most important and ecologically diverse marine environments. At the base of the reef ecosystem are dinoflagellate algae, which live symbiotically within coral cells. Efforts to understand the relationship between alga and coral have been greatly hampered by the lack of an appropriate dinoflagellate genetic transformation technology. By making use of the plasmid-like fragmented chloroplast genome, we have introduced novel genetic material into the dinoflagellate chloroplast genome. We have shown that the introduced genes are expressed and confer the expected phenotypes. Genetically modified cultures have been grown for 1 year with subculturing, maintaining the introduced genes and phenotypes. This indicates that cells continue to divide after transformation and that the transformation is stable. This is the first report of stable chloroplast transformation in dinoflagellate algae.
Introduction
Coral reefs are complex ecosystems, made up of many thousands of species. At the base of the ecosystem are dinoflagellate algae, frequently referred to as zooxanthellae. These single-celled algae live in symbiosis with corals as intracellular photosynthetic symbionts, providing fixed carbon to the host. Loss of the symbiotic alga results in coral bleaching, which is one of the most urgent and worsening worldwide ecological concerns. In 2016, 85% of the Great Barrier Reef was found to be affected by coral bleaching, a significantly higher proportion than had been previously identified (Hughes et al., 2017).
Change in sea water temperature is recognized as one of the environmental causes of coral bleaching (Spalding and Brown, 2015). It is likely that this results in disturbance of photosynthetic electron transfer in the dinoflagellate symbiont and consequent damage (Rehman et al., 2016), (Slavov et al., 2016). The PsbA (D1) reaction center protein of photosystem II is believed to be an important target of such damage (Warner et al., 1999). The key subunits of photosynthetic electron transfer chain complexes, including the PsbA protein, are encoded in the dinoflagellate chloroplast genome (Howe et al., 2008). There have been no reports to date of transformation of the dinoflagellate chloroplast genome, hampering attempts to study the mechanism of bleaching.
An alternative approach to transformation of the chloroplast might be to insert genes for proteins carrying chloroplast targeting sequences into the nucleus. There have been numerous attempts at stable nuclear transformation of dinoflagellates, but none has been clearly successful. An early report of the transformation of the dinoflagellates Amphidinium sp. and Symbiodinium microadriaticum mediated by silicon carbide whiskers, with selection for resistance to hygromycin or G418 and using b-glucuronidase (GUS) as a reporter, appeared to produce transformants after 12 weeks (ten Lohuis and Miller, 1998). However, there are no reports of successful use of this technique since the initial publication. In 2019, a series of experiments on Symbiodinium microadriaticum, using biolistics, electroporation and agitation with silicon carbide whiskers failed to introduce a chloramphenicol acetyl transferase (CAT) gene to the nuclear genome (Chen et al., 2019).
Despite this, Ortiz-Matamoros and co-workers reported transient expression of GFP in Symbiodinium using a plasmid designed for plant transformation introduced by treatment with glass beads and polyethylene glycol, and selection for resistance to the herbicide Basta (gluphosinate) (Ortiz-Matamoros et al., 2015a). However, transformed cells were not capable of cell division, and no genetic confirmation of transformation was carried out. Transformation with the same plasmid mobilized by the plant pathogen Agrobacterium was also reported, although the transformed cells again failed to divide (Ortiz-Matamoros et al., 2015b). The lack of stable expression of heterologous genes limits the use of these techniques for functional biochemical studies.
Here, we describe a method for stable transformation of the dinoflagellate chloroplast. The chloroplast genome of dinoflagellate species containing the carotenoid peridinin (which is the largest group and includes those forming symbionts with coral) is typically fragmented, comprising approximately 20 plasmid-like DNA molecules of 2-5 kbp known as 'minicircles' (Zhang et al., 1999), (Barbrook et al., 2014). Each minicircle typically carries a single gene, together with a conserved core region which is assumed to contain the origin of replication as well as the transcriptional start site (Howe et al., 2008). These minicircles have been shown to be localized to the chloroplast using in situ hybridization (Owari et al., 2014). Each chloroplast contains multiple copies of each minicircle, although the exact copy number varies according to the growth stage of a culture (Koumandou and Howe, 2007).
We exploited this unusual minicircular genome organization to create shuttle vectors for dinoflagellate chloroplast transformation. We created two artificial minicircles, both based on the psbA minicircle from the dinoflagellate Amphidinium carterae, as the A. carterae chloroplast genome is the best characterized amongst dinoflagellates (Barbrook et al., 2012;Koumandou et al., 2004;Barbrook et al., 2001). We replaced the psbA gene with a selectable marker (either a modified version of psbA which confers tolerance to the herbicide atrazine (Hirschberg and McIntosh, 1983), or a gene for chloramphenicol acetyl transferase (CAT), which confers resistance to chloramphenicol), and an E. coli plasmid backbone (to allow propagation in E. coli). We tested numerous transformation methods and were able to obtain sucessful introduction of these artificial minicircles into dinoflagellates using particle bombardment. Following selection, we could detect the presence of the artificial minicircles, and transcripts from them, using PCR and RT-PCR. We were able to detect the product of the introduced chloramphenicol acetyl transferase gene using immunofluorescence microscopy. Cultures under selection continued to divide and maintain the artificial minicircles for at least 1 year, indicating that transformation was stable. The availability of a method for dinoflagellate chloroplast transformation enables a range of studies on the maintenance and expression of this remarkable genome and the proteins it encodes, such as PsbA.
Construction of artificial minicircles
Two artificial minicircles were used in this study. The first, pAmpPSBA, was designed to confer atrazine tolerance. Tolerance to atrazine in plants can be conferred by a single residue change in the PsbA protein, where a Serine is mutated to a Glycine (Goloubinoff et al., 1984). We therefore cloned the A. carterae psbA minicircle into the E. coli vector pGEM-T easy (Promega) and introduced the necessary mutations into the psbA gene using site-directed mutagenesis, Figure 1. The second artificial minicircle, pAmpChl, was designed to confer chloramphenicol resistance. It is also based on the A. carterae psbA minicircle, but the psbA gene was excised and replaced by a A. carterae codon-optimized gene encoding chloramphenicol acetyl transferase. The plasmid backbone is E. coli pMA, Figure 1.
Glass beads and electroporation
There has been one previous report of stable transformation of dinoflagellates, using silicon carbide whiskers (ten Lohuis and Miller, 1998). This has never been reproduced, despite numerous attempts (Walker et al., 2005). Additionally, the whiskers are a significant health hazard, and, in other species, glass bead-mediated transformation has a higher transformation efficiency (Ortiz-Matamoros et al., 2018). There has been one previous report of transient transformation of dinoflagellates with glass beads (Ortiz-Matamoros et al., 2015a). We therefore first sought to replicate this finding, by transforming the A. carterae dinoflagellate chloroplast genome using glass beads. We used the artificial minicircle pAmpPSBA followed by selection with atrazine. The experiment was carried out three times, but no viable cells were seen following selection (i.e. following the addition of atrazine, all cells died at the same time that untransformed cells died).
Polyethylene glycol has been reported to increase glass bead transformation efficiency. We therefore added polyethylene glycol and repeated the transformation (in triplicate). No live cells were recovered following selection, indicating that this method did not give rise to stable transformants. A second attempt (in triplicate) was also not successful.
Many eukaryotic protist species can be transformed using electroporation, and the Lonza Nucleofector system is used with many hard-to-transfect species. This includes Perkinsus marinus, a sister group to the dinoflagellates (Burkett and Vasta, 1997). The artificial minicircle pAmpPSBA was used to attempt to transfect A. carterae using the Lonza Nucleofector with several settings indicated as suitable for protist transformation (X-100, D-023, L-029 and EH 100), each in triplicate. We were never able to recover any transformants.
Biolistic transformation with pAmpPSBA
We next turned to biolistics, since this has been reliably used to transform the unicellular alga Chlamydomonas for 30 years (Boynton et al., 1988). Nine experiments to introduce the pAmpPSBA artificial minicircle into A. carterae using biolistic transformation were carried out, using a range of rupture disk pressures. Each experiment was carried out in triplicate, and included a single negative control line (cells subjected to biolistic treatment but without pAmpPSBA). The mean survival time for each culture under selection was assessed by bright field microscopy, and results are shown in Table 1. In six experiments, A. carterae cells shot with particles carrying pAmpPSBA showed greater mean survival time under selection conditions than untransformed cells, suggesting successful transformation (Experiments A2, A3, A4, A5, A7 and A8). One experiment (A6) was harvested prior to the death of the control strain, so no conclusions can be drawn on relative survival times. Finally, two experiments showed no difference in the length of time for which cells survived. The first (Experiment A1) was carried out using the lowest pressure rupture disks. In this experiment, control and experimental cells survived just 13 days, suggesting that cells had not been transformed, perhaps because an insufficient bombardment velocity had been applied. In the second experiment (A9), selection was carried out with 1 mg ml À1 atrazine, below the lethal concentration of 2 mg ml À1 . Both test and control cultures survived at least three months, with subculturing occurring at 8-week intervals.
Biolistic transformation with pAmpChl
Transformation attempts were also made with A. carterae using chloramphenicol resistance as selectable marker. Experiments were carried out with pAmpChl and 1550 p.s.i. rupture disks. In the first experiment, chloramphenicol (final concentration 10-50 mg ml À1 ) was applied after 3 days in liquid culture, to allow time for initial synthesis of chloramphenicol acetyl transferase (Experiments C1A-E), Table 2. No untransformed wild-type cells (i.e. shot with gold particles without DNA) survived after 15 days, whatever the chloramphenicol concentration. However, at 10 mg ml À1 chloramphenicol, cells shot with particles carrying the pAmpChl plasmid survived for at least 35 days, Experiment C1A). When chloramphenicol concentration was 30 mg ml À1 or greater, cells shot with particles containing the pAmpChl plasmid had died by day 15, (Experiment C1C-E), Table 2. Note that where appropriate, cells were subcultured after 28 days.
Detection of artificial minicircles using PCR
To test if the transformation construct could be recovered from putatively transformed cultures, DNA was isolated from atrazine-selected A. carterae cultures (experiment A5, two lines designated A5.1 and A5.2) by vortexing with glass beads. DNA was also isolated from wild-type A. carterae as a Table 1. Biolistic transformation of A. carterae with pAmpPSBA. Each experiment was carried out in triplicate, thus producing three potentially transformed lines. In addition, one line of cells was subjected to biolistic bombardment, but without the pAmpPSBA ('untransformed'). Note that cultures from experiments 5-9 were harvested for genetic analysis, and thus the listed survival time is the day of harvesting, labeled with *. negative control. In addition, a DNA purification was carried out with transformed cells (line A5.1), but without vortexing with glass beads ('unbroken cells') in order to test whether DNA remained adsorbed to the outside of cells. A positive control was included using the pAmpPSBA artificial minicircle. PCR was performed using the primers MC-pG-F and MC-pG-R ( Figure 1) which lie on either side of the junction between the psbA minicircle and the pGEM-T Easy vector. A single product was amplified from each of lines A5.1 and A5.2, with no product detected from either the wild type or the 'unbroken cells' (Figure 2A). This product matched the size of product from the positive control. Products from both lines were cloned and sequenced. The sequence was as expected from pAmpPSBA, as a chimaera between the psbA minicircle and the pGEM-T Easy vector, confirming that the atrazine-resistant A. carterae did indeed contain the pAmpPSBA sequence.
To test if the pAmpChl artificial minicircle was present, DNA was isolated from three chloramphenicol-selected A. carterae lines (C1A.1, C1A.2 and C1A.3) after 35 days of selection. PCR using primers within the chloramphenicol resistance gene (CAT-F and CAT-R) was carried out. A positive control was included using the pAmpChl artificial minicircle. A single product was obtained from lines C1A.1 and C1A.3, with no product detected from line C1A.2 ( Figure 2B). Products from lines C1A.1 and C1A.3 were cloned and sequenced. The sequence was the same as that expected from pAmpChl, which confirmed that the chloramphenicol-resistant A. carterae did indeed contain the pAmpChl sequence.
Relative copy number of artificial minicircles
As we utilized the backbone of an existing psbA minicircle to create pAmpChl we tested to see what proportion of the psbA minicircles contained the psbA gene and what proportion contained the CAT gene (and associated shuttle vector). We therefore designed primers immediately flanking the psbA gene (copy-F and copy-R), in a region that was common to both the psbA and pAmpChl minicircles, and carried out PCR. This should amplify either psbA or CAT. Analysis of the products using agarose gel electrophoresis revealed the presence of two bands, one corresponding to the psbA gene and one to the CAT gene, as shown in Figure 3. Taking into account the size of the products, the likelihood that the CAT-specific product would be generated more efficiently, being smaller, and the relative intensity of each band, it would appear that there are roughly similar numbers of each minicircle in the chloroplast.
Transcription of the artificial minicircles
In order to test if transcripts from the two artificial minicircles could be detected in the putatively transformed lines, total RNA was extracted and purified. cDNA was synthesized using RNA from Each experiment was carried out in triplicate, thus producing three potentially transformed lines. In addition, one line of cells was subjected to biolistic bombardment, but with gold particles lacking the pAmpChl ('untransformed'). For experiment 1, cells from each plate (three shot with gold particles carrying the plasmid and one with gold particles only) were divided into five separate samples, each incubated at a different chloramphenicol concentration. Note that cultures from experiments C1A, C2, and C3, were harvested for genetic analysis, and thus the listed survival time of lines still alive at that point is the day of harvesting, labeled with *. Experiment C4 was still alive at 57 weeks and is thus marked with +. atrazine-selected cultures from experiment A6 (lines A6.1, A6.2, A6.3 and untransformed) and random hexamer primers, followed by PCR with the specific primers MC-pG-F-II and MC-pG-R-II. cDNA was synthesized using RNA from chloramphenicol-selected culture lines C3.1, C3.2 and C3.3 and the gene-specific primer CAT-FSS, followed by a nested PCR strategy. Primers CAT-F and CAT-R were used in the first round of PCR (30 cycles). 1 ml of PCR product was used as template for the second round of PCR (10 cycles) with primers CAT-F-Nest and CAT-R-Nest. Negative controls, which omitted the reverse transcriptase, were included for all RT-PCRs. A positive control was included using the pAmpPSBA or pAmpChl artificial minicircle. RT-PCRs using RNA from the three atrazine-selected lines in Experiment A6 all yielded a band consistent with the size of the positive control, Figure 4A. The DNA in the bands was purified, cloned and sequenced. The sequence of all three matched the pAmpPSBA artificial minicircle, confirming that it was transcribed. The sequence spanned the site of the atrazine resistance mutations and included the expected sequence alterations. The negative control yielded no PCR products (data not shown). The same results were obtained for three lines in each of Experiments A7 and A8 (data not shown).
RT-PCRs using RNA from the three chloramphenicol-selected lines (C3.1-C3.3) yielded bands from two of the three cell lines in Experiment C3 ( Figure 4B). The PCR products were sequenced directly and shown to correspond to the pAmpChl minicircle. The negative control yielded no PCR product. Two of three lines in Experiment C2 yielded bands in RT-PCRs (data not shown). Artificial minicircle products localize to the chloroplast In order to confirm that we had transformed the chloroplast genome and not the nuclear genome, we carried out an immunofluorescence assay with an anti-chloramphenicol acetyl transferase antibody, together with a secondary antibody labeled with Alexa Fluor 405. The antibodies colocalized with the chloroplast, identified as the region of the cell with significant autofluorescence, as shown in Figure 5. The antibodies did not localize with the nucleus or other parts of the cell. This result shows that the CAT protein is expressed in the chloroplast (compare the merged image with the chloroplast autofluorescence and CAT images). As the CAT gene did not include a chloroplast targeting sequence (which would be necessary to target nuclearencoded proteins to the chloroplast), the pAmpChl artificial minicircle must be located in the chloroplast.
Stability of transformation
To test if the atrazine-resistance phenotype transformation of the dinoflagellate chloroplast was stable under low-level selection, cells were shot with gold particles carrying pAmpPSBA and cultured under continuous atrazine selection at 1 mg ml À1 (experiment A9 in Table 1). Cell counts increased over time, although at a rate much lower (~10%) than untransformed cells under no selection. An untransformed cell line was also maintained, which survived under the same atrazine concentration (1 mg ml À1 ). Both cell lines were subcultured at 8-week intervals. After 3 months, cells were harvested and DNA was isolated. PCR using the primers MC-pG-F and MC-pG-R was carried out (as above). A positive control PCR was included using the pAmpPSBA vector, and PCR with DNA isolated from the untransformed, wild type cells maintained at non-lethal atrazine concentration was included as a negative control ( Figure 6A). A single product, of expected size, was obtained using the three transformed cell lines, with no product detected from the untransformed cells. DNA sequencing confirmed that the products from all three lines corresponded to pAmpPSBA. This showed that the transformation of A. carterae with pAmpPSBA was stable at a non-lethal atrazine concentration.
To test if the chloramphenicol-resistance phenotype transformation of the dinoflagellate chloroplast was stable, three lines were generated by shooting with gold particles carrying pAmpChl and cultured under continuous chloramphenicol selection at 20 mg ml À1 (Experiment C4 in Table 2) with subculturing every 14 days. After 18 weeks, well after control untransformed cells had died, a sample of cells was harvested from each transformed line and DNA was isolated. A PCR reaction using the primers CAT-F-Nest and CAT-R-Nest was carried out on each sample. A positive control PCR was included using the pAmpChl vector ( Figure 6B). A single product, of expected size, was obtained for each of the three transformed cell lines. DNA sequencing confirmed that the products from all three lines corresponded to pAmpChl. No band was seen in a wild-type PCR carried out with the same primers on wild-type cells (data not shown), confirming that the band could only have arisen
Discussion
Here, we present evidence for the first stable transformation of the dinoflagellate chloroplast genome. By making use of the plasmid-like fragmented chloroplast genome and a biolistic system, we have introduced a modified version of an existing sequence as well as a heterologous gene. These genes are transcribed, and produce protein, as shown by immunofluorescence microscopy and the presence of an expected phenotype. Stable transformation was achieved with two separate artificial minicircles, one containing a modified psbA gene designed to confer atrazine tolerance and another encoding chloramphenicol resistance, with cultures surviving at least 1 year under selection. The copy number of the artificial minicircle is similar to that of native minicircles.
A. carterae cells transformed with the modified psbA gene (atrazine tolerance) survived under non-lethal concentrations of atrazine for at least 3 months, retaining the modified gene, indicating that the transformation is stable even under low levels of selection. The results suggest it is important to titrate the concentration of selective agents used. With an atrazine concentration of 2.5 mg ml À1 some of the transformed cultures did not survive more than a few days longer than untransformed ones, suggesting that the modified PsbA was at least partially inhibited at that higher atrazine concentration. In addition, it is possible that a background of minicircles containing wild-type psbA genes competed with the introduced artificial minicircles for replication or transcription factors, making it difficult for adequate levels of atrazine-insensitive PsbA to be maintained to cope with the higher atrazine concentration. With chloramphenicol concentrations of 30 mg ml À1 or above, the survival of transformed and untransformed strains was similar. However, at 20 mg ml À1 or lower the transformed cultures outlasted the untransformed ones, and some were able to survive apparently indefinitely.
There have been multiple previous reports of transformation in dinoflagellates. Although the initial report (using silicon carbide whiskers) described transformation as stable (ten Lohuis and Miller, 1998), it has never been successfully reproduced (Walker et al., 2005). Transient expression has brightfield chloroplast CAT merge Figure 5. A chloroplast localization for chloramphenicol acetyl transferase. Immunofluorescence microscopy using the A. carterae pAmpChl line. Cells (brightfield) showed significant autofluorescence in the chloroplast (red). A primary antibody specific for CAT with a secondary Alexa Fluor 405 antibody (blue) showed localization of CAT to the chloroplast (indicated by the overlay image labeled merge). DOI: https://doi.org/10.7554/eLife.45292.008 been carried out using glass beads (Ortiz-Matamoros et al., 2015a). However, we were unable to build on this result to obtain stable transformation, despite numerous attempts (and there have not been any subsequent reports of transformation using either glass beads or Agrobacterium). Instead, we found that microparticle bombardment gave stable transformants. Our results -using the same artificial minicircle in all experiments -suggest that the primary reason for failure of the majority of methods to give stable transformation is due to inability of the DNA to enter the cell, rather than inherent dinoflagellate genetics. The ability to modify the dinoflagellate chloroplast genome will be of enormous value in many areas of dinoflagellate biology. Modification of existing minicircles should allow us to study many other aspects of this highly unusual chloroplast genome, such as the promoter regions of the genes. For example, many chloroplast genes are down-regulated under high temperature stress (Gierz et al., 2017). Little is known about how transcription is regulated, or initiated (Barbrook et al., 2012), though it is assumed that initiation occurs in the conserved core region of the minicircle (Barbrook et al., 2001;Zhang et al., 2002). We also do not know how the minicircles are replicated (Barbrook et al., 2018), although again it is assumed the core region is important. It will now be possible to mutate this core region to determine which sections are important. The ability to express heterologous proteins will be of great value in studying a wide range of other aspects of dinoflagellate chloroplast biology. The ability to express modified forms of the PsbA protein will be of particular value in studying the role of this protein in the response by dinoflagellates to the disturbances that are believed to precipitate coral bleaching. Culturing of Amphidinium carterae A. carterae CCMP1314 (from the Culture Collection of Marine Phytoplankton) was cultured in f/2 medium on a 16 hr light/8 hr cycle, 18˚C at 30mE m À2 s À1 , as described previously (Barbrook et al., 2006). Cells were returned to this light regime immediately following all transformation methods.
Design of artificial minicircles
The pAmpPSBA artificial minicircle (predicted to confer atrazine tolerance) was prepared by PCR amplification of the wild-type psbA mincircle with outward facing primers from a point immediately downstream of the proposed poly-U addition site (Genbank AJ250262, fwd primer 1128-1155, rev primer 1127-1106) (Barbrook et al., 2012). The linear PCR product was purified and cloned into the pGEM-T Easy plasmid (Promega), which contains the ampicillin resistance marker and a bacterial origin of replication. The point mutations necessary to confer atrazine tolerance were introduced in a further round of PCR with Pfu polymerase and the following mutagenic primers, forward primer GTC TTATCTTCCAGTATGCTGGCTTCAACAACTCCCGTTCTC, reverse primer GAGAACGGGAGTTG TTGAAGCCAGCATACTGGAAGATAAGAC. This altered a TCC (Serine) codon to a GGC (Glycine) codon at position 260 of the PsbA protein (numbered as in AJ250262). The PCR products were treated with DpnI to digest any template DNA and then used to transform chemically competent E. coli JM109. Ampicillin selection was used to identify colonies containing pAmpPSBA, and plasmids were sequenced. A plasmid map is shown in Figure 1. The full vector sequence is given in Supplementary Data.
The pAmpChl vector was synthesized by GeneArt. This vector was based on a pMA vector backbone, and contained the psbA minicircle (as above), but with the psbA coding region removed and precisely replaced by an A. carterae chloroplast codon-optimized E. coli chloramphenicol acetyl transferase gene (CAT) (Barbrook and Howe, 2000), Figure 1. The full vector sequence is given in Supplementary Data.
Both vectors were propagated in E. coli under ampicillin selection, and isolated using the Promega Maxiprep plasmid purification protocol prior to transformation into A. carterae. The vectors were verified by DNA sequencing before use.
Glass bead mediated transformation
A. carterae cells (1.3 Â 10 7 ) were transformed with pAmpPSBA using glass beads, with or without polyethylene glycol, following the protocol as described by Ortiz-Matamaros et al (Ortiz-Matamoros et al., 2015a). A control, where cells were treated with glass beads but without pAmpPSBA, was carried out at the same time. Selection was applied (2 ug/ml atrazine) after 24 hr. Each experiment was carried out in triplicate (i.e. three reactions, three controls).
Biolistic transformation of A. carterae
Biolistic transformation was carried out using a Biorad Biolistics PDS-1000/He system, Biorad rupture disks, stopping screens and macrocarriers. Preparation of particles carrying DNA was carried out using Seashell Technology's DNAdel gold carrier delivery system and 550 nm diameter gold particles.
A. carterae cells were grown to early log growth phase before harvesting prior to transformation. For each transformation,~2.5Â10 7 cells (as determined by light microscopy utilizing a haemocytometer) were spotted onto the center of a 1% agarose f/2 medium plate and allowed to dry. 0.5 mg of gold particles and 0.5 mg of vector DNA were used for each plate to be transformed. Each plate was shot using the above-mentioned Biorad Biolistics PDS-1000/He system and rupture disks of either 1100 PSI, 1350 PSI or 1550 pounds per square inch (p.s.i.) (see Table 1 for details).
Cells were immediately resuspended in 30-50 ml fresh f/2 medium and allowed to recover before the addition of the selective agent. Cells shot using the pAmpPSBA artificial minicircle were allowed 16-24 hr to recover. Cells shot using the pAmpChl artificial minicircle were allowed 72 hr to recover. Cells were maintained in liquid culture as they do not grow on solid medium. Medium was replaced every 4 weeks (atrazine) or 2 weeks (chloramphenicol). Cells were subcultured (two fold dilution) every 8 weeks (atrazine) or 4 weeks (chloramphenicol). A step-by step protocol is described in 'Biolistic Transformation of Amphidinium' (https://www.protocols.io/view/biolistic-transformation-ofamphidnium-hnmb5c6).
Culture survival was assessed by microscopy. A spot of 50 ml was placed onto each of three microscope slides for each culture. After covering with standard coverslips, the entire volume for each was examined using a light microscope at x100 magnification. Cells were assessed as living if they showed more than simple Brownian motion. In general, 'dead' cells appeared to disintegrate shortly after movement ceased. If no living cells were found 3 days in a row, the culture was recorded as dead on day 3.
Extraction of DNA and RNA from A. carterae
Total RNA was isolated from A. carterae using the Trizol -chloroform method. Purification was carried out using the RNA clean-up with on-column DNase protocol of the Qiagen RNeasy kit as described (Rio et al., 2010) except that isopropanol precipitation was carried out overnight, rather than for 10 min.
DNA was released from cells prior to PCR by resuspending 5 Â 10 4 to 10 7 cells (depending on the number available) in 50 ml dH2O with~10-20 acid-washed 500 mm glass beads and vortexing for 10 min.
RT-PCR and PCR
First strand synthesis of the RNA was performed using Invitrogen Superscript IV using the manufacturer's protocol and either random hexamer primers or a gene-specific primer. Negative controls lacking reverse transcriptase were performed by the same method but replacing the reverse transcriptase enzyme with dH 2 O. PCR was carried out using Promega GoTaq polymerase according to the manufacturer's instructions, and annealing temperature, extension time and MgCl 2 concentration were varied as appropriate.
Cloning and sequencing of PCR products PCR products were separated by 1-1.5% agarose gel electrophoresis and visualized by staining with GelRed. PCR products were purified from excised gel pieces using the MinElute gel extraction kit (Qiagen). Some PCR products were directly sequenced after gel extraction whilst others were ligated into the pGEM-T Easy plasmid vector (Promega), following the manufacturer's instructions. The ligation mix was used to transform chemically competent Escherichia coli TG1, followed by overnight growth on 1.5% LB agar containing ampicillin at 100 mg/ml. Individual colonies were picked and grown overnight in LB containing ampicillin at 100 mg/ml. Plasmids were extracted from resulting cultures using the QIAprep Spin Miniprep Kit (Qiagen). All sequencing was carried out using an Applied Biosystems 3130XL DNA Analyser in the Department of Biochemistry, University of Cambridge sequencing facility.
Immunofluorescence microscopy
Five  10 5 cells were fixed in 1% paraformaldehyde in f/2 medium for 5 min. The reaction was quenched by the addition of glycine to a final concentration of 0.125 M for a further 5 min. Both steps were carried out with constant agitation. Cells were washed in f/2 medium three times for 5 min each, and then permeabilised by addition of 0.2% Triton-X-100 for 15 min, again with constant agitation, and washed three further times in f/2 medium. Blocking was carried out in 5% BSA in f/2 for 30 min under constant agitation.
Cells were incubated with a rabbit anti-chloramphenicol acetyl transferase as the primary antibody (Antibodies online, ABIN285051) at a final concentration of 1:500 in f/2% and 5% BSA for 1 hr under constant agitation, and washed three times in f/2 medium for 5 min. Cells were then incubated with an anti-rabbit Alexa Fluor 405 secondary antibody (raised in goat; ThermoFisher) at a final concentration of 1:1000 in f/2% and 5% BSA for 1 hr under constant agitation, washed in f/2 medium three times for 5 min, and mounted on VWR-polysine-coated slides using Vectashield mounting medium. Cells were visualized using a Nikon C2 confocal microscope. | 7,070 | 2018-10-24T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
Voltammetry coupled with impedance spectroscopy
Dynamic EIS (dEIS) is the joint use of cyclic voltammetry and electrochemical impedance spectroscopy. A method is planned for data evaluation which involves transformations yielding potential program invariant forms of certain characteristic functions. This way of calculation is illustrated by the analysis of two archetypes of electrochemical kinetics. A future task is to develop the related theories and to perform demonstration experiments aimed at establishing dEIS as a robust, general-purpose technique for characterizing electrochemical systems.
Introduction
There are two extremes of doing research. One is the work of pioneers-breaking the ground, finding the tools, setting the rules, and showing the directions for the followers. The other is the job of the consolidators: re-doing experiments to confirm (or disprove) the existence of phenomena and/or to provide reliable data, testing or falsifying parts of theories, and reformulating concepts to make the theories concise, that is, to make disciplines well-established. Pioneering is unpredictable, consolidation is a never-ending task.
The need for consolidation comes not only from new phenomena and materials but also from technical advancement. This is so also when classical electrochemical methods are considered. These methods, as summarized in a widely used, excellent textbook [1], obtained their mature forms before the era of digital data acquisition-in the age of flatbed recorders. At that time, evaluation of the charts was done by comparing coordinates of characteristic points (e.g., peak location and height) with the related equations. Nowadays, data acquisition methods and computational facilities allow us to include all (or at least many) data points into the analysis instead of a few characteristic points only and to perform complex analyses utilizing the information carried by the entire measured curve.
Accordingly, the theory on electrochemical kinetics-the subject of which are charge transfer processes affected also by adsorption and transport from/to the interface-should be reshaped, enabling one to extract as much information as possible from measured data. This is, certainly, a task to deal with also in the future. One possible way is outlined here.
The current state
A couple of classical, basic electrochemical measurement techniques exist based on measurement of potential E, and current j, as a function of time t (or frequency ω). Some of them, like cyclic voltammetry, CV, are performed in a broad potential range, while others-collectively called electrochemical impedance (or immittance) spectroscopy, EISusually in a narrow one.
The two methods, CV and EIS, are typically used for different purposes: with CV, we learn the qualitative features (e.g., what kind of reactions take place and what is the order-of-magnitude of currents), whereas EIS is used to get quantitative information on some known properties (e.g., what is the exact rate of a reaction). Just as the old scientistclassifying joke (about "impedance transformers" and "potential sweepers") implies, electrochemists are rarely familiar with both methods; the textbooks teach us the two methods separately and few if any connection is shown between the results of the two types of measurements. This is so, even though there exist some variants of CV and EIS which include some element of the other method (e.g., AC voltammetry or EIS measured sequentially on stepwise set potentials or harmonic analysis with large amplitude perturbations).
There are two links connecting the two methods: one is the experimental method of dynamic EIS (dEIS), and the other is a theoretical one: the potential program invariant (PPI) representation of the measured j(t) vs E(t) voltammetry data.
dEIS is a combination of CV and EIS: it is their simultaneous use. During dEIS measurements, the potential is scanned just like in the case of the classical CV; however, it is also perturbed by a small amplitude, periodic, noise-like voltage. This way, audio-frequency impedance spectra can be measured continuously as the potential is varied. dEIS is apparently a complicated method; yet, it does not require sophisticated instruments, only a commercially available data acquisition system, some modifications on the potentiostat plus the software. Such instruments have been developed at many places among them in our laboratory, for various applications which are beyond the present scope. Note that the implementation of dEIS to potentiostats will be an important (and probably profitable) task of instrument manufacturers.
One difference between CV (or in general: the j(t) vs E(t) voltammetry) and EIS is to be emphasized here: EIS has a standardized output in the sense that it produces spectra that are independent on the actual form of potential or current perturbation by which they have been measured. In contrast, the results of voltammetry measurements do depend on the E(t) function (scan rates and turning potentials). However, appropriate transformations may yield functions which have lost their connection with the potential program by which they have been measured-that is, they result in PPI functions. The classical example for such a transform is the semi-integrated form of cyclic voltammograms of reversible redox couples, as demonstrated five decades ago [2]. This is the archetype of the PPI representation of voltammograms taken in a broad potential range.
In principle, the same electric properties of a system can be determined either by a set of CVs taken with multiple scan rates or from EIS measured in broad frequency range at multiple potentials. In this sense, the methods CV and EIS are equivalents, and hence, in the ideal case, one should be able to transform a set of CVs to another set of EIS and vice versa. In the same vein, from dEIS measurements, both results (i.e., CV and EIS) should yield the same physical-chemical quantities as from complete sets of separate CVs or EIS. There exist numerous CV and EIS theories; however, only a few treat the two methods simultaneously with the aim of exploring the connections between them. A future task is to elaborate such theories. Two simple textbook cases might serve as an illustration.
Two examples of PPI representations
In the first example, we consider the diffusion-controlled charge transfer reaction with infinite reaction rates taking on a noble metal electrode of planar geometry. The CV of this "reversible system" is the well-known curve with asymmetric peaks and big hysteresis; this curve does depend on the scan parameters such as the scan rate (see Fig. 1a). However, applying a convolution procedure on the current-time curve, one can get the so-called semi-integrated form, M rev , so that the M rev vs E curve becomes hysteresis free and scan rate independent, i.e., a PPI function. The derivative dM rev /dE vs E, per se, is also PPI. If the same system is investigated by impedance measurements, one obtains spectra which can be analyzed in terms of a Randles circuit, with zero charge transfer resistance. The coefficient of the Warburg admittance, σ W,rev , exhibits a potential dependence which is just the same as that of dM rev /dE vs E obtained from the CVs, as it has been shown theoretically [3] and experimentally (Fig. 1b, [4]). To conclude, there is a PPI function, M rev , which appears in the theories of both CV and EIS in the context of charge transfer of reversible redox systems. M rev and dM rev /dE are the PPI function of large-and smallsignal ("global" and "local") responses.
The second example concerns the same system but with finite charge transfer rates. The CV of this "quasi-reversible system" is similar to that of a reversible one, but the peakseparation is scan rate dependent (see Fig. 2a). By the methods described in ref. [5], one can transform the set of CVs to two PPI functions: j inf , which characterizes charge transfer alone, and M rev , which is a property of the diffusional flux (see Fig. 2b.) For analyzing EIS of the same system, the equivalent circuit is a full Randles circuit. In the case of dEIS, the charge transfer resistance and the Warburg admittance coefficient depend on the scan rate. However, with a transformation described in ref. [6], one can get rid of the scan rate dependences and get PPI functions dj inf /dE and the σ W,rev . = dM rev /dE.
Outlook
Development and testing of such theories are in progress, for various electrochemical situations. Just as other theories and methods, they are by no means universal. Rather they are elements of a toolbox, just like the equivalent circuit elements are in impedance analysis. Nevertheless, these theories will yield three great advances over the existing ones: 1. Whereas the conventional voltammetry methods yield curves which depend both on electrochemical properties of the system and measurement parameters, the PPI functions depend on the former only. In this sense, they are the appropriate representations of the physical-chemical properties. 2. PPI functions are usually calculated by integral transforms; hence, these theories allow us to determine physical-chemical quantities from many data points rather than from single or a few ones (from peak heights, peak distances, etc.). Such procedures improve measurement accuracy and robustness in general. 3. Finally, dEIS theories involving PPI functions are expected to carry the same information in a much more concise way than the present (CV and EIS) ones known from textbooks.
The new theories are expected to consolidate the background of dEIS evaluations and to improve the accuracy of the results. This will, in turn, open new routes for the pioneers of electrochemistry. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. | 2,331 | 2020-06-07T00:00:00.000 | [
"Chemistry",
"Engineering"
] |
EVALUATING ACCESSIBILITY OF MALAYSIAN PUBLIC UNIVERSITIES WEBSITES USING ACHECKER AND WAVE
Websites become essential means for most universities to communicate, exchange of relevant information and enable transactions among their stakeholders. Therefore, website accessibility accessible website is crucial to students to ensure equal access to of the university’s information regardless of their physical disabilities and other possible limitations. This study reports the web accessibility of 20 Malaysian public universities based on AChecker and WAVE. The results suggest a relatively low level of compliance to the guidelines as specified in WCAG 2.0 and Section 508. Among the aspects that deserve immediate attention are the provision of text alternatives for any non-text contents, keyboard accessibility and colour contrast. Other concerns such as navigation, adaptability, input assistance, compatibility, empty link and empty heading can be further improved. Regardless of low conformance, most websites extensively integrated some of the accessibility features as set out by Section 508. Overall, this study offers meaningful insights, particularly to web developers for better compliance with the standards while designing their websites.
INTRODUCTION
Evolving Internet functionalities since 1990s have enabled dramatic improvement for business communication.Nowadays, websites are gaining popularity as one of the potential means for disseminating or exchanging of information with various stakeholders.In fact, a website offer dual side benefits to both the information provider and the information requester (users).For the former, the website enables the business to effectively supply information to the targeted audience.As for the latter, the website features employed, ensure users to have better access to richer information/services apart from having a smooth navigational experience (Williams & Rattray, 2005).
Unlike the private sector, public institutions employ websites to improve the delivery of services to the citizens or other related stakeholders.Specifically, setting up a highly effective website is crucial for the information-intensive sector such as public universities.The website turns to be a resource centre for the stakeholders, which includes prospective and current students, potential and existing employees, visitors and the alumni.Furthermore, due to a growing trend among students to seek information for pursuing their higher degrees (Schimmel et al., 2010), an effective web-based communication channel helps institutions to address the students' information needs.Nevertheless, having necessary information available is not sufficient unless it is made accessible to the stakeholders (Kamoun & Almourad, 2014).
Accessibility represents one of the web-quality aspects that ensures effective use of the website, ease of navigation and understanding its structure despite having physical disabilities or other constraints (Shawn, 2006).The World Wide Web Consortium (W3C) (2009) states that an accessible website supports people with disabilities to perceive, understand, navigate, and interact effectively for equal chances to contribute to the web communication activities.Thus, web accessibility guarantees that all potential users can access web applications regardless of an individual's limitations or the context of use.The definitions imply that web accessibility is not only restricted to the needs of people with physical disabilities to use the web but also to include people with different skills, preferences and needs.This includes individuals who have slow Internet connection, suffer from temporary or age-related disabilities and have different technological capabilities such as browser type, screen sizes, or device type (Andrés, Lorca & Martínez, 2010;Providenti & Zai III, 2011).In short, web accessibility concerns are beyond the need of disabled people.More importantly, accessibility concerns the ability of users to navigate the website by taking into consideration the differences in their requirements and contextual environment.Apart from augmenting browsing experience, an accessible website also facilitates organisations to have effective website updating activities (Hofstader, 2004), offers better web visibility and increases traffics to its website (Williams & Rattray, 2005).As such, institutions have to design websites that are highly flexible to address all these diversities and potentials (W3C, 2005).
While earlier studies focused on different organisational contexts and employed various accessibility software, this study reports the accessibility of the Malaysian public universities websites.Specifically, the objectives of this study are twofold,.Firstly, to report the current status of public universities compliance on web accessibility as stipulated by the most recent guidelines i.e.WCAG 2.0 and Section 508 of the United States Rehabilitation Act 1973, secondly, to identify areas with highest incidences of non-compliance that could be useful sources to rectify the problems (as per reports of AChecker and WAVE automated tools).
Similar to other countries, the web accessibility issue deserves special attention in Malaysia.The Law of Malaysia (2008) on Person with Disabilities Act 2008 (Act 685) (Part IV) clearly specifies the equal rights of the disabled people to access any public facilities or services, which include basic facilities, infrastructure, transportation, education, employment and access to ICT-related services.Specifically, the government and the information providers have to provide ICT services in a format which is accessible to people with different kinds of disabilities without additional cost (Section 30, para 1 to 3).With the number of disabled people increasing to about 69%, from 314,247 in 2010 to 531,962 in 2014 the (Department of Social Welfare, 2014), government has taken their education as one of the priorities in the government agenda.As specified clearly in the Person with Disabilities Act (2008), education for persons with disabilities should form an integral part of national educational planning, curriculum development and school organization.As of 2013, about 1,572 of the 481,361 students enrolled in public higher institutions nationwide are those with certain disabilities (Ministry of Higher Education [MOHE], 2014).Thus, public universities are expected to comply with the Act so as to ensure equal access of information, resources, facilities and ICT (including website) to these group of students.
Investigation of the web accessibility of public universities also resembles the government's initiative to strengthen public sector efficiency via the Malaysian Government Portals and Websites Assessment (MGPWA).A project spearheaded by the Multimedia Development Corporation (MDeC) aims to heighten government service delivery amongst the public sectors.Conformance to W3C Disability Accessibility standard was among the criteria specified in the MGPWA guideline.
The outcome of this study facilitates web development activities amongst public universities.Web accessibility evaluation facilitates reengineer the process of the website after the stage for the purpose of continuous improvement of the website performance (Boldyref, 2002).Moreover, continuous efforts to understand users' requirements can be a starting point towards more effective website development strategies (Yates, 2005).In the context of public universities, having more accessible websites warrants equal access to educational related information for disabled people (Shawar, 2015), apart from taking care of diverse stakeholders needs for better web surfing experience.
WEB ACCESSIBILITY GUIDELINE AND STANDARDS
The World Wide Web Consortium (W3C), an international consortium that combines full-time employees and the public towards cumulative efforts to propose a standard for web accessibility.The Consortium initiated an exercise known as Web Accessibility Initiative (WAI), which later produced the Web Content Accessibility Guidelines (WCAG).The guideline becomes the defacto standard for evaluating accessibility of the web (Rømen & Svanaes, 2012).The first version of the guideline (WCAG 1.0) that came into picture in the late 1990s, offered guidelines in designing accessible websites regardless of the physical, sensory or cognitive abilities of the audience.While W3C does not impose any legal enforcement of the guideline, it does offer meaningful insight to make more accessible websites (Yates, 2005).
W3C officially released a newer version (WCAG 2.0) in 2008(WCAG 2.0) in (W3C, 2008)), which encompassed a wider range of recommendations for accessible web content.The guideline was grounded on four principles, perceivable, operable, understandable, and robust (Peters & Bradbad, 2010).As WCAG (2008) outlined perceivable principle concerned the presentation of information and user interface components that facilitate them to perceive.Meanwhile, operable denotes of the operability of user-interface components and navigation functionalities.The third principle focused on understandable presentation of information and user-interface operation.Finally, the robust principle stressed the reliable interpretation of the web content regardless of the user agents used.The WCAG 2.0 guideline further extended the accessibility conformance into five distinct aspects; conformance level, full page, complete processes, only accessibility-supported ways of using technologies and noninterference (W3C, 2009).
The revised guideline comprised of 12 guidelines and 61 success criteria with three levels of conformance, i.e.Priority 1 (Level A), Priority 2 (Level AA) and Priority 3 (Level AAA) (W3C, 2008).Level A (the lowest compliance) specified the compulsory elements of the web for people with disability to get access to the materials provided.Meanwhile, Level AA listed the advanced requirements that likely removed significant accessibility barriers for a wider group of audience to be able to access the web content.The highest conformance level (Level AAA) stipulated other advanced features that ensured widest accessibility of the web among the audience.Meeting the requirements of the least Priority 1 was crucial for most organisations in order to offset between an economic performance and a social/moral performance (Williams & Rattray, 2005).All in all, the revised requirements made WCAG 2.0 more educational, able to produce reliable results, testable elements, and produce unambiguous interpretation (Ribera, Porras, Boldu, Termens, Sule, & Paris, 2009).
Similarly, Section 508 of the United States Rehabilitation Act 1973 also addressed the issue of equal access amongst the disabled.The act required the US federal agencies to provide equal access to electronic and information technology applications.The Act stressed on 16 web components in designing and presenting accessible websites (WebAiM, 2013).
AUTOMATED TOOLS FOR WEB ACCESSIBILITY
Considering the complexity of the web functionalities, the emergence of automated tools complement the existing standards/guidelines to objectively evaluate web accessibility compliance.At present, there are various automated tools to assist quick and objective reviews of web accessibility.Accessibility Valet, AChecker, Cynthia Says, EvalAccess, FAE, MAGENTA, OCAWA, TAW, WAVE and Web Accessibility Checker are the top-10 free tools that are available on the net (Source: http://usabilitygeek.com/10-free-web-basedweb-site-accessibility-evaluation-tools).The tools that are capable of quickly generating assessment results, are easy to use and offer great usability explain their popularity (Lujan-Mora, Navarrete, & Penafiel, 2014).In addition, automated tools that highly correspond to the applicable web accessibility standards/guidelines further enhance reliability and relevancy of the results generated.Table 1 shows the selected tools and their embedded web accessibility standards.While automated tools are highly useful to objectively predict web accessibility, part of the results reported require human judgment or manual inspection of the web (Peters & Bradbad, 2010).Hence, interpretation of the results generated from an automated tool deservescareful scrutiny and should be interpreted within its limitations.
Most Popular Automated Accessibility Tools and Applicable Standards
Several studies have evaluated web accessibility of higher learning institutions.Using HERA and WAVE as the assessment tools, Adepoju and Shehu (2014) reported substantial incompliance of 36 Nigerian federal universities' websites on accessibility guidelines.Similarly, an examination of 74 Spanish universities reported low to moderate levels of compliance on accessibility standard (Chacón-Medina, Chacón-López, López-Justicia, & Fernández-Jiménez, 2013).Laitano (2015) who diagnosed Argentine public universities, websites based on WCAG 2.0 guidelines indicated serious accessibility issues for most parts of the websites, particularly on the syntax, presentation of the web content and non-text content, and visual readability of the text.Finally, a cross-countries study on selected higher education institutions in the Middle East and England concluded that website accessibility in developed countries was better off than the developing economies (Shawar, 2015).
Despite the claim that e-government practice is improving in Malaysia and the websites are showing higher maturity (Abdul Rashid, Othman, & Hassan, (2014), studies on web accessibility indicate otherwise.Abdul Aziz, Wan Mohd Isa and Nordin's (2010) investigation on 120 websites of higher education institutions pointed out various accessibility and usability issues.Lastly, a longitudinal study on Malaysian public higher institutions' websites using three automated tools showed little improvement of web accessibility over a two-year period (Abuaddous, Jali & Basir, 2013).Other than the studies on higher institutions, several studies on other sectors are worth mentioning.Abdul Latif and Masrek (2010) reported the current status of Malaysian e-government websites and attained webmasters' perception on web accessibility.As the study reported, all the websites did not even meet the lowest accessibility compliance level (Priority 1).Viewing the issue from the disabled group perspective, another study found that the use of images, hyperlinks and page layouts on the website failed to meet the requirements of visually impaired persons (Ramayah, Jaafar & Mohd Yatim, 2010).
With respect to the assessment tools, previous works have employed different tools and approaches for evaluating web accessibility.Table 2 offers an insight of the selected tools used in earlier studies.The data indicates popularity and diversity of tools used as objective measure of web accessibility.Extending from previous works, this study evaluates the current state of web accessibility compliance of Malaysian public universities as outlined by WCAG 2.0 and Section 508.The next section elaborates in greater detail the data collection process and is followed by the findings and discussion of the results.
METHODOLOGY
This study examined the web accessibility of 20 Malaysian public universities' websites.Considering its popularity, it being freely available and its extensive integration of accessibility standards (WCAG 2.0 and Section 508), this study employed the Accessibility Checker (AChecker) and Web Accessibility Versatile Evaluator (WAVE) as preferred automated tools.As per the WCAG 2.0 standard, websites need to comply with Level A requirements that allow disabled persons to use the websites.This corresponds to the Provider-Based Evaluation (ProBE) 2015 assessment guidelines.ProBE is an initiative by the Multimedia Development Corporation (MDeC) for websites' selfassessment manuals and web maintenance tools for Malaysian government agencies (including the public universities).Although the guidelines specify web accessibility as non-mandatory criteria, their requirement is getting more important over time as one of the strategies to increase web usage and improve user experience.
The authors carried out website accessibility assessment using both automated tools from Monday, 21 March 2016 to Tuesday, 22 March 2016.The assessment particularly focused on the respective institution's homepage.Assessing multiple webpages offers richer information, but this study restricted the assessment to the homepage for a quick review of web accessibility, in view of the dynamic nature of the web.Moreover, as the homepage denotes the entrance to any website, it could be the most up-to-date section of the website and the most frequently maintained by the webmaster (Providenti & Zai III, 2011).The Next section offers a comparison of the web accessibility evaluation result across all institutions based on the automated tools employed.
Web Accessibility Results Based on AChecker
AChecker evaluated the web accessibility based on two primary guidelines, namely WCAG 2.0 and Section 508.As Table 3 indicates, none of the websites examined passed the lowest accessibility test as set out in WCAG 2. 0 (Level A), except three institutions (IIUM, UMS and UTEM) that reported conditional passes.In addition, only the IIUM website reported a conditional pass for WCAG 2.0 Level AA and Level AAA requirements respectively.The assessment based on Section 508 guidelines showed that none of the websites passed the test.The high number of errors reported for each level of test is partly because AChecker functions on the most recent standard i.e.WCAG 2.0 and Section 508 that are highly extensive (Adepoju & Shehu, 2014).
Overall, the results provide evidence that the level of compliance with respect to web accessibility requirements is relatively very low amongst the public universities in Malaysia.The results seem consistent with what has been reported by similar studies worldwide such as in the Middle East countries (Shawar, 2015), Nigeria (Adepoju & shehu, 2014), Argentina (Laitano, 2015) and Malaysia (Abdul Latif & Masrek, 2010;Abdul Aziz et al., 2010;Abuaddous et al., 2013).AChecker classifies accessibility-related problems into three categories.(a) known problems, i.e. the problems that have been identified with certainty as accessibility barriers, (b) likely problems, i.e. the problems that have been identified as probable barriers, but requires further subjective assessment and (c) potential problems which refer to the problems that A Checker is unable to detect, which demands for manual inspection of the web.For known problem, AChecker inspects 12 primary issues with several criteria being specified to evaluate each of the issues concerned.The Known Problems identified are crucial, thus, requiring immediate attention of the web developer.Table 4 reports the known problems as generated by AChecker for all levels of WCAG 2.0 specification.
As per Level A requirements, one of the issues that deserves further attention is keyboard accessibility with 149 incidences of errors (40 percent) for all websites.Among the concerns over keyboard accessibility is on pairing of attributes.In case the website contains an onmousedown attribute, it must also contain an onkeydown attribute, and if it contains an onmouseout attribute, it must also contain an onblur attribute.Meanwhile, if the website contains an onmouseover attribute, it must also contain an onfocus attribute; and if it contains an onmouseup attribute, it must also contain an onkeyup attribute.Adepoju & Shehu, 2014;Hackett & Parmanto, 2005;Pribeanu et al., 2012).
The presence of text alternative enables the content to be presented to the users as per its original content (Abdul Latif & Masrek, 2010).Its presence is more paramount if the non-text element complements a webpage content or it is associated to a function of the page (Providenti & Zai, 2007).Thus, this error calls for special attention by the web developer.
As for the Level AAA category, the criteria that requires serious attention are the selection of colour (distinguishable), particularly the selection of colour between the text and the background, link text and background, active link text and background or visited link text and background.The optimal contrast ratio must be at least 7:1 for easily distinguishable content.Meeting the specified contrast ratio facilitates users to easily recognise the text and to separate the foreground from the background of the web page.Application of lower colour contrast causes difficulty for certain users to view or to read the web content (Roggio, 2016).The issue of contrast colours is also reported as the second highest occurrence as per Level AA category (333 incidences or 33%).Level AA requires a minimum contrast ratio level of 5:1.
Other problems that deserve further attention include navigability, adaptability, input assistance and compatibility.A navigable website provide ways to assist users to navigate, to find content, and to determine where they are while browsing in the website.Adaptability is essential through which the web administrator ensures separation of information and structure from the presentation.Provision of input assistance helps users to avoid and correct mistakes, while compatibility ensures the website supports current and future user agents, including assistive technologies.The issues marked as not applicable (n/a) represent the issues not being classified as known problems.
Instead, AChecker classifies these issues as likely problems or potential problems.
By referring to Section 508 requirements, AChecker reports three issues out of 16 standards listed in Table 5.Similar to Bakhsh and Mehmood's (2012) finding, the first issue that requires further attention is the use of script with 240 occurrences of errors (61% of total errors) reported for non-compliance of this standard.The standard requires every script elements that occurs within the body must be followed by a noscript section.The second issue is about the text equivalent (or text alternative as per previous findings under WCAG 2.0 Level AA,) in which all images must have an alternate text (34% of total reported errors).People who are unable to visually view the image depend highly on this feature.Additionally,the image element should not have an alternate attribute value of null or whitespace if the image element is contained by an A element and there is no other link text.If the image is used as a link, then it must provide an alternate text that describes the link destination.
Similar to the findings of related studies (Adepoju & Shehu, 2014;Providenti & Zai, 2007), the third issue that requires the web administrator's attention is accessible form.Select element or input element that contains a type attribute value of text, image, password, checkbox, file or radio must have an associated label element.Thus, web administrators must add a label element that surrounds the control's label.They must set the for attribute on the label element to the same value as the id attribute of the control and/or add a title attribute to the input element and/or create a label element that contains the input element.
Web Accessibility Errors
The overall results revealed that almost all the websites reported at least one error except IIUM, UPNM and UTeM that passed the web accessibility test.Nine of the websites reported between one and 10 errors, two of the websites reported between 11 and 20 errors while the remaining two websites contained 33 and 87 errors respectively.Table 7 reveals the summary of errors for all the websites.Table 8 reports the details of the errors detected by WAVE that requires the web administrator's immediate attention.The most commonly found error amongst the websites was an empty link, i.e. the link provided in the website contains no text.This corresponds to several other studies (Adepoju & Shehu, 2014;Shawar, 2015).To rectify this error, the webmaster should provide the text within the link that describes the functionality and/or target of the link.
The second highest incidence of errors is related to missing alternative text for images.Basically, an alternative text provides a textual alternative to non-text content in the website and it becomes a barrier to accessibility especially for screen-reader users (Smith, 2007).The third error with the highest occurrences refers to the empty heading, in which the heading contains no content.The web administrator must ensure that all headings contain informative content.According to WAVE, some users, especially keyboard and screen-reader users, often navigate the web based on the heading elements.Hence, an empty heading will produce no information and may cause confusion to the users.Table 8 presents the remaining errors reported by WAVE together with the explanation of its purpose and importance.
CONCLUSION, LIMITATIONS AND FUTURE RESEARCH
In line with the increased reliance on the website to disseminate and communicate relevant information to various stakeholders, public universities in Malaysia should ensure their websites are accessible to all users regardless of their abilities and specific requirements.In response, this paper examined and reports the current status of web accessibility amongst 20 Malaysian public universities.This study employed two automated tools available online, i.e.AChecker and WAVE.Overall, this study indicated a relatively low level of conformance to the web accessibility standards/guidelines.Most of the websites examined did not even pass the minimum web accessibility requirement as stipulated in WCAG 2.0 (Level A).Certain aspects of the website design demand the immediate attention of webmasters to further enhance the accessibility of the website, namely provision of text alternatives for any non-text content, keyboard accessibility and colour contrast.Other issues such as such as navigation, adaptability, input assistance, compatibility, empty link, and empty heading also deserve further attention to enhance web accessibility.Nevertheless, some of the accessibility features and structural elements were being extensively applied by most of the websites.This study provides meaningful insights particularly to the web developers for continuous improvement of their respective university websites for better compliance with the established web accessibility standards.Continuous enhancement of the web accessibility features a better position of the institution based on website ranking tools such as webometric.
Despite the insightful findings reported, interpretation of the results requires careful consideration on the limitations of the study.First, the results of the present study may not be representative of other categories of universities or organisations.Different categories of institutions or organisation may have been established with different objectives and strategic directions.
Nevertheless, the areas of concern are relatively relevant to all web developers or any organisation as part of their improvement initiatives.Secondly, this study employed WAVE and AChecker as the preferred tools to produce an objective evaluation of web accessibility.Obviously, there may be slight variations across the different tools considering the scope and guidelines adopted by each automated tool.This study also did not consider the aspects of the automated tools reports that require manual inspection or confirmation.
Future research could probably focus on the different scopes or type of organisations to promote greater awareness on web accessibility among the web developers.Alternatively, other tools also can be applied to compare the results among them or a combination of automated and manual assessments could be another interesting area to explore.
Table 2
Application of Automated Tools in Web Accessibility Research
Table 3
AChecker Results Summary Based on WCAG 2.0 and Section 508 by University Legend: R=Result, K=Known problems, L=Likely problems, P=Potential problems, F=Fail, P=Pass, C=Conditional pass
Table 4
Known Problems as per WCAG 2.0 Latif & Masrek 2010;2012) Mehmood (2012), errors reported from Level AA are the highest among the three assessment components.Level AA recommends the provision of text alternatives for any non-text content such as image, area, embed and input elements.Higher incidences of such error (35%) are in parallel with previous studies (AbdulLatif & Masrek 2010;
Table 5
Problems as per Section 508 Table6shows the summary of the web accessibility results generated by WAVE, a tool embedded in the Google Chrome browser.This tool provides visual feedback about the accessibility of the web content by injecting icons and indicators onto the page.All analysis were entirely carried out within the browser.WAVE analyses the web accessibility errors, alerts, features, structural elements, HTML5 and Accessible Rich Internet Applications (ARIA) and contrast errors based on WCAG 2.0 (Level A), WCAG 2.0 (Level AA) and Section 508.
Table 7
Web Accessibility Errors as Reported by WAVE
Table 8
WAVE Errors Details | 5,761 | 2016-11-28T00:00:00.000 | [
"Education",
"Computer Science"
] |
Effect of Multi-Pass Power Spinning on Microstructure Homogenization and Mechanical-Property Strengthening of Ti2AlNb-Based Alloy Using Welded Tube Blank
Long seamless tubes of Ti2AlNb-based alloys are difficult to manufacture through conventional forming methods. In this study, a multi-pass power spinning process was first utilized to fabricate thin-walled tube of Ti-22Al-24Nb-0.5Mo alloy using welded thick tube blank, assisted by on-line electro-magnetic induction heating to maintain high spinning temperature during the whole spinning process. After six-pass hot power spinning at 950 ± 30 °C, the microhardness difference of BM (base metal), HAZ (heat affect zone) and FZ (fusion zone) became much smaller, and the microhardness fluctuation ΔHV dropped to 32 from 122 of the as-welded joint due to the phase composition and microstructure homogenization. The grain size of B2 phase was refined to 0.4/0.6 μm from 2.7/10.8 μm of the as-received BM/FZ, respectively. Meanwhile, the B2 phase <111>B2//ND texture of the as-received rolled sheet weakened during multi-pass spinning due to recrystallization, which co-existed with <001>B2//ND texture in final pass. The ultimate tensile strength in axial/tangential direction was increased to 1245/1299 MPa from 1206/1010 MPa of the as-received rolled sheet, respectively, mainly due to the effect of fine grain strengthening. This study provides an effective way to manufacture high-performance tubular workpieces with low cost and high efficiency.
Introduction
Based on the superior properties, such as high specific strength, outstanding oxidation and splendid creep resistance [1][2][3][4], the Ti 2 AlNb-based alloys have presented amounts of brilliant performances in aviation and aerospace field, and the corresponding thin-walled tubular workpieces exhibit increasing application potential in aviation and automotive industries [5,6]. Especially, during the long-term service, the temperature of the aviation components can reach 650 • C or even 750 • C [5,6], which make the Ti 2 AlNb-based alloys good candidates for high-temperature load-bearing structural components. As typical multiple phase alloys, the Ti 2 AlNb alloys have quite a few microstructural morphologies [7][8][9][10]. Usually, the Ti 2 AlNb alloys consist of the O phase (Cmcm symmetrical structure, Ti 2 AlNb), the bcc B2 phase (ordered Pm3m symmetric structure)/β phase (disordered structure), and the hcp α 2 phase (DO19 P63/mmc symmetrical structure, Ti 3 Al) [1]. Based on the Boehlert research [11], the tensile elongation of Ti-22Al-25Nb fluctuated between 0 and 16% at room-temperature, and the yield strength could be manipulated between 650 and 1600 MPa [12,13]. Du et al. [14] reported that the tensile strength of the multilayer structure of Ti 2 AlNb-based alloy decreases from 981 MPa to 684 MPa at 650 • C, because of the precipitation of coarsened O laths during furnace cooling.
As considering the mechanical property, i.e., high strength accompanies low ductility, the as-rolled Ti 2 AlNb-based alloy sheets are usually formed into complex-shaped workpieces by hot forming at high temperature over 900 • C, and even higher than 950 • C or 970 • C [14]. Besides, superplastic forming and hot gas forming were used to produce the complex-shaped components of Ti 2 AlNb-based alloys, such as Ti-22Al-24Nb box-shaped component [15] and cup-shaped Ti-22Al-24.5Nb-0.5Mo component [16]. However, no study on hot forming of long thin-walled tubes of Ti 2 AlNb-based alloys has been reported until now. It is well known that power spinning is an outstanding process to manufacture thinwalled tubular components with high performance [17,18]. However, the large-diameter tube billet of Ti 2 AlNb-based alloys for power spinning is quite difficult to prepare through conventional extrusion and ring forging because of their inferior hot workability. The combination of hot pressing and welding from sheet plate to form Ti 2 AlNb-based alloy tubes shows high production efficiency and low cost, while the poor weldability deteriorates the strength and ductility of welded tubes of Ti 2 AlNb alloys. As Chen et al. reported, the Ti-22Al-25Nb EBW (electronic beam welding) joints exhibited no ductility at 650 • C and failed at 333-387 MPa in fusion zone (FZ) as the inter granular failure, indicating especially infirm B2 phase grain boundaries in the FZ [12]. Li et al. reported that the full-penetrated EBW weld joint of thick plate suffered about 1000 MPa tensile stress in the center of the FZ, resulting in the risk for cracking [19]. Compared with welding sheets, welding tubes need more complex clamping fixture and welding process, while the strength and ductility reduced more significantly [20]. In addition, there is no research concerning power spinning of thin-walled tubes of Ti 2 AlNb-based alloys available to date. Therefore, it is difficult to produce thin-walled Ti 2 AlNb-based tubes with satisfied shape and admirable mechanical property by way of tube welding and power spinning process.
In this study, a multi-pass hot power spinning process using welding thick tube blanks of Ti-22Al-24Nb-0.5Mo alloy was first designed and performed to form thin-walled tubular parts without cracks successfully. The microstructure evolution and mechanical properties of base metal and EBW joint during hot spinning were investigated, which exhibited a homogenized strengthening effect with the increase in thickness reduction. This work is expected to provide a feasible way to fabricate high performance thin tubular workpiece of Ti 2 AlNb alloys with low cost and high efficiency.
Material Selection and Experimental Procedure
The hot rolled Ti-22Al-24Nb-0.5Mo (at%) sheets with 6 mm in thickness were provided by IMR (Institute of metal research, Chinese Academy of Sciences). The dimension of the pre-spun welded tube was 100 mm (diameter) × 6 mm (thickness) × 80 mm (length) in this study. To make welded tubes, two pieces of as-received Ti 2 AlNb sheets were hot pressed to semicircular parts at 920 ± 30 • C, and then connected by electronic beam welding (EBW). The spun tube workpiece was formed by 6-pass hot flow spinning under the temperature 950 ± 30 • C. Subsequently, the spun workpiece was cooled in the air, followed by annealing treatment at the temperature 800 • C for 3 h, to relieve the residual stress. Figure 1 shows the whole manufacture process.
Temperature Control System
The ductility exhibited high correlation with the temperature as well as strain rate closely of the Ti 2 AlNb-based alloy, as reported by our previous research [21]. In hot spinning of Ti 2 AlNb alloy, the spinning temperature should be kept relatively stable in 900-1000 • C. In this study, the temperature control system was designed to heat tube billet and control the temperature change during hot spinning, as shown in Figure 2, which consisted of an electromagnetic induction heating device and an infrared temperature monitoring system, (FLIR Systems, Inc., Arlington, WA, USA). During hot spinning, the electromagnetic induction coil moved synchronously with the roller along axial direction to heat tube billets, 950 ± 30 • C. Before the onset of the hot spinning process, the preheating
Temperature Control System
The ductility exhibited high correlation with the temperature as well as strain r closely of the Ti2AlNb-based alloy, as reported by our previous research [21]. In h spinning of Ti2AlNb alloy, the spinning temperature should be kept relatively stable 900-1000 °C. In this study, the temperature control system was designed to heat tube b let and control the temperature change during hot spinning, as shown in Figure 2, wh consisted of an electromagnetic induction heating device and an infrared temperatu monitoring system, (FLIR Systems, Inc., Arlington, WA, USA). During hot spinning, t electromagnetic induction coil moved synchronously with the roller along axial dir tion to heat tube billets, 950 ± 30 °C. Before the onset of the hot spinning process, t preheating temperature of the mandrel and tube billets should be higher than 650 and 750 °C, respectively.
Temperature Control System
The ductility exhibited high correlation with the temperature as well as strain rate closely of the Ti2AlNb-based alloy, as reported by our previous research [21]. In hot spinning of Ti2AlNb alloy, the spinning temperature should be kept relatively stable in 900-1000 °C. In this study, the temperature control system was designed to heat tube billet and control the temperature change during hot spinning, as shown in Figure 2, which consisted of an electromagnetic induction heating device and an infrared temperature monitoring system, (FLIR Systems, Inc., Arlington, WA, USA). During hot spinning, the electromagnetic induction coil moved synchronously with the roller along axial direction to heat tube billets, 950 ± 30 °C. Before the onset of the hot spinning process, the preheating temperature of the mandrel and tube billets should be higher than 650 °C and 750 °C, respectively.
Spinning Process Scheme
The spinning experiments were performed on a CNC spinning machine with two rollers symmetrically distributed around the mandrel. Backward flow spinning method was adopted in this study. During tube spinning, the thinning ratio per pass ψ i was chose as 20%. The total thickness reduction (ψ f = (t 0 − t f )/t 0 , where t 0 was the initial wall thickness of the tubular blank and t f was the final wall thickness of the spun parts), was chosen as Table 1 listed the main process parameters of flow spinning used in the spinning experiment. The strain rate in the present hot power spinning experiment was 0.2 s −1 according to the following equation [22]. .
where v 0 was feed speed, α ρ was front angle of roller, ψ t was reduction, t 0 was initial tube blank thickness, n was rotation speed of mandrel, f was roller feed rate.
Mechanical-Property Evaluation and Microstructures Characterization
To analyze the mechanical properties of spun tube, tensile specimens with gauge length of 4 mm, thickness of 1.0 mm and width of 1.2 mm were cut along axial and tangential direction from the spun tubes, respectively, see Figure 3. The room temperature tensile test with a strain rate of 1 × 10 −3 s −1 was carried out on an Instron 5569 electronic testing machine, which was repeated five times. Both the axial and tangential tensile tests were carried out on base metal of the multi-pass spun workpiece, see Figure 3. The EBSD analysis was performed on a scanning electron microscope (SEM, Quanta 200FEG) at a step size of 0.1 µm with voltage 30 kV, and post processed with the TSL OIM Analysis 6.1.3 software. The electropolished solution of EBSD samples was perchloric acid, methyl alcohol and butanol (ratio, 6:60:34). The polishing temperature is −20 • C, and the polishing time is 120 s. The microhardness test was carried out on a HVS-1000A microhardness tester (Laizhou Huayin Testing Instrument Co., Ltd., Laizhou, China). A Loading force 1000 g with a dwell time of 15 s was chosen based on ASTM standard E384-99. The TEM observation was carried out on a Talos F200x transmission electron microscope (FEI Company, Hillsborough, OR, USA). The TEM samples were mechanically thinned to 70 µm, and then perforated by ion milling.
Based on SEM, TEM results, the width and length as well as various phase volume fractions were determined (repeated 5 times) by means of the software of Image-Pro Plus, which has often been used to calculate mean values of complicated microstructure [23][24][25].
Initial Microstructures
The cross-sectional macroscopic morphology of the as-weld joint was presented Figure 4a. The microstructure of the base metal (BM), B2 + O + α2 phases, was depic in Figure 4b. The area fraction of dark α2, grey lenticular O+ rim-O, and the light phase was estimated to be 13.2%, 31.4%, 55.4%, respectively, as shown in Table 2. From BM towards fusion zone (FZ), the area fraction of prior equiaxed α2 phase creased, accompanying with the transformation O→B2. Boehlert et al. [26] systematica analyzed the phase diagram of the Ti-22Al-xNb alloy, indicating that α2-Ti3Al phase p ferred to nucleate earlier than the fine plate-like O phase as the temperature decreas during furnace cooling. Hence, the heat affected zone (HAZ) and BM mainly consis of B2 + O + α2 phases [27].
About 95.5% B2 phase and 4.5% α2 phase could be found in FZ due to the huge h input and the material remelting [28]. As previous researches reported, the phase co position of Ti-22Al-25Nb (at%) alloy was closely related with the cooling rate. If the co
Initial Microstructures
The cross-sectional macroscopic morphology of the as-weld joint was presented in Figure 4a. The microstructure of the base metal (BM), B2 + O + α 2 phases, was depicted in Figure 4b. The area fraction of dark α 2 , grey lenticular O + rim-O, and the light B2 phase was estimated to be 13.2%, 31.4%, 55.4%, respectively, as shown in Table 2. Plus, which has often been used to calculate mean values of complicated microstructure [23][24][25].
Initial Microstructures
The cross-sectional macroscopic morphology of the as-weld joint was presented in Figure 4a. The microstructure of the base metal (BM), B2 + O + α2 phases, was depicted in Figure 4b. The area fraction of dark α2, grey lenticular O+ rim-O, and the light B2 phase was estimated to be 13.2%, 31.4%, 55.4%, respectively, as shown in Table 2. From BM towards fusion zone (FZ), the area fraction of prior equiaxed α2 phase decreased, accompanying with the transformation O→B2. Boehlert et al. [26] systematically analyzed the phase diagram of the Ti-22Al-xNb alloy, indicating that α2-Ti3Al phase preferred to nucleate earlier than the fine plate-like O phase as the temperature decreased during furnace cooling. Hence, the heat affected zone (HAZ) and BM mainly consisted of B2 + O + α2 phases [27].
About 95.5% B2 phase and 4.5% α2 phase could be found in FZ due to the huge heat input and the material remelting [28]. As previous researches reported, the phase composition of Ti-22Al-25Nb (at%) alloy was closely related with the cooling rate. If the cool- From BM towards fusion zone (FZ), the area fraction of prior equiaxed α 2 phase decreased, accompanying with the transformation O → B2. Boehlert et al. [26] systematically analyzed the phase diagram of the Ti-22Al-xNb alloy, indicating that α 2 -Ti 3 Al phase preferred to nucleate earlier than the fine plate-like O phase as the temperature decreased during furnace cooling. Hence, the heat affected zone (HAZ) and BM mainly consisted of B2 + O + α 2 phases [27].
About 95.5% B2 phase and 4.5% α 2 phase could be found in FZ due to the huge heat input and the material remelting [28]. As previous researches reported, the phase composition of Ti-22Al-25Nb (at%) alloy was closely related with the cooling rate. If the cooling rate increased higher than 120 K/s, the single B2 phase microstructure tended to occur [29,30]. Boehlert et al. reported that, for Ti-23Al-27Nb (at%) alloy, single B2 phase could be obtained by the way of 1090 • C solution heat-treated and subsequent quench treatment [26]. In addition, due to higher content of Nb element, the transformation of β → a 2 was impeded effectively since the β phase stabilization were enhanced and higher cooling rate was needed. Thus, the FZ microstructure of Ti-24Al-17Nb (at%) alloy laser-welded-joint was primarily composed of B2 phase [31]. Figure 5 presents the microstructure evolution of spun tube. The phase fractions of the post-welded spun joints were summarized in Table 2. The lenticular O phase in BM was elongated gradually during 1st to 3rd pass spinning and then dynamically recrystallized during 4th to 6th pass spinning. After 6th pass spinning, the lenticular O phase transformed into global morphology with average diameter~0.8 µm. The content of O phase fluctuated between 31.4% and 43.8%, and α 2 phase varied between 13.2% and 12.1% in BM during the whole spinning process.
Microstructure Evolution during Multi-Pass Power Spinning
The microstructure of FZ changed obviously during multi-pass hot spinning, see Figure 5. After 1st and 2nd pass spinning, the content of O phase in FZ reached 47.1% and 47.6%, respectively, and that of α 2 phase reached 10.3% and 10.7%, respectively, attributed to B2 → B2 + O + α 2 . After 3rd pass spinning, the equiaxed O and α 2 phases occurred, while the content of O and α 2 phases only changed little during 3rd to 6th pass spinning.
Different from FZ, the abnormal coarsened a 2 phase could be seen in HAZ, see Figure 5a-f. Besides, the α 2 phase content in HAZ changed slightly from 12.8% to 12.1% during the whole process of 6-pass spinning. As Boehlert et al. [26] mentioned, the α 2 phase occurred in α 2 + B2 or α 2 + B2 + O phase region was difficult to be absorbed or transformed completely during the subsequent aging treatment at lower temperature. The rim-O phase was partly accountable over this phenomenon, which impeded the Nb element diffusion between α 2 and bcc phases [26]. Besides, after 6th pass spinning, the O phase content in HAZ reached 45.7%, which was much greater than that of the as-welded HAZ (5.7%) due to the transformation of B2 → B2 + O + α 2 at 950 ± 30 • C. Similar to FZ, when the spinning process increased over 3 passes (ψ t = 49%), the globalization of O phase in HAZ became quite evident with mean diameter~0.6 µm after 6th pass spinning. The phase composition and phase content among FZ, HAZ and BM converged gradually with the increase in thinning ratio, as listed in Table 2, due to microstructure homogenization of the post-welded spun workpiece.
Microstructure Evolution of BM
For further revealing the deformation mechanism in BM of the spun tube, the corresponding microstructures are characterized using EBSD and TEM in this study. Figure 6 presents the IQ + Phase maps (Image Quality + Phase) of the spun BM formed at different passes. Figure 7 shows the statistical data of B2 phase grain size as well as misorientation fraction. Actually, it is hard to discern α 2 and O phase based on EBSD result because of little difference between each other in crystal structure [32]. Thus, they were integrated as O phase, and marked in green color. During 1st to 3rd pass spinning (ψ t = 20-49%), the B2 and O phases were elongated gradually since the axial flow of metal dominated plastic deformation of tube spinning, and the high-angle boundary (HAB, misorientation > 15 • ) fraction of B2 phase was less than 63.1%, see Figure 6b-d. Beyond 4th pass spinning (ψ t = 59%), amounts of spheroidized recrystallization grains were formed, and the HAB fraction was higher than 81.3%, which was consistent with the analysis result of B2 phase grain morphology, see Figures 6e-g and 7e-g. With the increase in thinning ratio, the LAB (misorientation < 15 • ) usually evolved into HAB by absorbing dislocations. Due to severe plastic deformation and dynamic recrystallization during hot flow spinning, the B2 phase grain size dropped from 2.7 µm to 0.4 µm after six-pass spinning (ψ t = 74%), as shown in Figures 6 and 7.
For further revealing the deformation mechanism in BM of the spun tube, the corre-sponding microstructures are characterized using EBSD and TEM in this study. Figure 6 presents the IQ + Phase maps (Image Quality + Phase) of the spun BM formed at different passes. Figure 7 shows the statistical data of B2 phase grain size as well as misorientation fraction. Actually, it is hard to discern α2 and O phase based on EBSD result because of little difference between each other in crystal structure [32]. Thus, they were integrated as O phase, and marked in green color. During 1st to 3rd pass spinning (ψt = 20-49%), the B2 and O phases were elongated gradually since the axial flow of metal dominated plastic deformation of tube spinning, and the high-angle boundary (HAB, misorientation > 15°) fraction of B2 phase was less than 63.1%, see Figure 6b-d. Beyond 4th pass spinning (ψt = 59%), amounts of spheroidized recrystallization grains were formed, and the HAB fraction was higher than 81.3%, which was consistent with the analysis result of B2 phase grain morphology, see Figures 6e-g and 7e-g. With the increase in thinning ratio, the LAB (misorientation < 15°) usually evolved into HAB by absorbing dislocations. Due to severe plastic deformation and dynamic recrystallization during hot flow spinning, the B2 phase grain size dropped from 2.7 μm to 0.4 μm after six-pass spinning (ψt = 74%), as shown in Figures 6 and 7. For the purpose of further investigating the orientation characteristics of the grains in the Ti 2 AlNb-based spun workpieces, the corresponding inverse pole figures (IPFs) of B2 and O phase were obtained based on the EBSD data, see Figures 8 and 9, respectively. The B2 phase of the as-received alloy possesses a strong <111> B2 texture in the normal direction (ND), i.e., <111> B2 //ND, see Figure 8a. The slip systems {110}<111> B2 and {112}<111> B2 , were activated when the spinning process operated. Meanwhile, <001> B2 texture formed gradually in the normal direction, i.e., <001> B2 //ND, see Figure 8b. It can be found in Table 3, all the B2 phase Schmid factors were higher than 0.4 under 0-6pass, which demonstrated that the two slip systems had a good slip deformation capacity, see Table 3. Meanwhile, as the <001> B2 texture evolved, the <100> O texture emerged simultaneously, see Figure 9c. As B. Shao et al. [1] reported that if the slip plane and direction of O phase were the same as B2 phase, it is favorable for O phase slip. After 5th pass spinning, the <001> B2 texture weakened, and the <111> B2 texture emerged again since dynamic recrystallization occurred evidently during the hot spinning process, see Figure 8f,g. For the purpose of further investigating the orientation characteristics of the grains in the Ti2AlNb-based spun workpieces, the corresponding inverse pole figures (IPFs) of B2 and O phase were obtained based on the EBSD data, see Figures 8 and 9, respectively. The B2 phase of the as-received alloy possesses a strong <111>B2 texture in the normal direction (ND), i.e., <111>B2//ND, see Figure 8a. The slip systems {110}<111>B2 and {112}<111>B2, were activated when the spinning process operated. Meanwhile, <001>B2 texture formed gradually in the normal direction, i.e., <001>B2//ND, see Figure 8b. It can be found in Table 3, all the B2 phase Schmid factors were higher than 0.4 under 0-6-pass, which demonstrated that the two slip systems had a good slip deformation capacity, see Table 3. Meanwhile, as the <001>B2 texture evolved, the <100>O texture emerged simultaneously, see Figure 9c. As B. Shao et al. [1] reported that if the slip plane and direction of For identifying the phase structure as well as morphology evolution of spun tube of Ti 2 AlNb alloy, TEM micrographs in BM of the initial welded tube, 3rd and 6th pass spun workpiece were shown in Figures 10-12, respectively. The TEM analysis indicates that the initial microstructure of initial welded tube contained B2, O and α 2 phase, which could be observed by the diffraction pattern of selected area, see Figure 10a,b. Three prominent features based on TEM micrographs were shown in three local enlarged images: splitting boundaries between O phase, see Figure 10c; dislocation tangles (DTs) in B2 phase, see Figure 10d; dislocation pile-up in O phase, see Figure 10e. In addition, the length and width of acicular O phase were about 200 nm and 30 nm, respectively, see Figure 10d.
For identifying the phase structure as well as morphology evolution of spun tube of Ti2AlNb alloy, TEM micrographs in BM of the initial welded tube, 3rd and 6th pass spun workpiece were shown in Figures 10-12, respectively. The TEM analysis indicates that the initial microstructure of initial welded tube contained B2, O and α2 phase, which could be observed by the diffraction pattern of selected area, see Figure 10a,b. Three prominent features based on TEM micrographs were shown in three local enlarged images: splitting boundaries between O phase, see Figure 10c; dislocation tangles (DTs) in B2 phase, see Figure 10d; dislocation pile-up in O phase, see Figure 10e. In addition, the length and width of acicular O phase were about 200 nm and 30 nm, respectively, see Figure 10d. Figure 11 shows the prominent deformation features (fractured O phase, Elongated B2, lamella-O, equiaxed-B2) of the 3rd pass spun tube (ψt =49%) based on TEM micrographs. First, the brittle O phase was fractured and crushed into granular shape, as pointed by blue arrows in Figure 11c. Second, plastic deformation mainly operated in the B2 phase matrix, and plenty of elongated B2 phases could be seen in Figure 11d, as pointed by purple arrows. Besides, a large number of dislocation lines (DLs) and dislocation tangles (DTs) could be found in B2 matrix, see Figure 11d,f. Third, lamella O phase were still could be found along the axial direction of spun tube since the moderate deformation degree could not be used to crashed O phase thoroughly, see Figure 11e. Last, plenty of equiaxed-B2 phase appeared due to partial dynamic recrystallization during hot spinning, see Figure 11f. Figure 14 illustrates the EBSD result of IQ + Phase maps in FZ of welded spun workpieces formed at different passes. The mean grain size of B2 phase was 10.8 μm, see Figure 14a. As the thinning ratio increased, the mean grain size decreased dramatically caused of dynamic recrystallization. Compared with BM, dynamic recrystallization took place much earlier in FZ. After 1st pass spinning, amounts of recrystallized B2 grains with diameter 2.9 μm could be found in the IQ + Phase map with the HAB fraction of 75.26 %, see Figure 14b. After 6th pass spinning, the mean grain size of B2 in FZ decreased to 0.6 μm with the HAB fraction of 83.54%, see Figure 14g. Figure 11 shows the prominent deformation features (fractured O phase, Elongated B2, lamella-O, equiaxed-B2) of the 3rd pass spun tube (ψ t =49%) based on TEM micrographs. First, the brittle O phase was fractured and crushed into granular shape, as pointed by blue arrows in Figure 11c. Second, plastic deformation mainly operated in the B2 phase matrix, and plenty of elongated B2 phases could be seen in Figure 11d, as pointed by purple arrows. Besides, a large number of dislocation lines (DLs) and dislocation tangles (DTs) could be found in B2 matrix, see Figure 11d,f. Third, lamella O phase were still could be found along the axial direction of spun tube since the moderate deformation degree could not be used to crashed O phase thoroughly, see Figure 11e. Last, plenty of equiaxed-B2 phase appeared due to partial dynamic recrystallization during hot spinning, see Figure 11f. Figure 12 presents the TEM results of the 6th pass spun tube with the thinning ratio of 74%. The ultra-fine grains of equiaxial O, α 2 and B2 phase can be seen in Figure 12a. The local enlarged figures of Figure 12a were shown in Figure 12c-e. The fine O phase with 500 nm average grain size distributed homogeneously, as estimated in Figure 12c. The red arrows and green arrows in Figure 12d,e illustrated the dislocation lines (DLs) and dislocation tangles (DTs), respectively. The DLs and DTs were found in equiaxial grain interior, and faded in grain boundaries, indicating that the dislocations of the spun microstructure were in an equilibrium state. Similar to other SPD processes [33,34], the grain size could be refined by dynamic recrystallization based on the high dislocation density accumulated by flow spinning [35]. In addition, see Figure 13a-c, accompanied with lenticular O phase disappeared, the B2 phase microstructure was gradually refined to 0.4 µm after six passes spinning due to the refinement mechanism dynamic recrystallization, which corresponding to the EBSD result of Section 3.2.1. Figure 12. TEM result of the 6th pass spun tube (ψt = 74%): (a) bright field image, (b) selected area electron diffraction of (a), (c-e) local enlarged image with three features of (a). * The position marked by * corresponds to the diffraction pattern of Figure 12b. Figure 14 illustrates the EBSD result of IQ + Phase maps in FZ of welded spun workpieces formed at different passes. The mean grain size of B2 phase was 10.8 μm, see Figure 14a. As the thinning ratio increased, the mean grain size decreased dramatically caused of dynamic recrystallization. Compared with BM, dynamic recrystallization took place much earlier in FZ. After 1st pass spinning, amounts of recrystallized B2 grains with diameter 2.9 μm could be found in the IQ + Phase map with the HAB fraction of 75.26 %, see Figure 14b. After 6th pass spinning, the mean grain size of B2 in FZ decreased to 0.6 μm with the HAB fraction of 83.54%, see Figure 14g. Figure 14 illustrates the EBSD result of IQ + Phase maps in FZ of welded spun workpieces formed at different passes. The mean grain size of B2 phase was 10.8 µm, see Figure 14a. As the thinning ratio increased, the mean grain size decreased dramatically caused of dynamic recrystallization. Compared with BM, dynamic recrystallization took place much earlier in FZ. After 1st pass spinning, amounts of recrystallized B2 grains with diameter 2.9 µm could be found in the IQ + Phase map with the HAB fraction of 75.26 %, see Figure 14b. After 6th pass spinning, the mean grain size of B2 in FZ decreased to 0.6 µm with the HAB fraction of 83.54%, see Figure 14g.
Mechanical-Property Analysis during Multi-Pass Power Spinning
The Vickers hardness values of as-welded and post-welded spun joints were shown in Figure 15, indicating the variation from BM to FZ. Clearly, the microhardness curve of as-welded joint illustrated a bimodal feature. The microhardness profile was closely related to the phase compositions [12]. After six-pass spinning, the average microhardness in FZ dropped from HV 410 to HV 390, corresponding to the phase evolution of B2→B2 + O + α2 phases. Different from FZ, a markedly decrease in microhardness was observed in HAZ, and the peak value dropped from HV 469 to HV 409. Besides, the microhardness of BM increased from HV 347 to 377, mainly due to the effect of fine grain strengthening. The difference of the microhardness of BM, HAZ and FZ tended to be minimized with the increase in thinning ratio during hot flow spinning. It can be found in Figure 14,
Mechanical-Property Analysis during Multi-Pass Power Spinning
The Vickers hardness values of as-welded and post-welded spun joints were shown in Figure 15, indicating the variation from BM to FZ. Clearly, the microhardness curve of as-welded joint illustrated a bimodal feature. The microhardness profile was closely related to the phase compositions [12]. After six-pass spinning, the average microhardness in FZ dropped from HV 410 to HV 390, corresponding to the phase evolution of B2 → B2 + O + α 2 phases. Different from FZ, a markedly decrease in microhardness was observed in HAZ, and the peak value dropped from HV 469 to HV 409. Besides, the microhardness of BM increased from HV 347 to 377, mainly due to the effect of fine grain strengthening. The difference of the microhardness of BM, HAZ and FZ tended to be minimized with the increase in thinning ratio during hot flow spinning. It can be found in Figure 14, after six-pass spinning the microhardness fluctuation ∆HV dropped to 32 from 122 of the as-welded joint, which should be ascribed to the phase composition and microstructure homogenization caused by multi-pass power spinning.
Furthermore, due to the co-effect of high spinning temperature, heat affect and complicated stress condition, the aspect ratio of the O phase platelets (AROP) reduced from 6.8 to 1.4 when the six spinning passes were finished. Jiao et al. reported that the high AROP was beneficial for improving ductility [36]. Therefore, after six-pass spinning, the total elongation in axial/tangential direction was decreased to 8.1%/7.3% from 14.3%/14.6% of the as-received alloy, respectively, mainly due to the reduction in AROP and B2 content as well as work hardening. However, the multi-pass power spinning process could homogenize the microstructure and improve the mechanical property of welded tube, indicating the feasibility for manufacturing high-performance thinned walled tubular workpieces of Ti2AlNb-based alloy in lower cost and higher efficiency. Table 4. Comparison of mechanical properties between the 6-pass spun tube in axial and tangential direction (Base metal).
BM-Axial
BM-Tangential The tensile strength of the spun workpieces of base metal in axial direction and tangential direction were listed in Table 4. After six-pass spinning, the UTS (ultimate tensile strength) in axial/tangential direction was increased to 1245/1299 MPa from 1206/1010 MPa of the as-received rolled sheet, respectively, mainly due to the effect of fine grain strengthening. Besides, another factor should not be ignored, i.e., the reduction in B2 phase content from 55.4% of as-received sheet to 44.1% of spun tube in final pass, as well as the increase in O phase content from 31.4% to 43.8% correspondingly. Shao et al. reported that the B2 phase deformation ability was much better than that of O and α 2 phases since the BCC microstructure had more slip systems [1]. On the other hand, after six-pass spinning, the anisotropy of mechanical properties was reversed, the highest UTS changed from the axial direction to the tangential direction, which should be attributed to the co-effect of texture transformation and O phase morphology evolution caused by multiple-pass spinning. Table 4. Comparison of mechanical properties between the 6-pass spun tube in axial and tangential direction (Base metal).
Funding: This research was funded by the National Natural Science Foundation of China, grant number 51775137. | 7,744.8 | 2022-01-28T00:00:00.000 | [
"Materials Science"
] |
MRET: Multi-resolution transformer for video quality assessment
No-reference video quality assessment (NR-VQA) for user generated content (UGC) is crucial for understanding and improving visual experience. Unlike video recognition tasks, VQA tasks are sensitive to changes in input resolution. Since large amounts of UGC videos nowadays are 720p or above, the fixed and relatively small input used in conventional NR-VQA methods results in missing high-frequency details for many videos. In this paper, we propose a novel Transformer-based NR-VQA framework that preserves the high-resolution quality information. With the multi-resolution input representation and a novel multi-resolution patch sampling mechanism, our method enables a comprehensive view of both the global video composition and local high-resolution details. The proposed approach can effectively aggregate quality information across different granularities in spatial and temporal dimensions, making the model robust to input resolution variations. Our method achieves state-of-the-art performance on large-scale UGC VQA datasets LSVQ and LSVQ-1080p, and on KoNViD-1k and LIVE-VQC without fine-tuning.
INTRODUCTION
Video quality assessment (VQA) has been an important research topic in the past years for understanding and improving perceptual quality of videos.Conventional VQA methods mainly focus on full reference (FR) scenarios where distorted videos are compared against their corresponding pristine reference.In recent years, there has been an explosion of user generated content (UGC) videos on social media platforms such as Facebook, Instagram, YouTube, and TikTok.For most UGC videos, the high-quality pristine reference is inaccessible.This results in a growing demand for no-reference (NR) VQA models, which can be used for ranking, recommending and optimizing UGC videos.
Many NR-VQA models (Ying et al., 2021;Li et al., 2019;Tu et al., 2021;You and Korhonen, 2019;Wang et al., 2021) have achieved significant success by leveraging the power of deep-learning.Most existing deep-learning approaches use convolutional neural networks (CNNs) to extract frozen frame-level features and then aggregate them in the temporal domain to predict the video quality.Since frozen frame-level features are not optimized for capturing spatial-temporal distortions, this could be insufficient to catch diverse spatial or temporal impairments in UGC videos.Moreover, predicting UGC video quality often involves long-range spatial-temporal dependencies, such as fast-moving objects or rapid zoom-in views.Since convolutional kernels in CNNs are specifically designed for capturing short-range spatial-temporal information, they cannot capture dependencies that extend beyond the receptive field (Bertasius et al., 2021).This limits CNN models' ability to model complex spatial-temporal dependencies in UGC VQA tasks, and therefore it may not be the best choice to effectively aggregate complex quality information in diverse UGC videos.
Recently, architectures based on Transformer (Vaswani et al., 2017) have been proven to be successful for various vision tasks (Arnab et al., 2021;Dosovitskiy et al., 2021;Carion et al., 2020;Chen et al., 2021), including image quality assessment (Ke et al., 2021).Unlike CNN models that are constrained by limited receptive fields, Transformers utilize the multi-head self-attention operation which attends over all elements in the input sequence.As a result, Transformers can capture both local and global long-range dependencies by directly comparing video quality features at all space-time locations.This inspires us to apply Transformer on VQA in order to effectively model the complex space-time distortions in UGC videos.
Despite the benefits of Transformers, directly applying Transformers on VQA is challenging because VQA tasks are resolution-sensitive.Video recognition models like ViViT (Arnab et al., 2021) use fixed and relatively small input size, e.g., 224×224.This is problematic for VQA since UGC videos with resolution smaller than 224 are very rare nowadays (less than 1% in LSVQ (Ying et al., 2021)).Such downsampling leads to missing high-frequency details for many videos.As shown in Figure 1, some visible artifacts in the high resolution video are not obvious when the video is downsampled.Human perceived video quality is affected by both the global video composition, e.g., content, video structure and smoothness and local details, e.g., texture and distortion artifacts.But it is hard to capture both global and local quality information when using fixed resolution inputs.Similarly for image quality assessment, Ke et al. (2021) showed the benefit of applying the Transformer architecture on the image at the original resolution.Although processing the original high-resolution input is affordable for a single image, it is computationally infeasible for videos, due to Transformer's quadratic memory and time complexity.
To enable high-resolution views in video Transformers for a more effective VQA model, we propose to leverage the complementary nature of low and high resolution frames.We use the low-resolution frames for a complete global composition view, and sample spatially aligned patches from the high-resolution frames to complement the high-frequency local details.The proposed Multi-REsolution Transformer (MRET) can therefore efficiently extract and encode the multi-scale quality information from the input video.This enables more effective aggregation of both global composition and local details of the video to better predict the perceptual video quality.
As illustrated in Figure 2, we first group the neighboring frames to build a multi-resolution representation composed of lower-resolution frames and higher-resolution frames.We then introduce a novel and effective multi-resolution patch sampling mechanism to sample spatially aligned patches from the multi-resolution frame input.These multi-resolution patches capture both the global view and local details at the same location, and they serve as the multi-resolution input for the video Transformer.In addition to preserving high-resolution details, our proposed MRET model also aligns the input videos at different resolutions, making the model more robust to resolution variations.After the multi-resolution tokens are extracted, a factorized spatial and temporal encoder is employed to efficiently process the large number of spatialtemporal tokens.
The major contributions of this paper are summarized into three folds: • We propose a multi-resolution Transformer for video quality assessment (MRET), which makes it possible to preserve high-resolution quality information for UGC VQA.
• We propose a novel multi-resolution patch sampling mechanism, enabling the Transformer to efficiently process both global composition information and local high-resolution details.
Video Quality Assessment: Video quality assessment aims to quantify video quality.FR-VQA methods measure quality changes from pristine videos, and NR-VQA methods measure video quality without a pristine reference.For UGC videos that lack high-quality pristine reference, NR-VQA metrics are more applicable.Conventional NR metrics (Saad et al., 2014;Mittal et al., 2015;Li et al., 2016;Sinno and Bovik, 2019;Dendi and Channappayya, 2020;Korhonen, 2019;Tu et al., 2021) utilize distortion-specific features and low-level features like natural scene statistics (NSS).These feature-based NR-VQA methods mainly rely on hand-crafted statistical features summarized from limited data and are harder to generalize to diversified UGC videos.In the past few years, CNN-based NR metrics (Ying et al., 2021;Li et al., 2019;You and Korhonen, 2019;Wang et al., 2021) achieve great success in VQA using features extracted with CNNs.The features are then aggregated temporally with pooling layers or recurrent units like LSTM.The PVQ (Ying et al., 2021) method learns to model the relationship between local video patches and the global original UGC video.It shows that exploiting both global and local information can be beneficial for VQA.
Recent CNN-Transformer hybrid methods (You, 2021;Tan et al., 2021;Li et al., 2021;Jiang et al., 2021) show the benefit of using Transformer for temporal aggregation on CNN-based frame-level features.Since all these methods use CNN for spatial feature extraction, they suffer from CNN's limitation, i.e., a relatively small spatial receptive field.Moreover, these frame-level features are usually extracted from either fixed size inputs or a frozen backbone without VQA optimization.Our method is a pure Transformer-based VQA model and can be optimized end-to-end.Unlike models that use fixed small input, our proposed MRET model enables high-resolution inputs.The proposed multi-resolution input representation allows the model to have a full spatial receptive field across multiple scales.
Vision Transformers: The Transformer (Vaswani et al., 2017) architecture was first proposed for NLP tasks and has recently been adopted for various computer vision tasks (Arnab et al., 2021;Dosovitskiy et al., 2021;Carion et al., 2020;Chen et al., 2021;Ke et al., 2021).The Vision Transformer (ViT) (Dosovitskiy et al., 2021) first proposes to classify an image by treating it as a sequence of patches.This seminal work has inspired subsequent research to adopt Transformer-based architectures for other vision tasks.For video recognition, ViViT (Arnab et al., 2021) examines four designs of spatial and temporal attention for the pretrained ViT model.TimeSformer (Bertasius et al., 2021) studies five different space-time attention methods and shows that a factorized space-time attention provides better speed-accuracy tradeoff.Video Swin Transformer (Liu et al., 2022) extends the local attention computation of Swin Transformer (Liu et al., 2021) to temporal dimension, and it achieves state-of-the-art accuracy on a broad range of video recognition benchmarks such as Kinetics-400 (Kay et al., 2017) and Kinetics-600 (Kay et al., 2017).Since video recognition tasks are less sensitive to input resolution than VQA, most of the video Transformers proposed for video recognition tasks use relatively small resolution and fixed square input, e.g.224 × 224.
The objective for the VQA task is sensitive to both global composition and local details, and it motivates us to enable video Transformers to process frames in a multi-resolution manner, capturing both global and local quality information.
Overall Architecture
Understanding the quality of UGC videos is hard because they are captured under very different conditions like unstable cameras, imperfect camera lens, varying resolutions and frame rates, different algorithms and parameters for processing and compression.As a result, UGC videos usually contain a mixture of spatial and temporal distortions.Moreover, the way viewers perceive the content and distortions also impact the perceptual quality of the video.Sometimes transient distortions such as sudden glitches and defocusing can significantly impact the overall perceived quality, which makes the problem even more complicated.As a result, both global video composition and local details are important for accessing the quality of UGC videos.
To capture video quality at different granularities, we propose a multi-resolution Transformer (MRET) for VQA which embeds video clips as multi-resolution patch tokens as shown in Figure 3. MRET is comprised of two major parts, namely 1) a multi-resolution video embedding module (Section 3.2), and 2) a space-time factorized Transformer encoding module (Section 3.3).
The multi-resolution video embedding module aims to encode the multi-scale quality information in the video, capturing both global video composition from lower resolution frames, and local details from higher resolution frames.The space-time factorized Transformer encoding module aggregates the spatial and temporal quality from the multi-scale embedding input.
Multi-resolution Video Representation
Since UGC videos are highly diverse, we need to design an effective multi-resolution video representation for capturing the complex global and local quality information.To achieve that, we first transform the input video into groups of multi-resolution frames.As shown in Figure 3, the input frames are divided into groups of N .N is the number of scales in the multi-resolution input.We then resize the N frames into a pyramid of low-resolution and high-resolution frames.We preserve the aspect ratios of the frames during resizing, and we control the shorter-side length for each frame (Figure 4).Assuming the shorter-side length for the largest resolution is L, the resulting pyramid of frames will have shorter-side length L, ..., 2L N , L N accordingly.As a result, we will have a pyramid of N frames, scaling from 1× to 1 N × resolution.After obtaining the multi-resolution frames, we need a way to effectively and efficiently encode them as input tokens to the Transformer.Although low-resolution frames can be processed efficiently, processing the high-resolution frames in its entirety can be computationally expensive.For the higher-resolution frames, we propose to sample patches instead to save computation.Intuitively, the lower-resolution frames provide global views of the video composition, while the higher-resolution ones provide complementary local details.We want a patch sampling method that can best utilize the complementary nature of these multi-scale views.To achieve that, we propose to sample spatially aligned grids of patches from the grouped multi-resolution frames.In short, we use the lowest resolution frame for a complete global view, and we sample local patches at the same location from the higher-resolution frames to provide the multi-scale local details.Since the patches are spatially aligned, the Transformer has access to both the global view and local details at the same location.This allows it to better utilize the complementary multi-scale information for learning video quality.
Figure 4 and Figure 5 demonstrate how we sample spatially aligned grids of patches.Firstly, we choose a frame center, as shown by the red triangle in Figure 4.During training, the frame center is chosen randomly along the middle line for the longer-length side.For inference, we use the center of the video input.After aligning the frames, we then sample center-aligned patches from the frames.P is the patch size.For the smallest frame, we continuously sample the grid of patches to capture the complete global view.For larger frames, we sample linearly spaced-out patches to provide multi-scale local details.The center for the patches remain aligned at the same location, as shown by the yellow triangles in Figure 4.For the i-th frame (i = 1, ..., N ), the distance between patches can be calculated as (N − i) × P .Since the patches are center-aligned, they form a "tube" of multi-resolution patches for the same location.As a result, those multi-resolution patches provide a gradual "zoom-out" view, capturing both the local details and global view at the same location.
As shown in Figure 5, we then linearly project each tube of multi-resolution patch x i to a 1D token z i ∈ R d using learned matrix E where d is the dimension of the Transformer input tokens.This can be implemented using a 3D convolution with kernel size N × P × P .Each embedded token contains multi-resolution patches at the same location, allowing the model to utilize both global and local spatial quality information.Moreover, the multi-scale patches also fuse local spatio-temporal information together during tokenization.Therefore, it provides a comprehensive representation for the input video.
Factorized Spatial Temporal Transformer
As shown in Figure 3, after extracting the multi-resolution frame embedding, we apply a factorization of spatial and temporal Transformer encoders in series to efficiently encode the space-time quality information.Firstly, the spatial Transformer encoder takes the tokens from each frame group to produce a latent representation per frame group.It serves as the representation at this time step.Secondly, the temporal Transformer encoder models temporal interaction by aggregating the information across time steps.
Spatial Transformer Encoder
The spatial Transformer encoder aggregates the multi-resolution patches extracted from the entire frame group to a representation h t ∈ R d at its time step where t = 1, ..., T is the temporal index for the frame group.T is the number of frame groups.As mentioned in the previous section, for multi-resolution patches x i from each frame group, we project it to a sequence of multi-resolution tokens as z i ∈ R d , i = 1, ..., M using learnable matrix E where M is the total number of patches.We follow the standard approach of prepending an extra learnable "classification token" (z cls ∈ R d ) (Devlin et al., 2019;Dosovitskiy et al., 2021) and use its representation at the final encoder layer as the final spatial representation for the frame group.Additionally, a learnable spatial positional embedding p ∈ R M ×d is added element-wisely to the input tokens z i to encode spatial position.The tokens are passed through a Transformer encoder with K layers.Each layer k consists of multi-head self-attention (MSA), layer normalization (LN), and multilayer perceptron (MLP) blocks.The spatial Transformer encoder is formulated as: (1)
Temporal Transformer Encoder
The temporal Transformer encoder models the interactions between tokens from different time steps.We use the z cls token position output from the spatial Transformer encoder as the frame group level representation.As shown in Figure 3, each group of frames will be encoded as a single token h t , t = 1, ..., T .We then prepend a h cls ∈ R d token and add a separate learnable temporal positional embedding p t ∈ R T ×d to the temporal tokens.These tokens are then fed to the temporal Transformer encoder, which models the temporal interactions across time.The output at the h cls token is used as the final representation for the whole video.Similarly, the temporal Transformer encoder can be formulated as: Q is the number of layers for the temporal Transformer encoder.v is output from the h cls token position from the temporal encoder, which is used as the final video representation.
Video Quality Prediction
To predict the final quality score, we add an MLP layer on top of the final video representation v.The output of the MLP layer is regressed to the video mean opinion score (MOS) label associated with each video in VQA datasets.The model is trained end-to-end with L 2 loss.
Initialization from Pretrained Models
Vision Transformers have been shown to be only effective when trained on large-scale datasets (Dosovitskiy et al., 2021;Arnab et al., 2021) as they lack the inductive biases of 2D image structures, which needs to be imposed during pretraining.However, existing video quality datasets are several magnitudes smaller than large-scale image classification datasets, such as ILSVRC-2012 ImageNet (Russakovsky et al., 2015) (we refer to it as ImageNet in what follows) and ImageNet-21k (Deng et al., 2009).As a result, training Transformer models from scratch using VQA datasets is extremely challenging and impractical.We therefore also choose to initialize the Transformer backbone from pretrained image models.
Unlike the 3D video input, the image Transformer models only need 2D projection for the input data.To initialize the 3D convolutional filter E from 2D filters E image in pretrained image models, we adopt the "central frame initialization strategy" used in ViViT (Arnab et al., 2021).In short, E is initialized with zeros along all temporal positions, except at the center N/2 .The initialization of E from pretrained image model can therefore be formulated as:
scores in [0.0, 5.0] and 960p fixed resolution.LIVE-VQC contains 585 videos with MOS scores in [0.0, 100.0] and video resolution from 240p to 1080p.We use KoNViD-1k and LIVE-VQC for evaluating the generalization ability of our model without fine-tuning.Since no training is involved, we use the entire dataset for evaluation.
Implementation Details
We set the number of multi-resolution frames in each group to N = 4.The shorter-side length L is set to 896 for the largest frame in the frame group.Correspondingly, the group of frames are rescaled with shorter-side length 896, 672, 448, and 224.We use patch size P = 16 when generating the multi-resolution frame patches.For each frame, we sample a 14 × 14 grid of patches.Unless otherwise specified, the input to our network is a video clip of 128 frames uniformly sampled from the video.
The hidden dimension for Transformer input tokens is set to d = 768.For the spatial Transformer, we use the ViT-Base (Dosovitskiy et al., 2021) model ( 12Transformer layers with 12 heads and 3072 MLP size), and we initialize it from the checkpoint trained on ImageNet-21K (Deng et al., 2009).For the temporal Transformer, we use 8 layers with 12 heads, and 3072 MLP size.The final model has 144M parameters and 577 GFLOPs.
We train the models with the synchronous SGD momentum optimizer, a cosine decay learning rate schedule from 0.3 and a batch size of 256 for 10 epochs in total.All the models are trained on TPUv3 hardware.Spearman rank ordered correlation (SRCC) and Pearson linear correlation (PLCC) are reported as performance metrics.
Results on LSVQ and LSVQ-1080p
Table 1 shows the results on full-size LSVQ and LSVQ-1080p datasets.Our proposed MRET outperforms other methods by large margins on both datasets.Notably, on the higher resolution test dataset LSVQ-1080p, our model is able to outperform the strongest baseline by 7.8% for PLCC (from 0.739 to 0.817).This shows that for high-resolution videos, the proposed multi-resolution Transformer is able to better aggregate local and global quality information for a more accurate video quality prediction.(Tu et al., 2021) 0.630 0.640 0.751 0.741 VSFA (Li et al., 2019) 0.734 0.772 0.784 0.794 PVQ (Ying et al., 2021) 0.770 0.807 0.791 0.795 LSCT-PHIQNet (You, 2021) 0.737 0.762 --MRET (Ours) 0.776 0.817 0.846 0.854 Table 2. Performance on KoNViD-1k and LIVE-VQC.Methods except LSCT-PHIQNet (You, 2021) in "w/o Fine-tune" group are trained on LSVQ.Blue and black numbers in bold represent the best and second best respectively.We take numbers from (Ying et al., 2021;Jiang et al., 2021;You, 2021;Tan et al., 2021;Liao et al., 2022) for the results of the reference methods.Our final method is marked in gray .Table 3. Ablation study results for multi-resolution input on LSVQ and LSVQ-1080p dataset.MRET uses multi-resolution input while "w/o Multi-resolution" uses fixed-resolution frames.Both models grouped the frames by N = 4 when encoding video frames into tokens.Blue and black numbers in bold represent the best and second best respectively on the same dataset.Our final method is marked in gray .
Performance on Cross Dataset
Since existing VQA datasets are magnitudes smaller than popular image classification datasets, VQA models are prone to overfitting.Therefore, it is of great interest to obtain a VQA model that can generalize across datasets.To verify the generalization capability of MRET, we conduct a cross-dataset evaluation where we train the model using LSVQ training set and separately eval on LIVE-VQC and KoNViD-1k without fine-tuning.As shown in Table 2, MRET is able to generalize very well to both datasets, and it performs the best among methods without fine-tuning.Moreover, its performance is even as good as the best ones that are fine-tuned on the target dataset.This demonstrates the strong generalization capability of MRET.Intuitively, the proposed multi-resolution input aligns the videos at different resolutions.Not only does it provide a more comprehensive view of the video quality, but it also makes the model more robust to resolution variations.As a result, MRET can learn to capture quality information for UGC videos under different conditions.
Spatial Temporal Quality Attention
To understand how MRET aggregates spatio-temporal information to predict the final video quality, we visualize the attention weights on spatial and temporal tokens using Attention Rollout (Abnar and Zuidema, 2020).In short, we average the attention weights of the Transformer across all heads and then recursively multiply the weight matrices of all layers.Figure 6 visualizes temporal attention for each input time step and spatial attention for selected frames.As shown by temporal attention for the video, the model is paying more attention to the second section when the duck is moving rapidly across the grass.The spatial attention also shows that the model is focusing on the main subject, i.e., duck in this case.This verifies that MRET is able to capture spatio-temporal quality information and utilize it to predict the video quality.
Effectiveness of Multi-resolution Frame Inputs
To verify the effectiveness of the proposed multi-resolution input representation, we run ablations by not using the multi-resolution input.The comparison result is shown in Table 3 as "MRET" and "w/o Multi-resolution" for with and without the multi-resolution frames respectively.For MRET, we resize the frames to [896,672,448,224] for shorter-side lengths.For the method "w/o Multi-resolution", we resize all the frames in the frame group to the same shorter-side length (224).The GFLOPs is the same for both models because the patch size and number of patches are the same.The multi-resolution frame input brings 1-2% boost in SRCC on LSVQ and 2-3% boost in SRCC on LSVQ-1080p.The gain is larger on LSVQ-1080p because the dataset contains more high-resolution videos, and therefore more quality information is lost when resized statically to a small resolution.Armed with the multi-resolution input representation, MRET is able to utilize both global information from lower-resolution frames and detailed information from higher-resolution frames.The results demonstrate that the proposed multi-resolution representation is indeed effective for capturing the complex multi-scale quality information that can be lost when using statically resized frames.Table 3 also shows that MRET performance improves with the increase of number of input frames since more temporal information is preserved.
After verifying that the multi-resolution representation is indeed more effective than fixed resolution, we also run ablations with different multi-resolution patch sampling methods (Table 4).For "Random", we first resize the frames to the 4-scale multi-resolution input, and then randomly sample the same number of patches from each resolution.For "High-res Patch on Last Frame", we use low-resolution patches for the first 3 frames (224×), and only sample high-resolution patches from the last frame (896×).MRET samples center-aligned patches from the 4-scale input, and it performs the best.This shows the proposed sampling method can more effectively utilize the complementary nature of the multi-resolution views.With the center-aligned multi-resolution patches, MRET is able to better aggregate both the global view, and the multi-resolution local details.
Number of Grouped Multi-resolution Frames N
In Table 5 we run ablations on the number of grouped frames N when building the multi-resolution video representation.The experiment is run with 60 frames instead of 128 since smaller N increases the number of input tokens for the temporal encoder and introduces high computation and memory cost.For MRET, we use multi-resolution input for the grouped frames and for "w/o Multi-resolution", we resize all the frames to the same 224 shorter-side length.For all N , using multi-resolution input is better than a fixed resolution.It further verifies the effectiveness of the proposed multi-resolution input structure.For multi-resolution input, the performance improves when increasing N from 2 to 5, but the gain becomes smaller as N grows larger.There is also a trade-off between getting higher resolution views and the loss of spatio-temporal Table 7. Ablation study results for frame sampling method.Our final method is marked in gray .
information with the increase of N , since the area ratio of sampled patches becomes smaller as resolution increases Overall, we find N = 4 to be a good balance between performance and complexity.
Pretrained Checkpoint Selection
Compared to CNNs, Transformers impose less restrictive inductive biases which broadens their representation ability.On the other hand, since Transformers lack the inductive biases of the 2D image structure, it generally needs large datasets for pretraining to learn the inductive priors.In Table 6, we try initializing the spatial Transformer encoder in MRET model with checkpoints pretrained on different image datasets, including two image classification (Class.)datasets, and one image quality assessment (IQA) dataset.ImageNet-21k is the largest and it performs the best, showing that large-scale pretraining is indeed beneficial.This conforms with the findings in previous vision Transformer works (Arnab et al., 2021;Dosovitskiy et al., 2021).LIVE-FB (Ying et al., 2020) is an IQA dataset on which PVQ (Ying et al., 2021) obtain their 2D frozen features.Since IQA is a very relevant task to VQA, pretraining on this relatively small IQA dataset leads to superior results than ImageNet.This shows that relevant task pretraining is beneficial when large-scale pretraining is not accessible.
Frame Sampling Strategy
We run ablations on the frame sampling strategy in Table 7.For our default "Uniform Sample", we sample 128 frames uniformly throughout the video.For "Front Sample", we sample the first 128 frames.
For "Center Clip" we take the center clip of 128 frames from the video.On LSVQ and LSVQ-1080p dataset, uniformly sampling the frames is the best probably because there is temporal redundancy between continuous frames and uniformly sampling the frames allows the model to see more diverse video clips.Since most of the videos in the VQA dataset are relatively short, uniformly sampling the frames is good enough to provide a comprehensive view.
CONCLUSION AND FUTURE WORK
We propose a multi-resolution Transformer (MRET) for VQA, which integrates multi-resolution views to capture both global and local quality information.By transforming the input frames to a multi-resolution representation with both low and high resolution frames, the model is able to capture video quality information at different granularities.To effectively handle the variety of resolutions in the multi-resolution input sequence, we propose a multi-resolution patch sampling mechanism.A factorization of spatial and temporal Transformers is employed to efficiently model spatial and temporal information and capture complex space-time distortions in UGC videos.Experiments on several large-scale UGC VQA datasets show that MRET can achieve state-of-the-art performance and has strong generalization capability, demonstrating the effectiveness of the proposed method.MRET is designed for VQA, and it can be extended to other scenarios where the task labels are affected by both video global composition and local details.The limitation of Transformers is that it can be computationally expensive, and thus costly to make predictions on long videos.In this paper, we focus on improving the performance of the VQA model and we leave it as future work to improve its efficiency and to lower the computation cost.One potential direction is to employ more efficient Transformer variants, such as Reformer (Kitaev et al., 2020) and Longformer (Beltagy et al., 2020) where the attention complexity has been greatly reduced.Those efficient Transformers can be adopted as a drop-in replacement for the current spatial and the temporal Transformer used in MRET.
Figure 1 .
Figure 1.Video quality is affected by both global video composition and local details.Although downsampled video frames provide the global view and are easier to process for deep-learning models, some distortions visible on the original high resolution videos may disappear when resized to a lower resolution.
Figure 2 .
Figure 2. The proposed multi-resolution Transformer (MRET) for VQA.To capture both global composition and local details of video quality, we build a multi-scale video representation with patches sampled from proportionally resized frames with different resolutions.
Figure 3 .
Figure 3. Model overview for MRET.Neighboring video frames are grouped and rescaled into a pyramid of low-resolution and high-resolution frames.Patches are sampled from the multi-resolution frames and encoded as the Transformer input tokens.The spatial Transformer encoder takes the multi-resolution tokens to produce a representation per frame group at its time step.The temporal Transformer encoder then aggregates across time steps.To predict the video quality score, we follow a common strategy in Transformers to prepend a "classification token" (z cls and h cls ) to the sequence to represent the whole sequence input and to use its output as the final representation.
Figure 4 .
Figure 4. Multi-resolution patch sampling.We first rescale the N frames to L, ..., 2L N , L N for the shorter side and uniformly sample grid of patches from the multi-resolution frames.P is the patch size.Patches are spatially aligned.The patches at the same location in the grid provide a multi-scale view for the same location.
Figure 5 .
Figure5.Multi-resolution video frames embedding.We extract center-aligned multi-resolution patches, and then linearly project the spatially aligned "tubes" of patches to 1D tokens.
Figure 6 .
Figure 6.Visualization of spatial and temporal attention from output tokens to the input.The heat-map on the top shows the spatial attention.The chart on the bottom shows the temporal attention.Higher attention values correspond to the more important video segments and spatial regions for prediction.The model is focusing on spatially and temporally more meaningful content when predicting the final video quality score.
Table 4 .
Ablation for multi-resolution patch sampling method.Our final method is marked in gray .
Table 5 .
Ablation study results for number of grouped frames N on the LSVQ-1080p dataset.MRET uses multi-resolution input while "w/o Multi-resolution" use fixed resolution frames.Models here are trained with 60 input frames instead of 128.
Table 6 .
Results for initializing MRET model from checkpoints pretrained on different image datasests.Our final method is marked in gray . | 7,145.6 | 2023-03-13T00:00:00.000 | [
"Computer Science"
] |
Lexical Simplification with Neural Ranking
We present a new Lexical Simplification approach that exploits Neural Networks to learn substitutions from the Newsela corpus - a large set of professionally produced simplifications. We extract candidate substitutions by combining the Newsela corpus with a retrofitted context-aware word embeddings model and rank them using a new neural regression model that learns rankings from annotated data. This strategy leads to the highest Accuracy, Precision and F1 scores to date in standard datasets for the task.
Introduction
In Lexical Simplification (LS), words and expressions that challenge a target audience are replaced with simpler alternatives. Early lexical simplifiers Carroll et al., 1998) combine WordNet (Fellbaum, 1998) and frequency information such as Kucera-Francis coefficients (Rudell, 1993). Modern simplifiers are more sophisticated, but most of them still adhere to the following pipeline: Complex Word Identification (CWI) to select words to simplify; Substitution Generation (SG) to produce candidate substitutions for each complex word; Substitution Selection (SS) to filter candidates that do not fit the context of the complex word; and Substitution Ranking (SR) to rank them according to their simplicity.
The most effective LS approaches exploit Machine Learning techniques. In CWI, ensembles that use large corpora and thesauri dominate the top 10 systems in the CWI task of SemEval 2016 (Paetzold and Specia, 2016d). In SG, Horn et al. (2014) extract candidates from a parallel Wikipedia and Simple Wikipedia corpus, yielding major improvements over previous approaches (Devlin, 1999;Biran et al., 2011). Glavaš anď Stajner (2015) and Paetzold and Specia (2016f) employ word embedding models to generate candidates, leading to even better results.
In SR, the state-of-the-art performance is achieved by employing supervised approaches: SVMRank (Horn et al., 2014) and Boundary Ranking (Paetzold and Specia, 2015). Supervised approaches have the caveat of requiring annotated data, but as a consequence they can adapt to the needs of a specific target audience.
Recently, (Xu et al., 2015) introduced the Newsela corpus, a new resource composed of thousands of news articles simplified by professionals. Their analysis reveals the potential use of this corpus in simplification, but thus far no simplifiers exist that exploit this resource. The scale of this corpus and the fact that it was created by professionals opens new avenues for research, including using Neural Network approaches, which have proved promising for many related problems.
Neural Networks for supervised ranking have performed well in Information Retrieval (Burges et al., 2005), Medical Risk Evaluation (Caruana et al., 1995) and Summarization (Cao et al., 2015), among other tasks, which suggests that they could be an interesting approach to SR. In the context of LS, existing work has only exploited word embeddings as features for SG, SS and SR.
In this paper, we introduce an LS approach that uses the Newsela corpus for SG and employs a new regression model for Neural Ranking in SR that addresses the task in three steps: Regression, Ordering and Confidence Check.
Hybrid Substitution Generation
Our approach combines candidate substitutions from two sources: the Newsela corpus and retrofitted context-aware word embedding models.
SG via Parallel Data
The Newsela corpus 1 (version 2016-01-29.1) contains 1,911 news articles in their original form, as well as up to 5 versions simplified by trained professionals to different reading levels. It has a total of 10,787 documents, each with a unique article identifier and a version indicator between 0 and 5, where 0 refers to the article's original form, and 5 to its simplest version.
To employ the Newsela corpus in SG, we first produce sentence alignments for all pairs of versions of a given article. To do so, we use paragraph and sentence alignment algorithms from (Paetzold and Specia, 2016g). They align paragraphs with sentences that have high TF-IDF similarity, concatenate aligned paragraphs, and finally align concatenated paragraphs at sentence-level using the TF-IDF similarity between them. Using this algorithm, we produce 550,644 sentence alignments.
We then tag sentences using the Stanford Tagger (Toutanvoa and Manning, 2000), produce word alignments using Meteor (Denkowski and Lavie, 2011), and extract candidates using a strategy similar to that of Horn et al. (2014). First we consider all aligned complex-to-simple word pairs as candidates. Then we filter them by discarding pairs which: do not share the same POS tag, have at least one non-content word, have at least one proper noun, or share the same stem. After filtering, we inflect all nouns, verbs, adjectives and adverbs to all possible variants. We then complement the candidate substitutions from the Newsela corpus using the following word embeddings model.
SG via Context-aware Word Embeddings
Paetzold and Specia (2016f) present a state-ofthe-art simplifier that generates candidates from a context-aware word embeddings model trained over a corpus composed of words concatenated with universal POS tags. We take this approach a step further by incorporating another enhancement: lexicon retrofitting. Faruqui et al. (2015) introduce an algorithm that allows for typical embeddings to be retrofitted over lexicon relations, such as synonymy, hypernymy, etc. To retrofit the context-aware models from (Paetzold and Specia, 2016f), we concatenate the words in WordNet (Fellbaum, 1998) with their universal POS tag, create a dictionary containing mappings between word-tag pairs and 1 https://newsela.com/data their synonyms, then use the algorithm described in (Faruqui et al., 2015).
We train a bag-of-words (CBOW) model (Mikolov et al., 2013b) of 1,300 dimensions with word2vec (Mikolov et al., 2013a) using a corpus of over 7 billion words that includes the SubIMDB corpus (Paetzold and Specia, 2016b), UMBC webbase 2 , News Crawl 3 , SUBTLEX (Brysbaert and New, 2009), Wikipedia and Simple Wikipedia (Kauchak, 2013). We retrofit the model over WordNet's synonym relations only. We choose this model training configuration because it has been shown to perform best for LS in a recent extensive benchmarking (Paetzold, 2016).
For each target word in the Newsela vocabulary we then generate as complementary candidate substitutions the three words in the model with the lowest cosine distances from the target word that have the same POS tag and are not a morphological variant. As demonstrated by Paetzold and Specia (2016a), in SG parallel corpora tend to yield higher Precision, but noticeably lower Recall than embedding models. We add only three candidates in order increase Recall without compromising the high Precision from the Newsela corpus.
Unsupervised Substitution Selection
We pair our generator with the Unsupervised Boundary Ranking SS approach from (Paetzold and Specia, 2016f). They learn a supervised ranking model over data gathered in unsupervised fashion. Candidates are ranked according to how well they fit the context of the target word, and a percentage of the worst ranking candidates is discarded.
For training, the approach requires a set of complex words in context along with candidate substitutions for it. To produce this data, we generate candidates for the complex words in all 929 simplification instances of the BenchLS dataset (Paetzold and Specia, 2016a) using our SG approach. The selector assigns label 1 to the complex words and 0 to all candidates, then trains the model over this data. During SS, we discard 50% of candidates with the worst rankings. We chose this proportion through experimentation. As features, we use the same described in (Paetzold and Specia, 2016f).
Neural Substitution Ranking
Our approach performs three steps: Regression, Ordering and Confidence Check.
Regression
In this step, we employ a multi-layer perceptron to determine the ranking between candidate substitutions. The network (Figure 1) takes as input a set of features from two candidates, and produces a single value that represents how much simpler candidate 1 is than candidate 2. If the value is negative, then candidate 1 is simpler than 2, if it is positive, candidate 2 is simpler than 1. Our network has three hidden layers with eight nodes each. For training we use the LexMTurk dataset (Horn et al., 2014), which contains 500 instances composed of a sentence, a target complex word and candidate substitutions ranked by simplicity. Let c 1 and c 2 be a pair of candidates from an instance, r 1 and r 2 their simplicity ranks, and Φ(c i ) a function that maps a candidate c i to a set of feature values. For each possible pair in each instance of the LexMTurk dataset we create two training instances: one with input [Φ(c 1 ) , Φ(c 2 )] and reference output r 1 − r 2 , and one with input [Φ(c 2 ) , Φ(c 1 )] and reference output r 2 − r 1 . We train our model for 500 epochs. We use the same n-gram probability features from SubIMDB used by (Paetzold and Specia, 2015). Hidden layers use the tanh activation function, and the output node uses a linear function with Mean Average Error.
Ordering
Once the model is trained, we rank candidates by simplicity. Let M (c i , c j ) be the value estimated by our model for a pair of candidates c i and c j of a generated set C. During the ordering, we calculate the final score R(c i ) of all candidates c i (Eq. 1).
Then, we simply rank all candidates based on R: the lower the score, the simpler a candidate is.
Confidence Check
Once candidates are ranked, in order to increase the reliability of our simplifier, instead of replacing the target complex word with the simplest candidate, we first compare the use of this candidate against the original word in context, which can be seen as a Confidence Check.
The target t is only replaced by the simplest candidate c if the language model probability of the trigram S j−1 j−2 t, in which S j−1 j−2 is the bigram of words preceding t in position j of sentence S, is smaller than that of trigram S j−1 j−2 c. This type of approach has been proved a reliable alternative to simply adding the target complex word to the candidate pool during ranking (Glavaš andŠtajner, 2015).
To calculate probabilities, we train a 5-gram language model over SubIMDB, since its word and n-gram frequencies have been shown to correlate with simplicity better than those from other larger corpora (Paetzold and Specia, 2016b). We henceforth refer to our LS approach (SG+SS+SR) as NNLS.
Substitution Generation Evaluation
Here we assess the performance of our SG approach in isolation (NNLS/SG), and when paired with our SS strategy (NNLS/SG+SS), as described in Sections 2 and 3. We compare them to the generators of all approaches featured in the benchmarks of Paetzold and Specia (2016a): Devlin , Biran (Biran et al., 2011), Yamamoto (Kajiwara et al., 2013), Horn (Horn et al., 2014), Glavas (Glavaš andŠtajner, 2015) and Paetzold (Paetzold and Specia, 2015;Paetzold and Specia, 2016f). These SG strategies extract candidates from WordNet, Wikipedia and Simple Wikipedia articles, Merriam dictionary, sentencealigned Wikipedia and Simple Wikipedia articles, typical word embeddings and context-aware word embeddings, respectively. They are all available in the LEXenstein framework (Paetzold and Specia, 2015).
We use two common evaluation datasets for LS: BenchLS (Paetzold and Specia, 2016a), which contains 929 instances and is annotated by English speakers from the U.S, and NNSEval (Paetzold and Specia, 2016f), which contains 239 instances and is annotated by non-native English speakers. Each instance is composed of a sentence, a target complex word, and a set of gold candidates ranked by simplicity. We use the same metrics featured in (Paetzold and Specia, 2016a), which are the well known Precision, Recall and F1. Notice that, since these datasets already provide target words deemed complex by human annotators, we do not address CWI in our evaluations.
The results in Table 1 reveal that our SG approach outperforms all others in Precision and F1 by a considerable margin, and that our SS approach leads to noticeable increases in Precision at almost no cost in Recall.
Substitution Ranking Evaluation
We also compare our Neural Ranking SR approach (NNLS/SR) to the rankers of all aforementioned lexical simplifiers. The Devlin, Biran, Yamamoto, Horn, Glavas and Paetzold rankers exploit Kucera-Francis coefficients (Rudell, 1993), hand-crafted complexity metrics, a supervised SVM ranker, rank averaging and Boundary Ranking, respectively. In this experiment we disregard the step of Confidence Check, since we aim to analyse the performance of our ranking strategy alone.
The datasets used are those introduced for the English Lexical Simplification task of SemEval 2012 (Specia et al., 2012), to which dozens of systems were submitted. The training and test sets are composed of 300 and 1,710 instances, respectively. Each instance is composed of a sentence, a target complex word, and a series of candidate substitutions ranked by simplicity. We use TRank, the official metric of the SemEval 2012 task, which measures the proportion of instances for which the candidate with the highest goldrank was ranked first, as well Pearson (p) correlation. While TRank best captures the reliability of rankers in practice, Pearson correlation shows how well the rankers capture simplicity in general. Table 2 reveals that, much like our SG approach, our Neural Ranker performs well in isolation, offering the highest scores among all strategies available.
Full Pipeline Evaluation
We then evaluate our approach in two settings: with (NNLS) and without (NNLS-C), the Confidence Check (Section 4.3). The evaluation datasets used are the same described in Section 5, and the metrics are: • Accuracy: The proportion of instances in which the target word was replaced by a gold candidate.
• Precision: The proportion of instances in which the target word was either replaced by a gold candidate or not replaced at all.
Error Analysis
In this Section we analyse NNLS to understand the sources of its errors. For that, we use PLUMBErr (Paetzold and Specia, 2016c;Shardlow, 2014), a method that assesses all steps taken by LS systems and identifies five types of errors: • 1: No error during simplification.
• 5: Replacement does not simplify the word.
Errors of type 2 are made during CWI, 3 during SG/SS, and 4 and 5 during SR. We pair ours, Devlin's, Horn's, Glavas' and Paetzold's simplifiers with two CWI approaches: one that simplifies everything (SE), and the Performance-Oriented Soft Voting approach (PV), which won the CWI task of SemEval 2016 (Paetzold and Specia, 2016e). Table 3 shows the count and proportion (in brackets) of instances in BenchLS in which each error was made. It shows that our approach correctly simplifies the largest number of problems, while making the fewest errors of type 3A and 4. However, it can be noticed that NNLS makes many errors of type 5. By analysing the output produced after each step, we found that this is caused by the inherently high Precision of our approach: by producing a smaller number of spurious candidates, our simplifier reduces the occurrences of ungrammatical and/or incoherent substitutions, but also disregards many candidates that are simpler than the target complex word. Nonetheless, this noticeably increases the number of correct simplifications made.
Conclusions
We introduced an LS approach that extracts candidate substitutions from the Newsela corpus and retrofitted context-aware word embedding models, selects them with Unsupervised Boundary Ranking, and ranks them using a new Neural Ranking strategy.
We found that: (i) our generator achieves the highest Precision and F1 scores to date, (ii) our Neural Ranking strategy leads to the top scores on the English Lexical Simplification task of Se-mEval 2012, (iii) and their combination offers the highest Precision and Accuracy scores in two standard evaluation datasets. An error analysis reveals that our LS approach makes considerably fewer grammaticality/meaning errors than former stateof-the-art simplifiers.
In future work, we aim to investigate new architectures for our Neural Ranking model, as well as to test our approach in other NLP tasks. An implementation of our Substitution Generation, Selection and Ranking approaches can be found in the LEXenstein framework 4 . | 3,645.6 | 2017-04-01T00:00:00.000 | [
"Computer Science",
"Linguistics"
] |
Adoption of Unified Communications and Collaboration from the Perspective of Diffusion of Innovation and Service-Dominant Logic: A Preliminary View
Unified Communications and Collaboration (UC&C) is a platform that has become increasingly popular in recent years and used in organizations. The service could increase flexibility, interoperability, efficiency, and productivity in managing business processes. As the number and variety of UC&C tools and services increases, many organizations have developed strategic plans and allocated budget to implement UC&C. However, the adoption of some UC&C tools is still below than expected. As a service that continues to evolve in tandem with new developments in technology, it is thus important to investigate the factors that impact upon the adoption of UC&C tools and services. Via a review of the literature, this paper will posit an initial model for the adoption of UC&C. This initial model is developed by adopting the views of two important theories which are the theory of Diffusion of Innovation (DOI) and service-dominant logic (SDL). The DOI theory is chosen since UC&C represents evolvement in the communications technology space, while SDL is chosen because of its focus on value co-creation and user-centeredness, which is an important factor that will influence adoption. The key factors influencing the adoption of UC&C embodied in the model will provide useful insights for managers in implementing UC&C in their organizations. Keywords— adoption theory; service science; Diffusion Of Innovation (DOI); Service-Dominant Logic (SDL); Unified Communication And Collaboration (UC&C).
I. INTRODUCTION
The rapid growth of the internet and communications technology (ICT) worldwide has encouraged organizations to migrate towards unified communications and collaboration (UC&C). UC&C is the technology which combines traditional communications and computer technology into a platform for more effective communications exchange. UC&C also can be described as a set of technological innovations that are incorporated into a software package [1]. In addition, the UC&C services provide the integration of real-time communication services across multiple devices and media types.
In the past few years, UC&C are becoming popular and are implemented in organizations for optimizing the efficiency of employees. UC&C has continuously evolved parallel with the improvement in technology. Nowadays more tools are becoming a part of the evolving suite of UC&C tools. A review of the literature reveals that research in UC&C can be loosely classified into two broad areas. One is the research concerning the technical or technological aspects of UC&C which is in abundance. The other stream concerns empirical studies that investigate the adoption of UC&C. In this aspect, the literature reveals that there is still limited work in this area [2]- [4]. Based on the works of Silic & Back [5] and several others who investigate the adoption and use of UC&C in organizations in order to understand better the user's decision making process in adopting UC&C, it can be concluded that there is still much work to be done in the area. Thus this paper will focus on the latter aspect by researching on the factors influencing UC&C adoption in order to ensure that the UC&C will continue to be relevant and useful for organizations. This paper's contributions will be two-fold. Firstly, the paper will show that the lens through which UC&C should be viewed through must be the lens of service innovation and secondly based on this view, we will propose a UC&C adoption model.
A. Unified Communications and Collaboration (UC&C)
With the growth of technology, communication services have also grown since it was introduced in 1880, with the advent of the telegraph. The evolution of UC&C has five stages: 1) Separate Communications; 2) Mixed Communications; 3) Unified Communications; 4) Social Communications and Collaboration; 5) Unified Communications and Integrated Collaboration. UC&C is the integration of software and hardware to provide real-time communication services [6]. The services of UC&C include voice mail, email, text message, fax, instant messaging, video, and web conferencing [7], [8]. According to Lassman and Pray [9], there are four major phases in the execution of UC&C initiative: 1) strategize and plan; 2) select and solution; 3) deploy, and 4) operate and evolve.
UC&C is perceived in various ways by different organizations. The definition of UC&C in the industry context is specified in the services that are offered by vendors [10]. It focuses more on the business process, product functionalities, and services. From the academic perspective, technically, it highlights on protocol, features, and frameworks [10]. UC&C is beneficial because it improves message exchange and problem-solving between employees; it increases the efficiency of communication; it helps develop closer relationships with client and partners; it supports recordings of meetings, reduces telephone service costs and also traveling costs [6]. The main sectors that use UC&C are the financial, education, health, manufacturing, banking, public sector, information technology, retail, logistics, pharmaceutical and business sector.
B. Existing Models for the Adoption of UC&C
This subsection explores the existing models and theories that have been used for understanding, predicting and explaining human behavior in adopting new technologies in particular UC&C technology. Various researches well research technology acceptance and adoption studies over the years. Among the well-known technology adoption models according to Olushola and Abiola [11] are technology acceptance model (TAM) [12], theory of planned behaviour (TPB) [13], unified theory of acceptance and use of technology (UTAUT) [14], Diffusion of Innovation (DOI) theory [15], and the Technology, Organizations and Environment (TOE) framework [16].
In understanding the issues that lead to or hinder UC&C adoption, a literature review and analysis was conducted. It is a systematic process that involves searching, selecting, evaluating, and synthesizing documents which were then examined and interpreted for useful information [17]. This document review is carried out to identify the underlying theories and factors that influence the adoption of UC&C. The details of the review procedure are given in Table I. A literature search with keyword UC&C or "united communications and collaboration" using the Google Scholar search engine returned 352 hits concerning various aspects of UC&C ranging from patents in UC&C implementation and deployment to UC&C use and adoption. However, in searching specifically for UC&C adoption models reveal only 9 relevant sources (see Table 2). These nine studies are then classified based on the underlying theories, frameworks or models used, unit of analysis and method of inquiry. Further analysis for each paper reveals the factors that influence UC&C adoption for each study. As can be seen in Table 2, influencing factors of adoption and use of UC&C are varied. Using a quantitative approach, a total of 296 employees from 50 different countries are surveyed by Silic & Back [5]. It was discovered that use behavior, effort expectancy, facilitating conditions, intention to use, performance expectancy and social influence are factors that influence adoption. In a similar study, Silic & Back [18] also found out that culture has an impact on technology adoption and use of UC&C. Bakker [1] discovered that organizational, inter-organizational, technical and economic perspectives are essential drivers of adoption. Other factors are given in detail in Table II. In summarizing the nine studies on UC&C adoption, it is important to note the following:
1)
No one theory is used throughout by the researchers. Silic & Back [5], taking inspiration from technology acceptance studies, employed UTAUT as the underlying theory while others used TOE, DOI, or combinations of different perspectives, and even Kairu [22] did not mention any theory used in his research. The unit of analysis also varies. Some studies focus on users' adoption of UC&C and others focus on organizational adoption. It is also interesting to note that the method used is either the case study approach or the quantitative approach or a mixture of both. The case study approach is used when the researcher is investigating adoption issues in specific companies like General Motors Bolton [2] and Nokia [22] and a small and medium company in Taiwan [6].
2)
It is observed that there are common factors, like organizational culture and, user elements (behavior, habits, attitudes, workstyles, user skills and ease of use) that are uncovered by different studies.
There is a dearth of research in UC&C adoption studies. This state of events is mentioned by Bolton [2]. The leading researcher in the area is Silic and Back [5], [18], [20] a total of three articles. However, UC&C technologies are continuing to evolve with its cloud implementation in the form of Unified Communication as a Service (UCaaS) [23], and, the importance of unified communications is also underlined by Dery et al. [24] as an initiative in digitizing the workplace in a company's digital innovation exercise. Thus we contend that UC&C adoption studies are just beginning and is essential to be carried out. The above summary of the literature points out to a need for a further investigation into UC&C adoption. In line with previous researches, it is hypothesized that this further investigation should be based on the identification of an appropriate theory or theories. In this aspect, our guiding principles (in choosing the appropriate theory) are based on two main characteristics of UC&C. The first is that it is innovative. This can be seen from the fact that it is continuing to evolve from the telegraph in the 1880s to a cloud-based service in the form of UCaaS. Thus, in agreement with Bakker [1] the diffusion of innovation (DOI) theory which describes the process flow for guiding the intention of users in adopting new technologies can be used as a basis for discussions on UC&C adoption. The other is the role of the users in the adoption process. As highlighted by 2) above, user elements are essential and central to the success of adoption. In light of the issues which concerns the users and the use of UC&C, it is understood that the service-dominant logic Vargo & Lusch [25] with its focus on value in use will prove valuable as a guiding theory in investigating the users' elements and roles in the adoption of UC&C. In a recent study by [26] and [27] have asserted that SDL theory gives preference to users in determining usage and viewing the value based on the value in use. Thus it is suggested that both servicedominant logic and DOI are used as a lens through which we view UC&C adoption.
In subsequent sections, a review of both DOI and Service Dominant Logic will be given.
C. Diffusion of Innovation (DOI)
Particular community groups adopt innovative technology through the process of diffusion. Rogers [15], via the Diffusion of Innovation (DOI) theory, pointed out that the spread of an innovation is governed by the following four factors (see Fig. 1): (1) the innovation itself which is an idea, object, or practice that is perceived as something new by individuals or units of adoption, (2) the communication channels by which the information on the innovation is spreading, (3) time, which is the rate of the diffusion of innovation or the relative speed with which members or units of adoption adopt it and, (4) the social system, which consists of individuals, organizations, or agencies sharing the same culture and potential adopters of the innovations. Fig. 1 Four factors that govern the diffusion of innovation [15] It is logical to assume that the adoption of innovations is contingent upon the adoption of the innovation at the individual level. Rogers [15] underlines five phases of an individuals' adoption process.
The five phases are knowledge, persuasion, decision, implementation, and confirmation. In the knowledge phase, individuals are vulnerable to innovation. At this stage, individual features have the most substantial influence. At the persuasion phase, an individual will form a negative or positive perception of the new technology. The third phase of the innovation-decision process, individuals will decide to take or reject the new technological innovations. The next phase, which is the implementation phase, is when the individual uses the innovation. Finally, at the confirmation phase, occurs when the individual has decided to adopt or reject innovation. This process of innovation is illustrated in Table III.
Phase Individual's actions or roles Knowledge
Occurs when an individual learns about the existence of the new technology and a clear understanding of how it works.
Persuasion
Occurs when an individual becomes interested in trying the new technology.
Decision
Occurs when an individual is involved in the activities that lead them to accept or reject the innovation. Implementation Occurs when an individual begins to use the new technology.
Confirmation
Occurs when an individual evaluates the results of an innovation; finalizes the decision to continue or discontinue using the new technology.
Rogers [15] to explain why some innovations are successful, while others never become widely accepted. For an innovation to be successful, an innovation should have five distinct characteristics which are observability, relative advantage, compatibility, trialability, and complexity as follows [1], [15] :
1)
Compatibility: The extent to which new technology is perceived as consistent with the existing value, past experiences, and needs of potential individual adopters to be absorbed into the life of adopters.
2)
Complexity: The extent to which an innovation is perceived as relatively difficult to understand and use. The complexity of using the new technology will obstruct user to adopt it.
3)
Trialability: The extent to which an innovation may be experimented with. The difficulty of using and trying on the innovation will be hampered user to adopt it.
4)
Observability: The extent to which the results of an innovation are visible to others. The visibility of innovation will influence communication among the individual's peers and personal networks and will, in turn, generate more positive or negative responses.
5)
Relative Advantage: The extent to which an innovation is perceived as being better than the idea it replaces.
The relationship between these factors and adoption is given in Fig. 2.
D. Service Science and Service Dominant Logic
Traditionally, services are differentiated from goods by contrasting the tangible characteristics of goods with the intangible characteristics of services. Classic examples of services include everything from banking, healthcare, and, logistics to hospitality, entertainment and more. Service science as a domain was introduced in 2004, by IBM [28]. It is a transdisciplinary area which was initially named service science, management, engineering, and design but was later shortened to or service science. In service science, service is defined as the application of specialized competencies (knowledge and skills) through deeds, processes, and performances for the benefit of another entity or the entity itself [29]. Vargo and Lusch [25], [30] provided the area of service science with a theoretical underpinning by the formulation of service-dominant logic. Service-dominant logic (SDL) is a logic which builds on eleven foundational premises (FPs) has been shown in Table IV. Under this logic, service is always the fundamental basis of exchange. The dominant service logic also removes the dichotomy between goods and service. In this logic, goods are just vehicles or mediums for service provision. In the foundational premises, the word service in its singular form is defined as the application of resources and competencies for the benefit of others. This means that service includes all economic activities in which individuals, organizations, and technologies work together, applying specific competencies and capabilities in order to create the most value together [31], [32]. Under this new notion, service is not something that is proposed by an organization to be consumed by its customers, it is, however, something that is co-created by all parties involved to achieve something of value. It is this notion of value co-creation that is central to the adoption of any technology, in this case, UC&C. We conjecture that the inability for the user to co-create value by using the tools in UC&C will hamper its adoption. Thus it is pertinent that the relevant foundational premises in the dominant service logic are taken into account in determining the factors that impact upon UC&C adoption. In identifying the relevant foundational premises, we mapped each of the premises to UC&C by articulating the implications from the FPs on UC&C adoption. This mapping is given in Table IV.
FPs
Specification Implications in the context of UC&C use
FP1
Service is the fundamental basis of exchange.
In utilizing UC&C tools, the actors involved will be able to provide a new service for others or will provide service to themselves in the form of improved work performance.
Improved service:
The degree to which service is improved or a new service is created from the introduced innovation.
FP3
Goods are distribution mechanisms for service provision.
UC&C suite of tools serves as distribution mechanisms for service provision. FP6 Value is co-created by multiple actors, always including the beneficiary.
The value UC&C is co-created by all actors involved.
Value co-creation capacity:
The degree to which value co-creation is enabled or allowed in the organizations. FP7 Actors cannot deliver value but can participate in the creation and offering of value propositions.
Actors cannot deliver the value of UC&C tools; they can only offer value propositions.
Efficient resource integration:
The degree of efficiency of the user in integrating all resources to extract value in use and realize the benefit from the innovation. FP9 All social and economic actors are resource integrators.
In extracting the value from UC&C use, all those involved must integrate all available resources (operand and operant). FP10 Value is always uniquely and phenomenologically determined by the beneficiary.
The value UC&C is the determined by those that benefit from them. FP11 Value co-creation is coordinated through actor-generated institutions and institutional arrangements.
The process of value co-creation in UC&C is coordinated via some institutional arrangements like policies.
Coordination Efficiency:
The degree to which the institutions or institutional arrangements coordinates value co-creation in the organizations.
III. RESULTS AND DISCUSSION
The journey that resulted in the proposal of the SDL based UC&C adoption model is best captured in Fig. 3. It started with the literature review on UC&C adoption models which was elaborated in the materials and methods section. From the literature review, it was found a lack of empirical studies in the UC&C service; there is a need for further investigation to ensure the service is positively adopted. Considering the UC&C service is evolving technology, DOI has been selected as the leading theory of the proposed model. User perspective is a preference in adopting new technology. Accordingly, SDL has been infused DOI factors. These two guiding theories are refined and synthesized to identify the elements which affect the adoption of UC&C services. In proposing the UC&C adoption model, our approach is to look at the elements proposed in the DOI model (see the elements in Fig. 2) and the elements extracted from service-dominant logic (see the last column in Table IV).
Infusing the concepts of service science into the DOI factors resulted in an SDL based definition of the elements given in bold texts in Table V.
DOI factors Definition SD concept SD infused DOI factors Relative advantage
The extent to which an innovation is perceived as being better than the idea it replaces.
Proposed value
Relative Value: The degree to which the proposed value of the innovation is perceived to have improved.
Compatibility
The extent to which new technology is perceived as consistent with the existing value, past experiences, and needs of potential individual adopters to be absorbed into the life of adopters.
Value in use Consistency: The degree to which the value in use of the innovation is consistent with the existing values of the adopter.
Complexity
The extent to which an innovation is perceived as relatively difficult to understand and use. The complexity of using the new technology will obstruct user to adopt it.
Integrability of resources:
The degree to which the user can integrate all resources in order to extract value in use of the innovation.
Trialability
The extent to which an innovation may be experimented with. The difficulty of using and trying on the innovation will be hampered user to adopt it.
Operant resource Experimentality: The degree to which the operant resource of the user has increased after trying or experimenting (with) the innovation.
Observability
The extent to which the results of an innovation are visible to others. The visibility of innovation will influence communication among the individual's peers and personal networks and will in turn generate more positive or negative responses.
service Service experience: The degree to which the resulting service from the innovation is seen or experienced by the all in the network of users.
It is observed that the definition of "Integrability of resources" given in Table V above is very similar to the definition of "resource integration efficiency" in Table IV. It is decided to use the new element, "resource integration efficiency" in our proposed SDL based UC&C adoption model. Therefore, this research proposes a model to understand the adoption of UC&C in organization better.
Individually, the relevant elements of the model are a relative value, consistency, resources integration efficiency, experimentality, service experience, improved service, value co-creation capacity, coordination efficiency. Hence the resulting elements of the proposed SDL based diffusion of innovation model are as for Table VI. Relative Value The degree to which the proposed value of the innovation is perceived to have improved.
Consistency
The degree to which the value in use of the innovation is consistent with the existing values of the adopter.
3.
Resource Integration Efficiency The degree to which the user can integrate all resources in order to extract value in use of the innovation.
Experimentality
The degree to which the operant resource of the user has increased after trying the innovation.
5.
Service Experience The degree to which the resulting service from the innovation is seen or experienced by the all in the network of users. 6 Improved Service The degree to which service is improved or a new service is created from the introduced innovation. 7.
Value Co-created Capacity The degree to which value co-creation is enabled or allowed in the organizations. 8.
Coordination Efficiency The degree to which the institutions or institutional arrangements coordinates value co-creation in the organizations.
The proposed model of this study is constructed as shown in Fig. 4. The proposed model consists of eight constructs that are extracted from the literature and the supporting theories. The researcher has been considered all five of the innovation characteristic from DOI theory, three factors from the SDL theory and security in developing new adoption model for UC&C. The eight constructs could be expected to influent the adoption of the UC&C services. The combination of these two theories will have an impact on the user's perception of the UC&C adoption.
Therefore, the hypotheses of this study are: H1: The improvement of the relative value of UC&C positively influences on the adoption of UC&C.
IV. CONCLUSIONS
In this article, a case for the development of a UC&C adoption model is put forward. The proposed model has its roots in the DOI theory and the SDL. We have articulated the reasons for choosing to combine these two theories and shown that by infusing the concepts from the dominant service logic and that of DOI, we can determine eight factors that will have an impact on the adoption of UC&C. It is acknowledged that the deliberations in this paper are conceptual and has to be further validated by a panel of experts. This paper ends with the eight statements of hypotheses that will be tested in the next step of the research. | 5,413.8 | 2018-10-28T00:00:00.000 | [
"Business",
"Economics"
] |
Performance Analysis of RF-Powered Cognitive Radio Networks with Integrated Ambient Backscatter Communications
Integrating ambient backscatter communications into RF-powered cognitive radio networks has been shown to be a promising method for achieving energy and spectrum efficient communications, which is very attractive for low-power or no-power communications. In such scenarios, a secondary user (SU) can operate in either transmission mode or backscatter mode. Specifically, an SU can directly transmit data if sufficient energy has been harvested (i.e., transmission mode). Or an SU can backscatter ambient signals to transmit data (i.e., backscatter mode). In this paper, we investigate the performance of such systems. Specifically, channel inversion power control and an energy store-and-reusemechanism for secondary users are adopted for efficient use of harvested energy. We apply stochastic geometry to analyze coverage probability and achievable rates for both primary and secondary users considering both communication modes. Analytical tractable expressions are obtained. Extensive simulations are performed and the numerical results show the validity of our analysis. Furthermore, the results indicate that the performance of secondary systems can be improved with the integration of both communication modes with only limited impact on the performance of primary systems.
Introduction
In recent years, the demand for smart systems (e.g., on-body sensing for e-Healthy) is growing fast.For such systems, the deployed sensors usually need to work continuously and transmit collected data for upper layer applications.Since most sensors have limited battery and limited spectrum resources, energy-efficient and spectrum efficient wireless communications are required.
Several techniques have been developed for achieving low-power or even no-power communications in a spectrum efficient manner, among which ambient backscatter communications and radio-frequency (RF) powered cognitive communications are two remarkable ones.In [1], authors investigate practical backscatter for on-body sensors by using the signals from Wi-Fi or Bluetooth.Such a backscatter system based on Wi-Fi is referred to as Wi-Fi backscatter [2].Be different from RF identification (RFID) which needs a dedicated signal emitter (RFID reader) [3], Wi-Fi backscatter does not need dedicated reader.However, it does not perform well in outdoor environment.While ambient backscatter communication [4] is also a type of passive communication which utilizes ambient RF signals (e.g., TV signals) to transmit data, no dedicated signal source is required, which makes nopower wireless communications possible.However, the communications are vulnerable since backscattered signals are usually weak and volatile.
Another technique, harvesting energy from ambient RF signals, has been proposed to support energy-efficient communications [5][6][7].Besides, RF-powered cognitive radio network (CRN) offers a method to utilize primary transmitter (PTs) signals as the energy source for secondary transmitters (STs).A main problem is that the transmission opportunities of RF-powered STs are limited by the harvested energy and channel availability.
In this work, we consider the integration of ambient backscatter communications into RF-powered cognitive radio networks in a similar way to [8].In this case, these two techniques could complement each other to jointly achieve the 2 Wireless Communications and Mobile Computing advantages while overcoming the individual shortcomings.Specifically, in such scenarios, the secondary transmitters can operate in two modes, that is, transmission mode and backscatter mode.In transmission mode, an ST can directly transmit data to its receiver if sufficient energy has been harvested and the channel is available.When the channel is busy, an ST can switch to backscatter mode which backscatters the ambient signals to transmit its own data.
In other words, in our cognitive radio network, the transmission of secondary transmitters falls into interweave paradigm to utilize white spaces of specific channels [9], and the needed energy is harvested from existing signals.As for backscatter communication, in some extent it falls into overlay paradigm since the backscattered signals from secondary transmitters to primary receivers can be ignored which will be stated in more detail in the following [9][10][11].As for the primary user activity model, which has a vital influence on cognitive users' performance, that is, directly determining spectrum access time of cognitive users, in this paper follows a simplified ON/OFF model [12] wherein total active (ON) duration and inactive (OFF) duration of a primary user are fixed.Besides, the details of spectrum sensing are omitted in this paper.Other complex and widely used primary user activity models and spectrum analysis can be found in [12][13][14][15].
Note that several existing works have been done for the integration of backscatter communications with cognitive radio networks or cellular networks.In [16], authors give an overview of backscatter assisted wireless powered communications and introduce a multiple access scheme in cognitive radio networks.The tradeoff analysis in RF-powered backscatter CRNs is provided in [17].However, only one single cell is considered and no performance analysis is provided.In [18], the integration of backscatter communications with heterogeneous cellular networks is proposed and analyzed.In [19], a backscatter network is analyzed by using stochastic geometry, but dedicated power beacons are deployed to support the communication, while, in [20,21], a single hybrid transmitter harvests energy from multiple ambient transmitters, transmits its own signal, or backscatters existing signals to a hybrid receiver, and its performance is analyzed.
In this paper, we investigate the performance of ambient backscatter communications in RF-powered cognitive radio network.Specifically, we propose an analytical framework based on stochastic geometry [22][23][24], with which the tractable expressions for coverage probability and achievable rates for both primary and secondary users considering both communication modes are obtained.We perform extensive simulations and the numerical results demonstrate the validity of the theoretical analysis.Also, the results indicate that secondary systems can achieve improved performance while having only limited impact on the primary systems, which show the effectiveness of integration.
The rest of the paper is organized as follows: Section 2 presents the comprehensive system model.Section 3 presents channel inversion power control and energy storage and reusing.Analytical expressions are given in Section 4. In Section 5, numerical results from analysis and simulations are described.Finally, Section 6 draws the conclusions.
System Model
2.1.Network Model.We consider a cognitive cellular network in which macro base stations (MBSs, i.e., PTs, ) serve primary cellular users (PRs, ) in the downlink while overlaid by cognitive secondary users.Each PR will connect to the nearest MBS.Besides, there are secondary transmitters (STs, ) equipped with energy storage and secondary receivers (SRs, ).An ST can communicate with an SR by either backscattering signals or emitting its own signals.Since backscattered signal is weak, the distance from backscattering node (i.e., ST) to receiving node (i.e., SR) is limited.It is shown in [4] that, for achieving 1 kbps information rate in outdoor environment, 2.5 feet is the maximum distance.Therefore, in this paper, for ease of analysis, we assume that an SR is at a constant small distance to its associated ST in an isotropic direction [19].
MBSs and PRs are modeled by homogeneous Poisson point processes (HPPP) Φ and Φ with intensities and λ , respectively.STs are uniformly distributed in annular regions with radii and centered at each MBS.Each ST can be loosely seen as a result of random and independent displacement of the MBS.In each annular region, the number of STs is ST ∼ (Λ), where (Λ) is the Poisson distribution with parameter Λ.The distribution model of STs then is similar to a Matern cluster process where Φ is the parent process [25].An inner radius > 0 of the annular region is considered for avoiding singularity of integral in the derivation process and the outer radius is related to circuit power constraint described in the following part.
We assume each PR is associated with its nearest MBS.The probability density function (pdf) of distance from a PR to its nearest MBS is [23] If a point is uniformly distributed within a circle with radius and is the distance to the center, the pdf of is () = 2/ 2 [26].By using conditional probability, we can get the pdf of distance from an ST to its MBS as Without loss of generality, according to Slivnyak's theorem [27], we analyze a typical PR ( 0 ) and a typical SR ( 0 ) located at the origin.The typical SR's corresponding ST and MBS are also similarly typical ST ( 0 ) and typical MBS ( 0 ), respectively.Figure 1 illustrates randomly generated positions of MBSs, PRs, and STs. Figure 2 shows the system model.one active PR in each channel of a cell, so active PRs in a generic channel form an HPPP Φ with intensity = by independent thinning.
Channel
We also assume that STs in each cell have equal probabilities to access a channel when it is idle, and no two STs share the same channel.However, STs in different cells may not have equal probabilities since numbers of STs in different cells may differ from each other.This means the thinned processes are not HPPPs.But for ease of analysis, in this paper, similar to the assumptions in [6,[28][29][30], we assume STs in different cells in a generic channel constitute an HPPP by random displacement and thinning.
Besides, if there are ST STs in one annular region, the probability that a channel in C is used by an ST is ST / ch , while if ST > ch , the probability is 1.Since ST ∼ (Λ), the probability that a channel is used averaged over ST is where Γ() = ∫ ∞ 0 −1 − d is the Gamma function and Γ(, ) = ∫ ∞ −1 − d is the upper incomplete Gamma function.So STs in a generic channel form an HPPP Φ ST with intensity ST = ch by independent thinning and random displacement.And ch is termed channel in use probability.
An extreme case is that Λ is high enough and ch = 1 holds.This equals the setting that only one channel is considered in the network and channels of all cells are used.In this case, besides ch , other details of the network remain unchanged, so does the analysis.
We assume that each channel experiences a constant noise power and exponential path-loss ℓ() = − (or ℓ( − ) = ‖ − ‖ − ) with a uniform exponent > 2, where is a distance, , ∈ R are two points, and ‖ ⋅ ‖ is the Euclidean norm.Independent Rayleigh fading is considered which remains constant within one time slot.The fading from a PR/ST to its corresponding MBS is ℎ ∼ exp( ℎ ), fading from an ST to its SR is ∼ exp( ), and interference fading from an MBS/PR/ST to a PR/SR is ∼ exp( ).Moreover, for convenience we set a time slot duration to be a unit time.
Communication Model.
When an MBS serves a PR, it may turn into busy or idle mode during each time slot.We further divide each time slot into minislots [19] and all minislots are synchronized among MBSs.In addition, we assume that each MBS randomly and independently turns into idle mode in one of the minislots.With such assumptions, when the MBS is transmitting, we consider that an ST performs energy harvesting (EH) and ambient backscatter communication (BC) in and minislots, respectively, and we have + +1 = .Minislots for EH and BC of each ST are also randomly and independently selected with equal probability.Besides, the ST performs traditional information transmission (IT) using the harvested energy when the MBS is idle.In the following, these two communication modes are termed BC mode and IT mode, respectively.An example of minislots assignment and selection when = 7 is given in Figure 3.
In a generic minislot, STs performing energy harvesting form an HPPP Φ eh with intensity eh = ( /) ST , which can be seen as the thinning of Φ ST since each ST may work in EH mode within a minislot with probability /.Similarly, STs in BC mode and IT mode form HPPPs Φ bc and Φit with intensities bc = ( /) ST and λit = ST /, respectively.Moreover, MBSs in busy mode also form an HPPP Φ bs with intensity bs = (1 − 1/).
We further assume that the SR knows about its ST's work mode so as to perform corresponding decoding.Saturation condition is also assumed where a data packet is always ready for transmission.
Besides, we assume that when an ST performs BC, a portion of the received power is used for BC, and the backscatter efficiency is .So the backscatter power of a generic ST is where ‖ − ‖ is the distance from the MBS to the ST, while the remaining portion 1− is stored as energy [19] along with received power in EH mode.And the conversion efficiency is .In this case, the harvested energy within a time slot is where eb = ( + (1 − ) ).
In order to perform IT, an ST must satisfy the circuit power constraint that is more than that consumed by the circuit, while the remaining power − is used for data transmission.Since a time slot lasts a unit time, we will also describe and as the power.The average harvested energy can be obtained as and the limitation of distance from an ST to an MBS, that is, , can be derived as where is based on the setting that STs distributed in the annular region should satisfy the circuit power constraint.
When fading is considered, the probability that harvested energy is above demand is where is a parameter representing energy demand, is the distance from an ST to its MBS, ≥ () is the pdf given in (2), ℎ is the fading, and where , ℎ , , and eb have been presented.Therefore, the constraint satisfaction probability sat of an ST is In the following, sat will also be termed power satisfaction probability.Taking this probability into account, STs in IT mode that have enough energy form an HPPP Φ it with intensity it = sat λit = sat ST /.In the following, these STs are called working STs.
When performing IT, the available energy is − , but for ease of analysis, in this paper we adopt the setting that a portion of the harvested energy can be used for active information transmission [31] and energy not used is ignored in different time slots.So the transmit power of a generic ST is where ‖ − ‖ is the distance from the MBS to the ST.
Interference Model.
In [32], authors mentioned interference regeneration that a backscatter node reflects all incident signals which leads to a square number of interference components for each SR.One effective solution is to adopt spread spectrum techniques.In this paper, we simplify this problem and assume that an ST backscatters only the signal from its corresponding MBS.
Backscattered Signal to an SR in IT Mode.
Backscattered signal from an ST to an SR in IT mode is considered as an extra path.We assume an SR in IT mode has the ability to eliminate multipath effect like PRs [4], so such interference is ignored.
Transmitted Signal to an SR in BC Mode
. Since an SR in BC mode senses and decodes backscattered signal by sensing changes in the signal caused by backscattering [4], we also ignore such interference.
Backscattered Interference to an SR in BC Mode
. Such interference comes from other STs ( ) in BC mode whose corresponding MBSs ( ) are in busy mode.Interfering STs come from Φ bc \ { 0 }.Such interference is firstly transmitted from an MBS to its ST, encountering path-loss ℓ( − ) and Rayleigh fading ℎ ∼ exp( ℎ ), then it is backscattered from an ST to the typical SR, encountering path-loss ℓ( − 0 ) and Rayleigh fading ∼ exp( ).The interference is denoted as where bc is the backscatter power of an ST in BC mode given in (4).
Transmitted Interference to an SR in IT
where it , bs represent two interference components, respectively, and it is the transmit power of an ST in IT mode given in (11).
Transmitted Interference to a PR. All interference to a
PR is identical to the interference to an SR in IT mode.We will analyze performance of a PR located at the origin encountering no interference from STs to make a comparison to reveal the effect on PRs.
Rate Model for Backscatter Link.
In [4], the bit rate (alternating sequence of ones and zeros) of the ambient backscatter prototype is related to the setting of circuit elements.Similar settings are also used in [8,17,18] and will be used in this paper, too.Besides, we assume if the signal-to-interferenceplus-noise-ratio (SINR) of backscatter communication is above a threshold, the predesigned rate can be achieved [20].Notations used in this paper are listed in the Notations.
Channel Inversion Power Control and Energy Storage and Reusing
Since energy is precious for secondary users, the harvested energy should be used more efficiently.So in this section we apply channel inversion power control to ST's active information transmission to avoid poor signal transmitted from STs.Moreover, since power control is applied, there is a higher probability that secondary users do not use up its energy.Therefore, we also propose a simple energy storage and reusing mechanism, to improve the utilization of harvested energy.
Channel Inversion Power Control.
In this part, we use channel inversion power control to let STs make less interference to primary users and conserve energy, while keeping their active transmission reliable.To be specific, an ST in IT mode will not use up its energy but transmit at a power to invert the path-loss to make sure that the average received power at its SR is equal to an SR's sensitivity in IT mode [33].Specifically, if the available energy is enough to support channel inversion power control, the ST will transmit at power , where is the distance between a secondary pair.So will highly impact the performance of STs in IT mode.Besides, if the available energy cannot support the power control to achieve the SR's sensitivity, the ST will not transmit.So an ST's transmit power using channel inversion power control is which can be rewritten as The probability that an ST transmits at power , which takes circuit power constraint into account, is where (⋅) is given in (8).For convenience, > /+ is termed power constraint with sensitivity and is termed sensitivity satisfaction probability.Since STs are independent from each other, working STs in IT mode adopting the power control form an HPPP Φ with intensity = λit = ST /.Any working ST will transmit at power and the remaining energy is ignored.
In the following, for convenience, when we analyze STs and PRs under power control, we still use notations which exist in normal settings, but with a slight difference when the notations involve STs in IT mode.For example, it under power control equals ∑ ∈Φ \{ 0 } ‖ − 0 ‖ − .
Energy Storage and Reusing.
As described above, an ST has a probability to get enough energy for transmitting in IT mode.If it does not get enough, in our previous settings, the unused energy is ignored and cannot be used in other time slots.Here we consider the setting that the energy can be stored, for a potential reusing.And we assume an ST has an energy storage component with capacity .Besides, energy storage follows channel inversion power control, and the remaining energy when an ST can perform active transmission is also stored.
If the energy demand for active transmission is , herein = / + , where is the circuit power consumption, the stored energy in a time slot is Note that we assume an ST can detect whether the harvested energy is enough for active transmission before trying to transmit.If not enough, the ST will not work and cost no energy.Besides, the energy consumption for detection is ignored, so the harvest energy is and we do not consider the case that < .From (8) we know the sensitivity satisfaction probability is ( ); hence the expectation of stored energy is For ease of analysis, we assume once an ST has stored enough energy for active transmission before a time slot, it will use that energy in the slot.But the unused part and harvested energy in that slot are ignored.Therefore, the expectation of stored energy after = ⌈ / ⌉ time slots is enough for another active transmission.So in the next time slot, an ST certainly has enough energy for active transmission, and the sensitivity satisfaction probability of the +1 slots increases to ( ( ) + 1)/( + 1).For simplicity we consider only the next time slot, so the increase of sensitivity satisfaction probability is where 1 − ( ) is the dissatisfaction probability, and this part means the improvement works only when harvested energy is not enough.
Although the reusing mechanism of stored energy is rough, it provides a view of reusing stored energy and increases the chance of active transmission.Some point processes (p.p.) described in the paper, along with their descriptions, intensities (inten.),and values are listed in Table 1 to provide a clear view.
Coverage Probability and Achievable Rate
We analyze the coverage probabilities and average achievable rates of an SR in different communication modes and a PR, by using Shannon formula.The average rate is Besides, since SINR ≥ 0, it is easy to derive where is the SINR threshold.So in the following, we will derive the coverage probability of the typical SR or PR in form where is the desired signal, is the interference, and is the noise.After deriving the coverage probabilities, average rates can be easily obtained.Firstly, we show some properties of the derivation processes which will be used for the following analysis and theorems.
Property 1.If the pdf of distance from an MBS to an ST is ≥ () in ( 2), the following expectation of can be derived easily: Property 2. If fading ℎ ∼ exp( ℎ ), the following expectation of ℎ can be derived easily: where is the path-loss exponent and Γ() = ∫ ∞ 0 −1 − d is the Gamma function.
Property 3. If is a random variable, the following integral of can be represented in nonintegral form: where is the path-loss exponent.Be similar to [26], we denoted the result in (25) as () with some differences.
Property 4. Given the Laplace transform L () of a continuous random variable , the pdf () of can be recovered by inverse Laplace transform as [21] Note that, in this paper, we focus on a generic channel, a typical primary receiver, and a typical secondary receiver.Unless otherwise stated, the following analyzed coverage probabilities and average rates are analyzed under the setting that the receivers exist already.
Coverage Probability of an SR in BC Mode.
Here we analyze the coverage probability of the typical secondary receiver working in backscatter mode; that is, its secondary transmitter backscatters signals for data transmission.The desired signal power is given in (4) and the interference is analyzed in Section 2.4.
Theorem 5. The coverage probability of an SR in BC mode located at the origin is
where and E[ −2 ], E[ℎ 2/ ], and () are given in (23), (24), and (25), respectively, where is the distance from a generic interfering ST to its MBS having the same pdf given in (2).
Proof.See Appendix.
Coverage Probability of an SR in IT Mode.
Here we analyze the coverage probability of the typical secondary receiver working in information transmission mode, that is, receiving signals transmitted by its secondary transmitter.The desired signal power is given in (11) and the interference is analyzed in Section 2.4.Note that the typical ST in IT mode also has a probability of satisfying circuit power constraint, so the average coverage probability of the typical SR is which can be further written as Since sat has been analyzed, in the following, we mainly derive it = P[SINR it > ] under the assumption that the typical ST satisfies the circuit power constraint, and coverage probability refers to it for convenience.
Theorem 6. The coverage probability of an SR in IT mode located at the origin is
where and is given in (28).
Proof.See Appendix.
Coverage Probability of a PR.
Here we analyze the coverage probability of the typical primary receiver with regard to interference from STs in IT mode.The desired signal comes from its MBS, and the interference is analyzed in Section 2.4.
Theorem 7. The coverage probability of a PR located at the origin is
where and is given in (28).
Proof.See Appendix.
Coverage Probability of a PR without Interference from STs.
Here we analyze the coverage probability of the typical primary receiver without interference from STs in IT mode.The desired signal still comes from the MBS and the primary receiver suffers only interference from other MBSs.
Theorem 8. The coverage probability of a PR considering no interference from STs is
where and are given in (34).
Proof.See Appendix.
Coverage Probability of an SR in IT Mode under Power
Control.
Here we analyze the coverage probability of the typical secondary receiver working in information transmission mode and adopting power control.The main differences are the transmit power of STs and their constraint satisfaction probability from sat to .Similarly to Section 4.2, coverage probability averaged over constraint satisfaction probability is and in the following, coverage probability refers to it , = P[SINR it > ] under the assumption that the typical ST satisfies the power constraint with sensitivity.
Theorem 9. The coverage probability of an SR in IT mode under power control is
where Proof.See Appendix.
Coverage Probability of a PR under Power Control
Theorem 10.The coverage probability of a PR suffering interference from STs in IT mode under power control is where , are given in (34), and Proof.See Appendix.
Numerical Results
We evaluate our analytical results by simulations.The simulation region is a square with side length of 10 km.The simulation results are obtained by averaging over 1000 runs.Unless otherwise stated, parameter values are listed in Table 2.
To generate uniformly distributed points in an annular region, we use a native method of generating a random point within a circle with radius , and if the random value is smaller than the inner radius , the point will be generated repeatedly until it is in the annular region.A mathematical method is to use inverse transform sampling.
Channel in Use Probability.
Figure 4 shows the channel in use probability ch given in (3).The simulation of channel selection is performed by random selection from all possible combinations.STs clustering around an MBS select idle channels using the method in Section 2.2; that is, they randomly select idle channels.We focus on the typical channel (numbered 1) and count how many typical channels are selected by STs clustering around all MBSs.The analytical and simulation curves overlap since the channel selection is simple, and the results are almost identical.
Power Satisfaction Probability and Sensitivity Satisfaction
Probability. Figure 5 shows the power satisfaction probability sat given in (10) the probability decreases, but not much.The reason is that a larger makes STs closer to MBSs, and STs then can harvest more energy to support a high satisfaction probability.
Figure 6 shows the sensitivity satisfaction probability given in (16), of an ST under power control.As SRs' sensitivity increases from −70 dBm to −50 dBm, decreases sharply.The reason is that does not change , but requires more energy.And when decreases further, the probability is already very low, so the change of satisfaction probability is small.The difference between these two figures also reveals the importance of .Moreover, is affected by circuit power constraint as well, which further affects the energy demand to transmit at .mode, and IT mode under power control.Since distance from an ST to its MBS and distance from an SR to its ST are both very short, signals decay not much.Besides, as described in Section 2.4, transmitted signals from MBSs and STs in IT mode make no interference to an SR in BC mode, so coverage probability of BC mode is much higher, while coverage probability of IT mode (under power control) is lower.Moreover, there is an intersection of curves of probabilities of IT mode and IT mode under power control.This happens since an SR in normal IT mode receives varied-power signals related to the ST's harvested energy.This means an ST may transmit at a high power, resulting in a high SINR, and may transmit at a low power, resulting in a low SINR.So the SINRs of SRs in normal IT mode range widely.On the other hand, an SR in IT mode under power control always receives constant-power (equals its sensitivity) signals from its ST, if fading is not considered.This means, there are few SRs having (extremely) low SINRs or high SINRs, which makes the curve shrink horizontally.
Coverage Probabilities.
Note that these probabilities are not averaged over sat or ; that is, we assume the typical ST has already satisfied the constraints.
Figure 8 shows the analytical and simulated coverage probabilities of a PR.The probabilities of a PR interfered or not by STs in IT mode change too little to be observed, so the figure shows that only two curves represent analytical and simulated results.Since an ST transmits by using harvested energy, the transmit power is relatively lower to the power of MBSs.Besides, only STs whose MBSs are idle can transmit, and the idle ratio is low (1/7 in our settings), so the number of interfering STs to a PR is small.These are the two main reasons why STs' interference to a PR is so low.As for an SR in IT mode, since it is very close to its ST, it still gets high SINR even when the ST transmits at low power.
Distributions of Interference Caused by STs in IT Mode and under Power Control.
As interference caused by STs in IT mode (and under power control) to a PR has quite limited effect on the PR's SINR, here we give the distributions of interference caused by STs.Following (26) in Property 4, (B.3) in Appendix, and (E.3) in Appendix, the analytical interference distributions caused by STs in IT mode and STs under power control are drawn in Figures 9 and 10, respectively.
Effect of Minislots Assignment on Coverage Probabilities.
The different assignments of minislots, that is, different and , result into different numbers of STs in backscatter mode and energy harvesting mode, that is, intensities bc and eh of Φ bc and Φ eh , respectively.The intensity bc does not affect backscatter power of an ST, because backscattering is instantaneous.But bc affects interference power since it changes number of STs in backscatter mode.However, as described in previous subsection, SINR of an SR in BC mode is much higher, which means change of interference power makes little effect on the SINR, where noise affects.So in both analytical and simulated results, when changes from 1 to 5, bc varies quite little (less than 3%) when 5 dB is selected as the SINR threshold.
As for coverage probabilities of an SR in other modes, they change quite little, too (less than 1% and 4% in analytical and simulated results, resp.).The reason is that when decreases, all STs harvest more energy (on average), so their transmit powers increase together.On the other hand, since we assume each MBS turns into idle in only one minislot, the minislots assignment does not affect number of STs in IT mode.Therefore, makes little change on their SINRs.As for PR, since STs' impact of it is very limited as shown in the previous subsection, its SINR changes little (less than 1% and 3% in analytical and simulated results, resp.) when varies.
Coverage Probabilities When Channels Are Fully Used.
When considering special case fully used channels, ch = 1 holds and other settings remained.The analytical and simulated results of coverage probabilities in this case are very close to the above probabilities under normal settings.The analytical average changes over different SINR thresholds are 0.11%, 0.01%, 0.01%, 0%, 0%, and 0.01% corresponding to BC mode, IT mode, IT mode under power control, a PR, a PR without STs, and a PR (STs under power control), respectively.And the simulated average changes are 0.90%, 1.81%, 0.83%, 0.83%, 0.47%, and 0.83%, respectively.So when ch = 4 and Λ changes from 3 to a high enough value, the coverage probabilities change little.
Average Rates of STs in IT Mode.
Figure 11 shows average rates of an ST in IT mode versus circuit power consumption , with different Λ.As increases, the rate becomes higher, too.This reason is that affects given in (7).A smaller means a smaller since STs have to be distributed closer to MBSs to get enough energy.However, since we assume a portion of the harvested energy can be used to perform information transmission, the power of the desired signal increases as well, resulting in a higher rate.Figure 12 shows average rates of an ST in IT mode under power control versus SRs' sensitivity , with different Λ.As increases, an SR receives a higher SINR because the desired signal is stronger, and there are less interfering STs.
Note that the rates are averaged over ch which represents the utility ratio of a channel by STs.Besides, since it is observed that STs in IT mode have limited interference to PRs, rates of a PR are omitted.
Average Stored Energy and Reusing.
A larger circuit power consumption or SRs' sensitivity makes energy demand higher, ( ) lower, smaller, and higher.So the impact on average stored energy in a time slot cannot be observed directly from (18).Figures 13 and 14 show the analytical and simulated results of versus and , respectively.
Figures 15 and 16 show the sensitivity satisfaction probability increase Δ given in (19).These results are computed using (19), and the simulation data needed are obtained from simulations.Besides, analytical results are computed in small granularity (about 100 data points in both figures), while a spot of simulated data points is drawn in circles.Under our parameter settings, when increases, Δ decreases sharply with a slow increase after that.But when increases, Δ varies like a staircase function and decreases overall after the beginning increase.The staircase follows which contains a ceiling function.Once the energy demand is large enough or averaged stored energy is less enough, increases by 1, and the satisfaction probability drops down.
Conclusions
In this work, we have analyzed the performance of ambient backscatter communications in RF-powered cognitive radio networks based on stochastic geometry.Besides, we have applied channel inversion power control to active information of secondary users.After that, a simple energy storage and reusing mechanism has been designed and analyzed to improve utilization of harvested energy.Analytical results for constraint satisfaction probabilities and coverage probabilities of secondary users and of primary users considering both communication modes of secondary users have been obtained.Besides, average rates of secondary users have been obtained based on coverage probability.As for energy reusing, we have analyzed average stored energy during a time slot and the increase of constraint satisfaction Wireless Communications and Mobile Computing 13 probability when power control is applied.The numerical results validate our theoretical analysis.Also, the results show performance improvement of secondary systems with only limited impact on the performance of primary systems.The analytical results and simulations demonstrate that integrating ambient backscatter communications into RF-powered cognitive radio network is a promising way to achieve energy and spectrum efficient wireless communications, which is suitable for certain Internet of things (IoT) applications.
Appendix A. Proof of Theorem 5
By definition of coverage probability with SINR threshold , we can start as where bc is the backscatter power of the typical ST given in (4), is the distance from the typical MBS to the typical ST, and is the constant distance from the typical ST to the typical SR.The inner probability can be derived as where = ( ℎ /( )) and L bc () is the Laplace transform of bc and can be derived as where the last step follows from the probability generating functional of PPP [23].The lower limit 0 follows from the fact that ST's positions are independent from each other, and SRs are in isotropic directions, so it could happen that another ST is close enough to an SR.The inner expectation of is By changing the variable = ( −1 bc ) 1/ , L bc () is further simplified as where where is the distance from an interfering ST to its corresponding MBS with pdf ≥ (), which means E[ −2 ] can be derived as (23).Besides, plugging , E[ 2/ bs ], and (25) into L bc () gives And plugging intermediate results into former equations will complete the proof.Note that here has no specific physical meaning, so for convenience, in other proofs, we still use the notation with different values when we derive Laplace transforms.
B. Proof of Theorem 6
Here we show parts of the proof since it is similar to Appendix.
Wireless Communications and Mobile Computing
By definition of coverage probability, where it is the transmit power of the typical ST in (11) and and are given in (A.1).By using the method in Appendix, it is easy to derive the inner probability as where = ℎ /( eb ).Since sum = it + bs , we can get L sum () = L it () ⋅ L bs ().Continually following the steps in Appendix, we can derive the two Laplace transform components as where Note that MBSs transmit at unit power, so there is no power notation before ‖ − 0 ‖.Finally, we can get the desired result by substituting intermediate results.
C. Proof of Theorem 7
Here we also give parts of the proof starting from definition of coverage probability: where is the distance from the typical MBS to the typical PR, () is given in (1), and sum is given in (13) where the main difference is the lower limit in the second equality, which follows from the fact that each PR is associated with its nearest MBS.The proof is similar to Appendix, so the remaining parts are omitted.
D. Proof of Theorem 8
Since the proof can be seen as parts of Appendix, we give only the definition of this coverage probability: where bs is given in (13).The result follows from removing the part of interference of STs in Appendix and proof details are omitted.
E. Proof of Theorem 9
When adopting power control, the desired signal power and interfering power from other STs in IT mode both change, so these two components are the main differences when compared to it in Appendix.By definition of coverage probability, where is the transmit power of the typical ST and is the constant distance from the typical ST to the typical SR.Different from Appendix, since the transmit power is constant, the derivation will be much simpler as the following:
F. Proof of Theorem 10
Since interfering power from other STs to a PR is the sole difference when STs adopt power control, only L it () differs from that in Appendix.Moreover, it is the same as (E.3) except the specific value of , and is the same as that in Appendix.So we omit the proof details.
𝛼:
Path-loss exponent : Backscatter portion in BC mode , : Signal to DC efficiency and backscatter efficiency : Intensity of a point process : Parameter of an exponential distribution (fading) : Available portion of harvested energy Φ: AP o i s s o np o i n tp r o c e s s ℎ: Fading of signal from an MBS to its ST or its PR : Fading of any interference signal : Fading of signal from an ST to its SR : C o v e r a g ep r o b a b i l i t y ch : Probability that a channel of a cell is used by an ST sat : Circuit power constraint satisfaction probability : Probability that an ST transmits at power , , , : Transmit, desired signal, interference, and noise power , : Inner and outer radii of the annular region L (): Laplace transform of r.v. C, : Channel set and a channel : M i n i s l o tn u m b e ro fat i m es l o t , : Backscatter and energy harvesting minislot number , : Harvested energy and circuit power consumption : SINR Threshold , , , : An MBS, a PR, an ST, and an SR (or their positions) : Sensitivity of an SR in IT mode : Transmit power considering power control ⋅ bs or ⋅ bs : Notations about busy MBSs ⋅ bc or ⋅ bc : Notations about STs in BC mode ⋅ it or ⋅ it : Notations about STs in IT mode or working STs ⋅ 0 : The typical entities (⋅ can be MBS, PR, ST, or SR).
Figure 1 :
Figure 1: Network composition illustration.Centers of dual-circles (annular regions), red dots, and blue triangles represent MBSs, primary receivers, and secondary transmitters, respectively.To ensure visibility, only one ST is showed in an annular region and all SRs are omitted.(Radii of annular regions in the illustration are enlarged.)
Figure 2 :Figure 3 :
Figure 2: System model of our network.ℎ, , refer to three types of fading described in Section 2.2.
Figure 4 :
Figure 4: Channel in use probability versus Poisson distribution parameter (i.e., average number of STs clusters around an MBS).
Figure 5 :Figure 6 :
Figure 5: Power satisfaction probability of an ST.
Figure 7 :
Figure 7: Analytical coverage probability of an ST in BC mode (bc), IT mode (it), and IT mode under power control (it ).
25 Figure 8 :Figure 9 :
Figure 8: Analytical and simulated coverage probabilities of a PR.
Figure 10 :
Figure 10: Analytical distribution of interference caused by STs in IT mode (under power control).
Figure 11 :Figure 12 :
Figure 11: Average rates of an ST in IT mode versus circuit power consumption.
Figure 13 :Figure 14 :
Figure 13: Average stored energy of an ST versus circuit power consumption.
Figure 15 :
Figure 15: Average stored energy of an ST versus circuit power consumption.
Figure 16 :
Figure 16: Average stored energy of an ST versus sensitivity.
Table 1 :
Some point processes and their descriptions.
Table 2 :
Parameter value list.
/ .Similarly L sum () = L it () ⋅ L bs (), and the two Laplace transform components can be derived as ∈Φ \{ 0 } − 0 − ) ] bs ()is the same as (B.5) except the specific value of .Substituting intermediate results will finish the proof. | 9,840.6 | 2018-04-02T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
Target Design in SEM-Based Nano-CT and Its Influence on X-ray Imaging
Nano-computed tomography (nano-CT) based on scanning electron microscopy (SEM) is utilized for multimodal material characterization in one instrument. Since SEM-based CT uses geometrical magnification, X-ray targets can be adapted without any further changes to the system. This allows for designing targets with varying geometry and chemical composition to influence the X-ray focal spot, intensity and energy distribution with the aim to enhance the image quality. In this paper, three different target geometries with a varying volume are presented: bulk, foil and needle target. Based on the analyzed electron beam properties and X-ray beam path, the influence of the different target designs on X-ray imaging is investigated. With the obtained information, three targets for different applications are recommended. A platinum (Pt) bulk target tilted by 25° as an optimal combination of high photon flux and spatial resolution is used for fast CT scans and the investigation of high-absorbing or large sample volumes. To image low-absorbing materials, e.g., polymers or organic materials, a target material with a characteristic line energy right above the detector energy threshold is recommended. In the case of the observed system, we used a 30° tilted chromium (Cr) target, leading to a higher image contrast. To reach a maximum spatial resolution of about 100 nm, we recommend a tungsten (W) needle target with a tip diameter of about 100 nm.
Introduction
Compared to electron imaging, in SEM-based nano-CT, the electron beam is focused on an X-ray target, leading to photon emission.In combination with a suitable X-ray detector, X-ray imaging can be realized by direct magnification.The method takes advantage of a fine electron spot size, leading to the formation of a small X-ray focal spot, resulting in a spatial resolution down to about 100 nm for X-ray imaging [1,2].Therefore, the system expands the modalities of SEM and is highly suitable for correlative microscopy [2][3][4][5].
Since SEM-based CT utilizes geometrical magnification for imaging, the hardware only consists of an X-ray target, a rotary sample stage available from SEM and an X-ray detector.The setup does not use X-ray optics tuned to a specific X-ray energy.In contrast, the X-ray target is easily accessible so it can be exchanged and adapted to the requirements of the measurement and no further changes to the system are needed.This allows for a variety of X-ray target modifications in terms of chemical composition, geometry and tilt.Varying these parameters allows for influencing X-ray beam characteristics like the energy distribution, intensity and focal spot size, which affects image quality parameters like the contrast and spatial resolution.In the past, different materials and shapes have been used, but a detailed correlation between the target design and its influence on X-ray imaging has not been investigated in detail.Research has been focused on improving the spatial resolution by minimizing the interaction volume of the electron and target and therefore reducing the target size.The target shape varies from bulk material [6], thin foils [7][8][9], to small wedge-shaped geometry, down to wire-shaped [10] and needle-shaped targets with a tip diameter of about 100 nm [3].Target materials like aluminum, titanium, tungsten, tantalum, gold or platinum-iridium alloys have also been mentioned but very few of them have been characterized [7,11].In addition, only the melting point has been taken into account, especially if the size of the target is very small.
We focus on an overall view of the target design, electron-matter interaction and its influence on X-ray beam properties with the goal to enhance the X-ray image quality.In the first step, the electron beam properties are determined as a base of the X-ray source.With knowledge about the electron beam and with the support of simulations, different targets with a varying geometry and chemical composition are analyzed experimentally in the second step.As a result, we present different target designs, leading to different X-ray beam properties suitable for various applications.
Experimental Setup
XRM-II nano-CT (ProCon X-ray, Sarstedt, Germany) is a system based on the field emission SEM JEOL JSM-7900F (JEOL, Tokyo, Japan) and illustrated in Figure 1.In addition to a secondary electron detector, the system is also equipped with an EDAX Element system (EDAX-AMETEK, Berwyn, PA, USA) for energy dispersive X-ray spectroscopy (EDS).Inside the electron column, the instrument is equipped with a probe current detector (PCD) to measure the electron beam current after passing through the optical system and aperture.There is a set of different aperture diameters that can be exchanged to vary the electron beam spot size and electron current.In addition to the PCD, a picoampere meter is integrated into the sample stage to measure the absorbed current of the sample (ACM).With an X-ray target manipulator inside the vacuum chamber and a photon counting X-ray detector (PCXD) (WidePIX from ADVACAM, Praha, Czech Republic) attached to it outside the vacuum chamber, a cone beam CT based on geometric magnification can be realized.The PCXD consists of 2 × 5 Medipix3 devices with a pixel size of 55 µm, leading to a total amount of 1280 × 512 pixels.The detector is constructed with a 1 mm CdTe sensor layer.
J. Imaging 2023, 9, x 2 of 19 imaging has not been investigated in detail.Research has been focused on improving the spatial resolution by minimizing the interaction volume of the electron and target and therefore reducing the target size.The target shape varies from bulk material [6], thin foils [7][8][9], to small wedge-shaped geometry, down to wire-shaped [10] and needle-shaped targets with a tip diameter of about 100 nm [3].Target materials like aluminum, titanium, tungsten, tantalum, gold or platinum-iridium alloys have also been mentioned but very few of them have been characterized [7,11].In addition, only the melting point has been taken into account, especially if the size of the target is very small.We focus on an overall view of the target design, electron-matter interaction and its influence on X-ray beam properties with the goal to enhance the X-ray image quality.In the first step, the electron beam properties are determined as a base of the X-ray source.With knowledge about the electron beam and with the support of simulations, different targets with a varying geometry and chemical composition are analyzed experimentally in the second step.As a result, we present different target designs, leading to different Xray beam properties suitable for various applications.
Experimental Setup
XRM-II nano-CT (ProCon X-ray, Sarstedt, Germany) is a system based on the field emission SEM JEOL JSM-7900F (JEOL, Tokyo, Japan) and illustrated in Figure 1.In addition to a secondary electron detector, the system is also equipped with an EDAX Element system (EDAX-AMETEK, Berwyn, PA, USA) for energy dispersive X-ray spectroscopy (EDS).Inside the electron column, the instrument is equipped with a probe current detector (PCD) to measure the electron beam current after passing through the optical system and aperture.There is a set of different aperture diameters that can be exchanged to vary the electron beam spot size and electron current.In addition to the PCD, a picoampere meter is integrated into the sample stage to measure the absorbed current of the sample (ACM).With an X-ray target manipulator inside the vacuum chamber and a photon counting X-ray detector (PCXD) (WidePIX from ADVACAM, Praha, Czech Republic) attached to it outside the vacuum chamber, a cone beam CT based on geometric magnification can be realized.The PCXD consists of 2 × 5 Medipix3 devices with a pixel size of 55 µm, leading to a total amount of 1280 × 512 pixels.The detector is constructed with a 1 mm CdTe sensor layer.The electron beam (blue) generated in the electron column (1) interacts with an X-ray target (here: needle-shaped target), (2) leading to X-ray emission (orange) used for imaging.X-rays are attenuated by the specimen mounted to a sample holder (3) and have to pass a Be window (4) and travel through air before hitting the X-ray detector (5).An EDS detector allows for chemical analysis for material characterization (6).The electron beam (blue) generated in the electron column (1) interacts with an X-ray target (here: needle-shaped target), (2) leading to X-ray emission (orange) used for imaging.X-rays are attenuated by the specimen mounted to a sample holder (3) and have to pass a Be window (4) and travel through air before hitting the X-ray detector (5).An EDS detector allows for chemical analysis for material characterization (6).
For nano-CT, the electron beam is focused on a target emitting a continuous X-ray spectrum, due to Bremsstrahlung and characteristic X-ray emission, which is attenuated by the sample volume right in front of the target.X-ray attenuation is determined by the chemical composition and density of the investigated sample.After penetrating the object, photons leave the vacuum chamber by passing through a 250 µm Beryllium (Be) window and hit the PCXD at a distance of about 426 mm to the X-ray source.Due to an energy threshold, the PCXD detects photons only above 5 keV.Further description of the CT mode of the XRM-II as well as volume reconstruction can be found in [2,3,5].The aforementioned equipment of the PCD, ACM, PCXD, EDX detector and electron imaging is used for experimental characterization.
Simulation of Electron-Target Interaction
In addition to experimental characterization, simulations are applied to evaluate the parameters of the X-ray source, which are not accessible via experiments.For this purpose, we used the simulation software CASINO (v2.48,Université de Sherbrooke, Québec, Canada) [12] and NIST DTSA-II (Lorentz 2020-05-18 revision, National Institute of Standards and Technology, Gaithersburg, MD, USA) [13].Both software are based on Monte Carlo simulation and provide a different scope of simulation tasks.CASINO is specially designed to simulate the electron trajectories of low-energy beam interaction in an SEM and is used to determine parameters like the backscatter coefficient or electron penetration depth.DTSA-II allows for the simulation of X-ray spectra as a result of the interaction with materials of a different geometry and chemical composition.Since DTSA is designed to simulate EDS, only Si(Li) and SDD detectors are available in the software, leading to a discrepancy in absorption efficiency compared to CdTe as the sensor material in the XRM-II.To reach a photon absorption of nearly 100%, like with CdTe, the sensor thickness is enhanced to a maximum of 100 mm.In both software, the setup of the components and electron beam parameters are adapted to the XRM-II.
How to Influence X-ray Beam Properties in SEM-Based Nano-CT
Multiple parameters influence the properties of X-ray radiation in SEM-based CT.The two most important parameters are the electron beam properties and X-ray target properties.The following sections outline a detailed characterization of the electron beam and X-ray target as well as the consequences of their interaction for X-ray imaging.
Electron Beam Characteristics
A precise knowledge of electron beam characteristics is necessary to quantify the electron-matter interaction of the X-ray source and as the input for electron simulations.Electron beam properties are mainly determined by the electron energy, electron current and electron spot size, which are investigated in the following.
The electron energy has a major influence on the emitted X-ray spectrum and X-ray focal spot.In SEM, the electron energy can reach 5-30 keV by selecting the corresponding acceleration voltage so the X-ray photon energy is limited to a maximum of 30 keV.The choice of acceleration voltage also determines the characteristic X-ray excitation, which should be about two to three times higher for an optimal photon outcome [14].This is important because characteristic X-ray photons contribute strongly to the X-ray intensity of the spectrum.Furthermore, the electron energy determines the mean free path and penetration depth of electrons in matter and leads to a specific extension of the electronmatter interaction volume.Its size correlates with the X-ray focal spot size and influences the spatial resolution in X-ray imaging, which is discussed later on.The electron beam spot size is another important parameter, which also slightly influences the electron energy, as explained later on.
The electron beam current is the second parameter to look at and strongly correlates with the used electron aperture.There are different apertures at two different positions installed in the electron beam path.One is an aperture at the top of the electron column with a diameter of 2 mm, which reduces misaligned electrons and is called the noise canceller (NC).The second one is the objective lens aperture (OLAP), positioned at the lower end of the column, which is a set of four apertures with different diameters between 30 and 1000 µm, as shown in Table 1.The electron beam current can be influenced in two ways: Firstly, by manually exchanging the aperture in the electron column, which limits the electron spot size and therefore results in a change in the electron beam current.Secondly, by incrementally changing the beam current controlled in the software (PC values), which causes a variation in the electron spot size as well.In terms of X-ray imaging, a high electron current is of interest to reach the maximum X-ray intensity.
Depending on the selected apertures, the electron beam current is measured with the PCD in the electron column and validated with a Faraday cup and ACM on the sample stage.The results are shown in Figure 2a for different probe current values, which can be selected in the software.The electron probe current (PC) measured with PCD perfectly matches the absorbed current (AC) measured with the Faraday cup except for the use of A0.For conventional SEM usage, this aperture is not used for imaging and only for beam alignment.The measured effect can be explained by the beam divergence and scattering of misaligned electrons.Despite the measured discrepancy, this fact can be neglected because only the maximum current (PC 18) is relevant for X-ray imaging and in this case, the PC and AC are identical.A1 results in a maximum electron beam current of about 330 nA.Smaller apertures are not useful for X-ray imaging and result in a very low X-ray intensity and extremely long image acquisition time.In the case of A0, the spot size increases and a maximum current of about 440 nA can be reached, which enhances to about 1 µA without the NC.Furthermore, the selected acceleration voltage influences the electron beam current, as shown in Figure 2b, for different apertures.For a high X-ray flux, only an acceleration voltage of 30 keV is most effective.Another advantage of a higher electron energy is the increasing probability of characteristic X-ray excitation.The only disadvantage of highenergy electrons is an increased interaction volume and X-ray focal spot.Nevertheless, the latter can also be minimized by the target geometry, as explained in Section 3.2.
The electron beam current is the second parameter to look at and strongly correlates with the used electron aperture.There are different apertures at two different positions installed in the electron beam path.One is an aperture at the top of the electron column with a diameter of 2 mm, which reduces misaligned electrons and is called the noise canceller (NC).The second one is the objective lens aperture (OLAP), positioned at the lower end of the column, which is a set of four apertures with different diameters between 30 and 1000 µm, as shown in Table 1.The electron beam current can be influenced in two ways: Firstly, by manually exchanging the aperture in the electron column, which limits the electron spot size and therefore results in a change in the electron beam current.Secondly, by incrementally changing the beam current controlled in the software (PC values), which causes a variation in the electron spot size as well.In terms of X-ray imaging, a high electron current is of interest to reach the maximum X-ray intensity.
Depending on the selected apertures, the electron beam current is measured with the PCD in the electron column and validated with a Faraday cup and ACM on the sample stage.The results are shown in Figure 2a for different probe current values, which can be selected in the software.The electron probe current (PC) measured with PCD perfectly matches the absorbed current (AC) measured with the Faraday cup except for the use of A0.For conventional SEM usage, this aperture is not used for imaging and only for beam alignment.The measured effect can be explained by the beam divergence and scattering of misaligned electrons.Despite the measured discrepancy, this fact can be neglected because only the maximum current (PC 18) is relevant for X-ray imaging and in this case, the PC and AC are identical.A1 results in a maximum electron beam current of about 330 nA.Smaller apertures are not useful for X-ray imaging and result in a very low X-ray intensity and extremely long image acquisition time.In the case of A0, the spot size increases and a maximum current of about 440 nA can be reached, which enhances to about 1 µA without the NC.Furthermore, the selected acceleration voltage influences the electron beam current, as shown in Figure 2b, for different apertures.For a high X-ray flux, only an acceleration voltage of 30 keV is most effective.Another advantage of a higher electron energy is the increasing probability of characteristic X-ray excitation.The only disadvantage of high-energy electrons is an increased interaction volume and X-ray focal spot.Nevertheless, the latter can also be minimized by the target geometry, as explained in Section 3. the target or the electron density, respectively.Estimating the electron spot size with the help of the spatial image resolution, as is common in SEM, is not suitable because of high electron currents, leading to thermal drift or the melting of the test objects.To determine the electron spot size, a modified method from Zhao et al. [15] was adapted to the provided hardware.As shown in Figure 3, a needle made of tungsten is used as the test object and placed on a brass sample holder, which is connected to an ACM.While the spot mode is activated, the AC is measured by moving the electron beam stepwise along a line over the W needle.The test object size (here: needle diameter) should be significantly larger than the electron beam step size to obtain a clear measuring signal.Moreover, the object should not be too large since a longer measurement time leads to thermal drift of the system and inaccurate measurements.A minimum object size is not given.In the presented case, we used a needle with a diameter of about 400 nm and an electron beam step size of 10-20 nm for measuring.As soon as parts of the electron beam hit the object surface, the AC reduces tremendously due to the high backscatter coefficient of W. As soon as the entire electron beam hits the target, the absorbed current reduces only slightly as long as electrons transmit the W target.The electron spot size s e can be calculated by subtracting the object width w o from the distance of the signal change ∆x: Another essential parameter of SEM-based CT is the electron spot size, which determines, in combination with the electron current, the number of electrons per area hitting the target or the electron density, respectively.Estimating the electron spot size with the help of the spatial image resolution, as is common in SEM, is not suitable because of high electron currents, leading to thermal drift or the melting of the test objects.To determine the electron spot size, a modified method from Zhao et al. [15] was adapted to the provided hardware.As shown in Figure 3, a needle made of tungsten is used as the test object and placed on a brass sample holder, which is connected to an ACM.While the spot mode is activated, the AC is measured by moving the electron beam stepwise along a line over the W needle.The test object size (here: needle diameter) should be significantly larger than the electron beam step size to obtain a clear measuring signal.Moreover, the object should not be too large since a longer measurement time leads to thermal drift of the system and inaccurate measurements.A minimum object size is not given.In the presented case, we used a needle with a diameter of about 400 nm and an electron beam step size of 10-20 nm for measuring.As soon as parts of the electron beam hit the object surface, the AC reduces tremendously due to the high backscatter coefficient of W. As soon as the entire electron beam hits the target, the absorbed current reduces only slightly as long as electrons transmit the W target.The electron spot size se can be calculated by subtracting the object width wo from the distance of the signal change Δx: se = Δx − wo. ( Since only the position of the signal change is used to calculate the beam spot size, the method is independent of the object shape.Moreover, the result delivers the maximum extension of the electron beam regardless of the electron density distribution or Gaussian distribution, respectively.Therefore, it will be an overestimation of the effective electron spot size.Different parameters like the aperture, PC value, electron energy and working distance (distance between the electron exit point and focus plane (WD)) influence the electron spot size.As discussed before, suitable parameters for X-ray imaging are PC 18, aperture A1 or A0 and an electron beam energy of 30 keV.The latter is also the best choice Since only the position of the signal change is used to calculate the beam spot size, the method is independent of the object shape.Moreover, the result delivers the maximum extension of the electron beam regardless of the electron density distribution or Gaussian distribution, respectively.Therefore, it will be an overestimation of the effective electron spot size.
Different parameters like the aperture, PC value, electron energy and working distance (distance between the electron exit point and focus plane (WD)) influence the electron spot size.As discussed before, suitable parameters for X-ray imaging are PC 18, aperture A1 or A0 and an electron beam energy of 30 keV.The latter is also the best choice because a higher electron energy leads to a smaller electron spot size due to a lower beam divergence.The mentioned electron beam parameters are selected to measure the electron beam spot size with the developed method depending on the WD.The results are depicted in Figure 4 and show a strong change in the electron spot size.Reducing the WD by about 10 mm minimizes the electron beam spot by about one-half for A0 and A1.In case the NC is removed from the beam path, the overall spot size is strongly increased but it can also be reduced by lowering the WD.The optical axis for X-ray imaging with the XRM-II is adjusted to a WD of 18 mm.Due to the presented data, the setup should be modified to obtain the smallest possible WD to reach the maximum efficiency.
depicted in Figure 4 and show a strong change in the electron spot size.Reducing the WD by about 10 mm minimizes the electron beam spot by about one-half for A0 and A1.In case the NC is removed from the beam path, the overall spot size is strongly increased but it can also be reduced by lowering the WD.The optical axis for X-ray imaging with the XRM-II is adjusted to a WD of 18 mm.Due to the presented data, the setup should be modified to obtain the smallest possible WD to reach the maximum efficiency.
Due to the exit window flange blocking the X-ray beam path, the optical axis is geometrically limited to a WD of 13 mm so the PCXD is only lifted up by about 5 mm.Because of the modification, a smaller electron beam spot and higher electron density are achieved so the X-ray intensity is enhanced in the case of a needle target and the X-ray focal spot size is minimized in the case of a bulk or foil target.To complete all the influences on the electron beam properties, the operator has to be mentioned as a parameter.Alignment of the focus and astigmatism is operator-dependent and has a significant influence on the electron beam shape and spot size hitting the target surface.
Geometry of X-ray Targets
In addition to the electron beam characterization, we investigated different target designs with a varying chemical composition, geometry and tilt angle.As previously discussed, a limitation of the X-ray focal spot size can be realized by reducing the acceleration voltage.A lower electron energy leads to a lower electron penetration depth and better spatial resolution for X-ray imaging will be achieved.At the same time, this would lead to disadvantages like a larger electron spot size (Figure 4a) due to the electron beam divergence and lower emission probability of characteristic X-rays.A more effective method is limiting the geometrical extension of the target.The target geometry and tilt determine the size of the electron-matter interaction volume, influencing the emitted Xray intensity and spatial resolution in X-ray imaging.
Due to the dimension of an X-ray target, we distinguished between the following geometries:
•
Bulk target: The target size is significantly larger than the interaction volume of the electrons and matter.No electrons will be transmitted through the target.
•
Foil target: A thin target layer below the size of the electron-matter interaction volume in one dimension.Typical dimensions are below 1 µm.Due to the exit window flange blocking the X-ray beam path, the optical axis is geometrically limited to a WD of 13 mm so the PCXD is only lifted up by about 5 mm.Because of the modification, a smaller electron beam spot and higher electron density are achieved so the X-ray intensity is enhanced in the case of a needle target and the X-ray focal spot size is minimized in the case of a bulk or foil target.
To complete all the influences on the electron beam properties, the operator has to be mentioned as a parameter.Alignment of the focus and astigmatism is operator-dependent and has a significant influence on the electron beam shape and spot size hitting the target surface.
Geometry of X-ray Targets
In addition to the electron beam characterization, we investigated different target designs with a varying chemical composition, geometry and tilt angle.As previously discussed, a limitation of the X-ray focal spot size can be realized by reducing the acceleration voltage.A lower electron energy leads to a lower electron penetration depth and better spatial resolution for X-ray imaging will be achieved.At the same time, this would lead to disadvantages like a larger electron spot size (Figure 4a) due to the electron beam divergence and lower emission probability of characteristic X-rays.A more effective method is limiting the geometrical extension of the target.The target geometry and tilt determine the size of the electron-matter interaction volume, influencing the emitted X-ray intensity and spatial resolution in X-ray imaging.
Due to the dimension of an X-ray target, we distinguished between the following geometries:
•
Bulk target: The target size is significantly larger than the interaction volume of the electrons and matter.No electrons will be transmitted through the target.
•
Foil target: A thin target layer below the size of the electron-matter interaction volume in one dimension.Typical dimensions are below 1 µm.
•
Needle target: The target geometry is similar to a needle and the interaction volume is strongly limited in two dimensions.The needle tip is pointing toward the X-ray detector.
The different target geometries are visualized in Figure 5, where incoming electrons are depicted in yellow and electron trajectories within the target material in green.Backscattered and transmitted electrons are represented in black.
resolution.The spatial resolution in the x-direction is independent of the target tilt angle and determined by a combination of the electron spot size and lateral electron propagation.The spatial resolution in the z-direction has to be distinguished depending on the target tilt angle.For low target tilt angles, the electron penetration depth determines the spatial resolution while for high target tilt angles, the electron spot size determines the spatial resolution.In terms of the X-ray intensity, the bulk target reaches a maximum since incoming electrons transfer their complete energy into the target.Foil (Figure 5b) or needle targets (Figure 5c) show different behaviors, which limit the spatial extent of electrons so only part of the electron energy is used for ionization and X-ray emission, before transmitting the target.Consequently, the focal spot size is reduced, the spatial resolution is enhanced but the X-ray intensity is decreased.In the case of a foil target, the X-ray focal spot size is determined by the same parameters as explained for the bulk target with one exception: for low target tilt angles, the electron penetration depth is limited by the foil thickness so the spatial resolution in the z-direction is also determined by the foil thickness.In the case of a needle target, the target geometry determines the X-ray focal spot size in the x-and z-direction.Another effect of the geometrical limitation of the X-ray focal spot size is an increase in the X-ray energy.Since high-energy photons are only emitted near the target surface, the energy of the emitted photon decreases with the increasing layer thickness and electron penetration depth, respectively.For example, reducing the layer thickness of a W foil target from 500 nm to 100 nm increases the mean X-ray photon energy by 1 keV, as simulations show.
Reducing the WD has no influence on the X-ray intensity while using a bulk or needle target, since the number of electrons hitting the target does not change.This is different for the use of a needle target.A smaller electron spot size at a low WD leads to an increased Regarding the bulk target, electrons reach the maximum penetration depth and the X-ray focal spot extends to its full size.In this case, the electron-target interaction volume (more details in Section 3.3.1),which is equivalent to the X-ray focal spot, approximately determines the spatial resolution in X-ray imaging.As soon as the target tilt and detector position are taken into account, the effective X-ray focal spot size, defined as the projection of the X-ray focal spot onto the detector surface, is the relevant parameter for spatial resolution.The spatial resolution in the x-direction is independent of the target tilt angle and determined by a combination of the electron spot size and lateral electron propagation.The spatial resolution in the z-direction has to be distinguished depending on the target tilt angle.For low target tilt angles, the electron penetration depth determines the spatial resolution while for high target tilt angles, the electron spot size determines the spatial resolution.In terms of the X-ray intensity, the bulk target reaches a maximum since incoming electrons transfer their complete energy into the target.
Foil (Figure 5b) or needle targets (Figure 5c) show different behaviors, which limit the spatial extent of electrons so only part of the electron energy is used for ionization and X-ray emission, before transmitting the target.Consequently, the focal spot size is reduced, the spatial resolution is enhanced but the X-ray intensity is decreased.In the case of a foil target, the X-ray focal spot size is determined by the same parameters as explained for the bulk target with one exception: for low target tilt angles, the electron penetration depth is limited by the foil thickness so the spatial resolution in the z-direction is also determined by the foil thickness.In the case of a needle target, the target geometry determines the X-ray focal spot size in the x-and z-direction.Another effect of the geometrical limitation of the X-ray focal spot size is an increase in the X-ray energy.Since high-energy photons are only emitted near the target surface, the energy of the emitted photon decreases with the increasing layer thickness and electron penetration depth, respectively.For example, reducing the layer thickness of a W foil target from 500 nm to 100 nm increases the mean X-ray photon energy by 1 keV, as simulations show.
Reducing the WD has no influence on the X-ray intensity while using a bulk or needle target, since the number of electrons hitting the target does not change.This is different for the use of a needle target.A smaller electron spot size at a low WD leads to an increased electron density hitting the target and therefore the X-ray intensity increases.For bulk and foil targets, the X-ray intensity can be influenced by the target tilt and will be discussed in detail in Section 3.3.1.Further investigations of the needle target are presented in Section 3.3.2.
Chemical Composition of X-ray Targets
The choice of target material determines the energy distribution of the emitted X-ray spectra, influences the behavior of attenuation and thereby the contrast of images.The characteristic X-ray line energy is highly important to select a suitable target material because characteristic photons strongly contribute to the X-ray spectrum.As listed in Table 2, the amount of characteristic photons varies from about 10-80% regarding the entire X-ray spectrum in the range of 5-30 keV and leads to a large difference in the overall X-ray intensity.The intensity of characteristic X-rays depends on their energy in combination with the excitation energy.As mentioned before, the excitation energy needs to be at least two to three times higher than the characteristic X-ray line energy to achieve the maximum intensity.Since the maximum electron energy in SEM is limited to 30 keV, the characteristic X-ray energy should not exceed an energy of around 15 keV for a high intensity.In addition, the detected X-ray energy is limited to 5 keV due to the PCXD's energy threshold.Further important parameters of the target material are the melting point and thermal conductivity, which determine the thermal stability, especially for small targets.Based on these facts, Table 2 shows potential target materials with different characteristic X-ray line energies in a range of around 5-15 keV, their melting point and their relevant characteristic X-ray line energy.Additionally, a maximum emission depth of a photon with an energy of 5 keV is listed.Elements with low melting points and thermal conductivity are only suitable for foil or bulk target geometry.
Table 2. Target materials, their melting point, relevant characteristic X-ray energy and the amount of characteristic photons excited with 30 keV electrons regarding the entire X-ray spectrum in the energy range between 5 and 30 keV.
Element
Melting Point [ Due to the high thermal energy input of the focused electron beam, parameters like the melting point and thermal conductivity are important regarding the target geometry.In contrast to a bulk target, which can be realized with every material, a needle target can only be realized by using materials with a very high melting point like W or in combination with a low electron flux.To take advantage of a smaller target volume, manufacturing a needle target embedded into a diamond substrate or a foil target sputtered onto a diamond substrate could also be realized.Thus, the spatial resolution for materials with a high electron penetration depth like Cr or Y will be enhanced.
The selected target material also influences the X-ray focal spot size especially in the case of a bulk material since the interaction volume is not limited by the target geometry.The incoming electron energy E 0 (in keV), atomic number Z, atomic mass number A (in u) and density ρ (in g/cm 3 ) of the target material determine the probability of the electronmatter interaction and therefore the electron penetration depth.In addition, only electrons with an energy above E c = 5 keV can excite photons used for imaging due to the PCXD's detector threshold.Castaing or Gaber and Fitting define the maximum X-ray emission depth z m (in nm) for incoming electrons perpendicular to the object surface as [16,17]: Based on the approximation, a maximum X-ray emission depth of a 5 keV photon is calculated for potential target materials and listed in Table 2.As an example, W or Pt strongly limit the electron penetration, leading to a maximum photon emission depth of about 1 µm, while the maximum photon emission depth of light metals like Cr is about three times higher.In the case of the XRM-II, one has to keep in mind that the approximation will be an overestimation due to different facts.First, X-ray imaging is determined by a volume of high photon density and not by photons with the lowest detectable energy.Second, the target is tilted for imaging, leading to a lower electron penetration depth and therefore a reduction in the X-ray focal spot size.Third, only the effective X-ray focal spot size (projection of the X-ray focal spot onto the detector surface) is relevant for imaging.
A further parameter influencing X-ray radiation is the X-ray beam path determined by the setup, which will be discussed in the following.Later on, we present the characterization of different target geometries and materials in terms of the X-ray intensity and spatial resolution in X-ray imaging.
X-ray Beam Path in SEM-Based CT
In X-ray imaging, the emitted and detected X-ray spectra vary because of the detection efficiency of the detector.In the case of the XRM-II, the beam path leads to additional photon absorption and is quantified in the following.
To calculate the emitted X-ray spectra for different target materials, the simulation software DTSA-II was used and the simulation parameters were adapted to the setup of the XRM-II (Monte Carlo simulation of a bulk; no detector window; maximum Si sensor thickness of 100 mm to reach 100% absorption; 0 • detector tilt angle; 45 • object tilt angle).The simulation results are represented in Figure 6a and show massive variations in the characteristic X-ray intensity.Materials with low characteristic line energies (Cr, Fe, Cu) show a high X-ray intensity since a large amount of electrons, even after multiple scatter events, are able to excite photon emission.Materials with high characteristic X-ray line energies (Y, Mo, W, Pt) show a low X-ray intensity because of a low probability of ionization and X-ray emission.A quantitative evaluation is listed in Table 3, where the intensity (sum of the photons of the X-ray spectrum) of the emitted and detected photons as well as the resulting total transmission (difference between the emitted and detected photons) can be found.In addition, the median of the X-ray energy is listed.Regarding the absolute emitted intensity of different target materials, Cr is by far the highest one.Nevertheless, the detected intensity is quite low because low-energy photons and especially the characteristic peak at 5.41 keV are strongly absorbed in the X-ray beam path.In total, about 40% of the photons in the range of 5-30 keV are detected.Compared to that, the absorption of high-energy photons is significantly lower, but the probability to excite Since the X-ray detector of the XRM-II is placed outside the vacuum chamber, a Be window of 250 µm, transparent for hard X-rays, is necessary to separate the vacuum chamber from its environment.In addition to passing through the Be window, X-ray photons need to travel about 236 mm through air and pass a 500 nm Al layer in front of the sensor material before detection.Since the X-ray energy used in SEM-based CT is rather low, the absorption of these three components should be taken into account when calculating the X-ray intensity.The values of the total attenuation are generated from XCOM (NIST) at standard atmospheric pressure (1013 hPa) while air is defined as a mixture of 78% N, 21% O and 1% Ar [18].Figure 6b shows the calculated transmission depending on the photon energy.The absorption efficiency of the sensor material made of 1 mm CdTe is nearly 100% for the relevant energy range and can be neglected in the examination.The overall effect of the three materials on the photon transmission is summed up in the detection efficiency and mainly influenced by air.
Using the detection efficiency and employing the Lambert-Beer law, detected photons attenuated by the beam path can be calculated and are represented in Figure 6c.Since the X-ray detector energy threshold cuts off photons below 5 keV, the relevant energy range is 5-30 keV.It can be clearly seen that the X-ray beam path in air reduces the photon flux by 50% in the energy range between 5 and 10 keV.
A quantitative evaluation is listed in Table 3, where the intensity (sum of the photons of the X-ray spectrum) of the emitted and detected photons as well as the resulting total transmission (difference between the emitted and detected photons) can be found.In addition, the median of the X-ray energy is listed.Regarding the absolute emitted intensity of different target materials, Cr is by far the highest one.Nevertheless, the detected intensity is quite low because low-energy photons and especially the characteristic peak at 5.41 keV are strongly absorbed in the X-ray beam path.In total, about 40% of the photons in the range of 5-30 keV are detected.Compared to that, the absorption of high-energy photons is significantly lower, but the probability to excite them is rather low.Therefore, the transmission of Y, Mo, W and Pt X-ray spectra is about 75-80%.The target materials W and Pt obtain the overall highest intensity due to a relatively high X-ray emission and transmission.Table 3.All values relate to an energy range of 5-30 keV.The table shows X-ray intensity of simulated X-ray emission and calculated detection normalized to the emitted intensity of Cr.In addition, absolute transmission and median of detected X-ray energy are listed.Transferring the beam path to a vacuum increases the X-ray intensity by about 20-50% depending on the target material.Consequently, the acquisition time for X-ray CT would decrease by about the same proportion.Unfortunately, the currently equipped PCXD is not suitable to be used in an ultra-high vacuum but adaptations to the system are possible.Similar results could be reached by flooding this part of the beam path with He, which also reduced the X-ray absorption tremendously.
Element
To increase the contrast in X-ray imaging, the energy distribution of the X-ray spectra plays a decisive role.An indicator to select the right target material that fits the needs of an investigated specimen is the median of the detected X-ray energy (5-30 keV) shown in Table 3. Due to the beam path, the detected X-ray energy is enhanced and reaches values between 5.5 and 13.4 keV for the selected target materials.This information is used to influence the image contrast, as discussed later.
The beam path can also influence the spatial resolution in X-ray imaging since air scattering can broaden the beam and thus worsen the resolution.This effect is especially strong at low photon energies.
Target Design
Based on the presented results, a set of targets are designed to fulfill different tasks: one with a high X-ray intensity for fast measuring, another for a high spatial resolution and a third to enhance the image contrast for an investigation of low-absorbing materials.
Bulk Target
To reach the maximum X-ray intensity, the obvious choice of geometry is a bulk target since the entire electron energy is used for X-ray emission.As shown in Table 3, the target should be made out of W or Pt due to a high detected X-ray intensity.An important parameter for the X-ray intensity and formation of the X-ray focal spot in a bulk target is the tilt angle, which is investigated experimentally.
In Figure 7, the experimental measurements of the X-ray intensity show a maximum at around a 25-30 • target tilt angle for all targets, whereas Pt leads to the highest measured intensity of 800 photons per minute (330 nA electron beam current) followed by W. Additional measurements show that the use of aperture A0 and removal of the NC lead to an intensity of around 2300 counts/min for a 25 • tilted Pt bulk target which is an increase of a factor of three.
Target Design
Based on the presented results, a set of targets are designed to fulfill different tasks: one with a high X-ray intensity for fast measuring, another for a high spatial resolution and a third to enhance the image contrast for an investigation of low-absorbing materials.
Bulk Target
To reach the maximum X-ray intensity, the obvious choice of geometry is a bulk target since the entire electron energy is used for X-ray emission.As shown in Table 3, the target should be made out of W or Pt due to a high detected X-ray intensity.An important parameter for the X-ray intensity and formation of the X-ray focal spot in a bulk target is the tilt angle, which is investigated experimentally.
In Figure 7, the experimental measurements of the X-ray intensity show a maximum at around a 25°-30° target tilt angle for all targets, whereas Pt leads to the highest measured intensity of 800 photons per minute (330 nA electron beam current) followed by W. Additional measurements show that the use of aperture A0 and removal of the NC lead to an intensity of around 2300 counts/min for a 25° tilted Pt bulk target which is an increase of a factor of three.
The target materials Cu, Y and Mo reach about 60%, Fe about 50% and Cr about 40% intensity compared to Pt.These materials lead to a longer measurement time for a similar SNR, but they are important due to a different photon energy distribution, which results in differences in the image contrast.The relative simulated intensity listed in Table 3 deviates from the measured intensity: materials with characteristic photon energy photons close to the detector threshold, especially Cr, Fe and Cu, show a lower intensity in experiments compared to simulations.This originates most likely from the detector threshold being set to 5 keV.While standard values are about 8-10 keV, the selected value of 5 keV seems to be unstable and and detector threshold shifts to higher energies.Thermal drift may also lead to a higher energy threshold.The target materials Cu, Y and Mo reach about 60%, Fe about 50% and Cr about 40% intensity compared to Pt.These materials lead to a longer measurement time for a similar SNR, but they are important due to a different photon energy distribution, which results in differences in the image contrast.The relative simulated intensity listed in Table 3 deviates from the measured intensity: materials with characteristic photon energy photons close to the detector threshold, especially Cr, Fe and Cu, show a lower intensity in experiments compared to simulations.This originates most likely from the detector threshold being set to 5 keV.While standard values are about 8-10 keV, the selected value of 5 keV seems to be unstable and and detector threshold shifts to higher energies.Thermal drift may also lead to a higher energy threshold.
To understand the behavior of the X-ray intensity depending on the target tilt angle, further investigations were performed via DTSA simulations.Within DTSA, the target and detector position as well as detector pixel size can be adapted according to the XRM-II.The X-ray intensity is calculated as a sum of the X-ray spectrum depending on the target tilt angle, which is changed in increments of 5 • between an absolute value of 5 • and 85 • .The intensity is evaluated for three different conditions: (i) emitted X-ray spectrum (energy range from 0-30 keV), (ii) emitted X-ray spectrum in the detectable energy range from 5-30 keV (detector energy threshold) and (iii) detected energy spectrum attenuated by the beam path in the energy range from 5-30 keV (see Section 3.2).The latter is calculated using the Lambert-Beer law for every simulated spectrum.
As an example, Figure 8a shows the distribution of the W X-ray intensity depending on the target tilt angle for the three conditions.The intensity distribution is influenced by the electron backscatter coefficient (BSC), X-ray energy and photon absorption in the target itself, which is depicted in Figure 8b,c.To explain the behavior, one has to distinguish between low and high target tilt angles.
To understand the behavior of the X-ray intensity depending on the target tilt angle, further investigations were performed via DTSA simulations.Within DTSA, the target and detector position as well as detector pixel size can be adapted according to the XRM-II.The X-ray intensity is calculated as a sum of the X-ray spectrum depending on the target tilt angle, which is changed in increments of 5° between an absolute value of 5°and 85°.The intensity is evaluated for three different conditions: (i) emitted X-ray spectrum (energy range from 0-30 keV), (ii) emitted X-ray spectrum in the detectable energy range from 5-30 keV (detector energy threshold) and (iii) detected energy spectrum attenuated by the beam path in the energy range from 5-30 keV (see Section 3.2).The latter is calculated using the Lambert-Beer law for every simulated spectrum.
As an example, Figure 8a shows the distribution of the W X-ray intensity depending on the target tilt angle for the three conditions.The intensity distribution is influenced by the electron backscatter coefficient (BSC), X-ray energy and photon absorption in the target itself, which is depicted in Figure 8b,c.To explain the behavior, one has to distinguish between low and high target tilt angles.For condition (I), the X-ray intensity distribution depending on the target tilt angle is rather symmetrical with its maximum at an angle of around 45° (black graph).Using low and high target tilt angles leads to a strong intensity reduction.
The intensity using low target tilt angles is reduced, due to photon absorption in the target, as shown in Figure 8c.Photons need to travel long distances through the target material itself to reach the PCXD, which is positioned at a 90° angle to the incoming electron beam.Therefore, most low-energy photons are absorbed in the target before reaching the surface and the overall intensity is low.This effect can be proven by the evaluated median of the X-ray energy depending on the target tilt angle.As shown in Figure 8b, the X-ray energy decreases strongly for low target tilt angles, meaning that only high-energy photons are detected.
For high target tilt angles, the X-ray energy only slightly influences the behavior of condition (I) since high-energy and also low-energy photons can leave the target (Figure 8c).The reduction in the X-ray intensity at high target tilt angles is caused by the electron back scattering (BSC) shown in Figure 8b.The latter is constantly increasing with an increasing tilt angle to about 80%, so more and more electrons do not contribute to the X- For condition (I), the X-ray intensity distribution depending on the target tilt angle is rather symmetrical with its maximum at an angle of around 45 • (black graph).Using low and high target tilt angles leads to a strong intensity reduction.
The intensity using low target tilt angles is reduced, due to photon absorption in the target, as shown in Figure 8c.Photons need to travel long distances through the target material itself to reach the PCXD, which is positioned at a 90 • angle to the incoming electron beam.Therefore, most low-energy photons are absorbed in the target before reaching the surface and the overall intensity is low.This effect can be proven by the evaluated median of the X-ray energy depending on the target tilt angle.As shown in Figure 8b, the X-ray energy decreases strongly for low target tilt angles, meaning that only high-energy photons are detected.
For high target tilt angles, the X-ray energy only slightly influences the behavior of condition (I) since high-energy and also low-energy photons can leave the target (Figure 8c).The reduction in the X-ray intensity at high target tilt angles is caused by the electron back scattering (BSC) shown in Figure 8b.The latter is constantly increasing with an increasing tilt angle to about 80%, so more and more electrons do not contribute to the X-ray emission.In summary, the target absorption and BSC result in a reduction in the X-ray intensity at low and high target tilt angles and lead to a maximum intensity at a tilt angle of around 45 • .Condition (II) reveals an overall strong reduction in the X-ray intensity since photons below the energy threshold of 5 keV including the characteristic line at 1.8 keV are eliminated from the spectrum.At target tilt angles smaller than 10 • , most low-energy photons are absorbed within the target, leading to the similar values of conditions (I) and (II) in Figure 8a.This behavior differs for higher target tilt angles.As shown for condition (I), photons with an energy below 5 keV can leave the target for higher tilt angles but the detector threshold now eliminates these photons, so the X-ray intensity decreases significantly.Moreover, the BSC leads to a further reduction in the X-ray intensity, so the overall maximum of the X-ray intensity shifts from a target tilt angle of 45 • to 25 • and perfectly matches the experimental results presented in Figure 7.
In terms of the X-ray energy of condition (II), the 5 keV cut-off leads to a strong increase in the X-ray energy that barely changes for different tilt angles.This is a strong indicator that photon absorption in the target is less relevant for the overall X-ray intensity of condition (II).
As soon as the beam path of the XRM-II is taken into account, and only the detected photons are evaluated (condition (III)), the intensity is further reduced, but the W spectrum is rather weakly attenuated because of its relatively high X-ray energy.
Based on Table 3, a target made of Cr and Mo represents materials with a low and high X-ray energy.In the following, we will focus on these two materials and also on Pt due to the high X-ray intensity and quantify the different image contrasts that can be reached.Therefore, a test object made of C, Al and Fe is placed into the beam path and images with different target materials are acquired.The exposure time is adjusted to an equal flatfield intensity for every target material.Figure 9 shows a radiography of the test object with different materials imaged and marked areas used to determine the CNR.To calculate the CNR, the method of Bechara [19,20] is applied and defined as: µ is defined as the expectation value and σ as the standard deviation of the signal.All parameters can be obtained graphically from the image histogram.With the approximation of a Gaussian distribution signal, σ is given by: ray emission.In summary, the target absorption and BSC result in a reduction in the Xray intensity at low and high target tilt angles and lead to a maximum intensity at a tilt angle of around 45°. Condition (II) reveals an overall strong reduction in the X-ray intensity since photons below the energy threshold of 5 keV including the characteristic line at 1.8 keV are eliminated from the spectrum.At target tilt angles smaller than 10°, most low-energy photons are absorbed within the target, leading to the similar values of conditions (I) and (II) in Figure 8a.This behavior differs for higher target tilt angles.As shown for condition (I), photons with an energy below 5 keV can leave the target for higher tilt angles but the detector threshold now eliminates these photons, so the X-ray intensity decreases significantly.Moreover, the BSC leads to a further reduction in the X-ray intensity, so the overall maximum of the X-ray intensity shifts from a target tilt angle of 45° to 25° and perfectly matches the experimental results presented in Figure 7.
In terms of the X-ray energy of condition (II), the 5 keV cut-off leads to a strong increase in the X-ray energy that barely changes for different tilt angles.This is a strong indicator that photon absorption in the target is less relevant for the overall X-ray intensity of condition (II).
As soon as the beam path of the XRM-II is taken into account, and only the detected photons are evaluated (condition (III)), the intensity is further reduced, but the W spectrum is rather weakly attenuated because of its relatively high X-ray energy.
Based on Table 3, a target made of Cr and Mo represents materials with a low and high X-ray energy.In the following, we will focus on these two materials and also on Pt due to the high X-ray intensity and quantify the different image contrasts that can be reached.Therefore, a test object made of C, Al and Fe is placed into the beam path and images with different target materials are acquired.The exposure time is adjusted to an equal flatfield intensity for every target material.Figure 9 shows a radiography of the test object with different materials imaged and marked areas used to determine the CNR.To calculate the CNR, the method of Bechara [19,20] is applied and defined as: µ is defined as the expectation value and σ as the standard deviation of the signal.All parameters can be obtained graphically from the image histogram.With the approximation of a Gaussian distribution of the signal, σ is given by: FWHM ≈ 2.35 σ (4) The resulting CNR is listed in Table 4.The use of different target materials leads to significant changes in the contrast.The target material Pt shows the highest contrast for imaging Al and Fe.These two materials represent light metals and period 4 elements, which are the predominant materials being investigated.Due to this performance, Pt is The resulting CNR is listed in Table 4.The use of different target materials leads to significant changes in the contrast.The target material Pt shows the highest contrast for imaging Al and Fe.These two materials represent light metals and period 4 elements, which are the predominant materials being investigated.Due to this performance, Pt is most suitable for applications in SEM-based nano-CT.A further useful target material is Cr since it delivers a high contrast for low-absorbing materials like polymers.The advantage of the high characteristic energy of Mo is not helpful in enhancing the image contrast and is unsuitable for most applications.To complete the characterization of different bulk targets, a Siemens star is imaged to obtain the spatial resolution.The electron beam current was set to 330 µA and the targets were tilted to around 25-30 • .To avoid deviations in the target tilt angle between different target elements, all targets are glued on one tilted target holder.Nevertheless, this leads to a varying distance (micrometer range) between the target and Siemens star for the different target elements and results in different magnifications and pixel sizes.To prevent a collision of the target and Siemens star, the maximum magnification is limited.Images are acquired with a pixel size of 50 nm using a Cr target, 63 nm using a Mo target and 45 nm using a Pt target.All three targets were used to acquire 30 projections by integrating 24 images with an exposure time of 2500 ms.Post-processing was performed in ImageJ by applying a median filter with a radius of one pixel.After that, all 30 single images were summed up pixel per pixel using z-projection [21].The images of the Siemens star are shown in Figure 10 where the yellow arcs represent the defined pattern size.The 600 nm measured electron spot size and 1 µm maximum emission depth of 5 keV photons determined in Sections 3.1.1and 3.1.3seem to be contradictory compared to the 125 nm spatial resolution in X-ray imaging but these values describe the X-ray focal spot size and not the effective X-ray focal spot size.The X-ray focal spot size is nearly equivalent to the interaction volume of the electron beam and target while the effective Xray focal spot size is defined as the projection of the focal spot size onto the detector surface.The latter is significantly smaller and the only relevant parameter since it determines the spatial resolution in X-ray imaging.There are different facts underlining that the effective focal spot size is significantly smaller: the method to measure the electron spot size is an overestimation since it results in the maximum electron focal spot size regardless of the Gaussian electron beam distribution.Strongly scattered electrons lead to a large measured electron spot size even though they do not contribute to the effective Xray focal spot.Moreover, the target tilt angle will lead to a lower electron penetration depth and 5 keV photons will only slightly contribute to X-ray imaging.The effective Xray focal spot will be formed by photons with a higher energy.Furthermore, photon absorption due to the target tilt is not taken into account.The sum of these factors leads to a significant reduction in the effective X-ray focal spot size and to the measured spatial The target material Cr shows a clear anisotropy of the spatial resolution.On the one hand, an electron beam spot size of about 600 nm and a low scatter angle of electrons in Cr limit the horizontal extent of the X-ray focal spot, leading to a vertical spatial resolution of about 200 nm.On the other hand, the vertical extent of the X-ray focal spot is rather large due to the relatively high electron penetration depth in Cr, resulting in a maximum X-ray emission depth of about 3 µm, as shown in Table 2. Since characteristic X-rays determine about 80% of the entire X-ray spectrum (see Table 2) and their energy with 5.4 keV is rather low, photons used for imaging will be emitted from deep inside the target.This large vertical extent of the X-ray focal spot leads to a low horizontal spatial resolution of about 500 nm.
With the use of a Mo bulk target, the details of about 170 nm can be resolved since the electron penetration depth is smaller compared to Cr.Moreover, the entire X-ray spectrum contributes to X-ray imaging and not mainly low energy characteristic photons, leading to a reduction in the effective X-ray spot size.In addition, only a volume of high photon density determines the spatial resolution in X-ray imaging.
The target material Pt reveals a sharp and high contrast image due to the high Xray intensity and narrow spatial X-ray distribution, leading to an image with less noise compared to Cr and Mo.Pt shows the lowest electron penetration depth for the investigated elements, limiting the X-ray focal spot so details in the size of about 125 nm can be resolved.
The 600 nm measured electron spot size and 1 µm maximum emission depth of 5 keV photons determined in Sections 3.1.1and 3.1.3seem to be contradictory compared to the 125 nm spatial resolution in X-ray imaging but these values describe the X-ray focal spot size and not the effective X-ray focal spot size.The X-ray focal spot size is nearly equivalent to the interaction volume of the electron beam and target while the effective Xray focal spot size is defined as the projection of the focal spot size onto the detector surface.The latter is significantly smaller and the only relevant parameter since it determines the spatial resolution in X-ray imaging.There are different facts underlining that the effective focal spot size is significantly smaller: the method to measure the electron spot size is an overestimation since it results in the maximum electron focal spot size regardless of the Gaussian electron beam distribution.Strongly scattered electrons lead to a large measured electron spot size even though they do not contribute to the effective X-ray focal spot.Moreover, the target tilt angle will lead to a lower electron penetration depth and 5 keV photons will only slightly contribute to X-ray imaging.The effective X-ray focal spot will be formed by photons with a higher energy.Furthermore, photon absorption due to the target tilt is not taken into account.The sum of these factors leads to a significant reduction in the effective X-ray focal spot size and to the measured spatial resolution of 125 nm.A quantification of the focal spot size of different target materials and its influence on the spatial resolution is planned for the future.
Using a Pt bulk target leads to the best performance by far regarding the X-ray intensity, spatial resolution as well as image contrast for most applications.Due to this fact, a Pt bulk target is recommended as a standard target with high photon flux.To image low-absorbing materials like polymers, a Cr bulk target will be useful to enhance the contrast.Negative consequences are a decreasing X-ray intensity and spatial resolution.To improve the spatial resolution, a Cr foil target will be prepared in the future.An additional diamond substrate will support heat flow to resist the high thermal electron beam input.
Needle Target
The focal spot size of the X-ray source influences the spatial resolution in X-ray imaging systems based on geometrical magnification.Since the minimum electron beam spot size is about 600 nm (Figure 4), the X-ray focal spot can be further reduced by a smaller X-ray target.Therefore, we used commercially available W needles with a tip diameter of around 100 nm intended for SEM-based nano probing (Figure 11a).W is most suitable because of its high melting point of over 3400 • C and its high thermal stability.Materials with lower melting points will change shape due to the high thermal energy input of the electron beam and the tip radius will enlarge.
The graph in Figure 11b represents the X-ray intensity [counts/minute/pixel] depending on the needle diameter, which was acquired while the electron beam was positioned on the needle tip center.For every data point, the electron beam spot is moved in the needle tip direction (negative y-direction) by a certain distance, so the electron beam hits the target at a position equivalent to the displayed needle diameter.For larger needle diameters, the interaction volume increases and consequently, the X-ray intensity increases.Moreover, Figure 11b shows the amount of electrons which are collected underneath the needle target with a Faraday cup.It represents transmitted and primary electrons not interacting with the target.The graph clearly shows that the X-ray intensity will increase as long as the number of transmitted electrons decreases because of a larger interaction volume so more electrons contribute to the X-ray emission.In case the electron beam is positioned too far away from the tip, emitted X-rays are absorbed by the target itself and the intensity reduces.The graph in Figure 11b represents the X-ray intensity [counts/minute/pixel] depending on the needle diameter, which was acquired while the electron beam was positioned on the needle tip center.For every data point, the electron beam spot is moved in the needle tip direction (negative y-direction) by a certain distance, so the electron beam hits the target at a position equivalent to the displayed needle diameter.For larger needle diameters, the interaction volume increases and consequently, the X-ray intensity increases.Moreover, Figure 11b shows the amount of electrons which are collected underneath the needle target with a Faraday cup.It represents transmitted and primary electrons not interacting with the target.The graph clearly shows that the X-ray intensity will increase as long as the number of transmitted electrons decreases because of a larger interaction volume so more electrons contribute to the X-ray emission.In case the electron beam is positioned too far away from the tip, emitted X-rays are absorbed by the target itself and the intensity reduces.
As shown in Figure 11a, only a certain amount of primary electrons hit the needle target, limiting the efficiency of the X-ray excitation which illustrates a high importance in enhancing the electron beam density by a reduction in the WD.As a result, the spatial resolution is enhanced due to a smaller interaction volume and also the X-ray intensity is enhanced since more primary electrons hit the target.Moreover, a high number of transmitted electrons will lead to additional X-ray generation below the target (e.g., sample holder or sample itself), reducing the X-ray image quality.
The spatial resolution of a W needle target was already determined in previous works and reached 80 nm in 2D and 100 nm in 3D [2].Consequently, a needle target made of W is most suitable for high-resolution imaging due to thermal stability and a small Xray focal spot size.
Conclusions
In SEM-based CT, three different target geometries with varying target volumes are presented: bulk target, foil target and needle target.The geometrical limitation realized by reducing the target volume leads to a minimization of the X-ray focal spot, resulting in an enhanced spatial resolution in X-ray imaging but at the same time to a decreasing Xray intensity.Furthermore, the electron beam itself has an influence on the X-ray As shown in Figure 11a, only a certain amount of primary electrons hit the needle target, limiting the efficiency of the X-ray excitation which illustrates a high importance in enhancing the electron beam density by a reduction in the WD.As a result, the spatial resolution is enhanced due to a smaller interaction volume and also the X-ray intensity is enhanced since more primary electrons hit the target.Moreover, a high number of transmitted electrons will lead to additional X-ray generation below the target (e.g., sample holder or sample itself), reducing the X-ray image quality.
The spatial resolution of a W needle target was already determined in previous works and reached 80 nm in 2D and 100 nm in 3D [2].Consequently, a needle target made of W is most suitable for high-resolution imaging due to thermal stability and a small X-ray focal spot size.
Conclusions
In SEM-based CT, three different target geometries with varying target volumes are presented: bulk target, foil target and needle target.The geometrical limitation realized by reducing the target volume leads to a minimization of the X-ray focal spot, resulting in an enhanced spatial resolution in X-ray imaging but at the same time to a decreasing X-ray intensity.Furthermore, the electron beam itself has an influence on the X-ray properties, whereas an electron beam current of 300 nA or higher is suitable for SEM-based nano-CT.A lower current increases the image acquisition time tremendously and leads to inefficient measurements.Moreover, the distance between the electron focusing unit and X-ray target (working distance) should be as small as possible, as experiments show.As a result, the electron beam and X-ray focal spot size are reduced to a minimum, leading to an enhanced spatial resolution in X-ray imaging.In the case of a needle target, the X-ray intensity is increased because a reduction in the electron beam spot size is also equivalent to an increase in the electron beam density so more electrons hit the target.A method to determine the electron beam spot size using a high electron current (>300 nA) is also presented.
The X-ray intensity changes strongly between the emitted and detected spectrum due to the XRM-II setup.In addition to the absorption caused by a Be window, air and an Al layer, the X-ray detector is only sensitive for energies above 5 keV.Transferring the beam path to a vacuum will increase the X-ray intensity by about 20-50% depending on the target material.Consequently, the acquisition time for X-ray CT would decrease by about the same amount.The X-ray intensity of the bulk and foil target is strongly influenced by the target tilt angle.Experiments and simulations reveal that absorption caused by the setup leads to a maximum shift in the X-ray intensity from 45 • to smaller target tilt angles of 25-30 • .
For SEM-based CT, we recommend three targets for different applications: • Platinum bulk target tilted by 25 • : The target is a perfect combination of X-ray flux and spatial resolution.Due to the low penetration of electrons, the X-ray focal spot size is small and a 2D spatial resolution of about 125 nm is reached.At the same time, a combination of 30 keV excitation energy and 9.4 keV characteristic X-ray line energy leads to the highest X-ray intensity.Due to these properties, a Pt target is suitable as a standard target for SEM-based CT.It is especially recommended for imaging high-absorbing or large samples and for fast CT scans.
•
Chromium bulk or foil target tilted by 30 • : Imaging low-absorbing materials like polymers, target materials with their characteristic line energy right above the energy threshold of the detector are recommended.Following these recommendations, Cr is chosen for the presented system.In addition to an enhanced image contrast, a lower X-ray intensity and spatial resolution have to be expected as disadvantages due to the low probability of X-ray excitation and high electron penetration depth.The spatial resolution will be enhanced by changing the geometry from bulk to foil, but only with a further loss of the X-ray flux.
•
Tungsten needle target: To reach the highest spatial resolution and reveal small details in a specimen, a needle target with a tip diameter of about 100 nm made of W is recommended.The geometry limits the X-ray focal spot size, and consequently, enhances the spatial resolution up to 80 nm (2D), while still delivering enough photons for imaging.Moreover, W is able to resist the high thermal energy input of the electron beam due to its high melting point.The spatial resolution and X-ray intensity can be adapted in a specific range according to the requirements of the measurement by changing the electron beam position onto the target tip.
Figure 1 .
Figure 1.Illustration of XRM-II nano-CT:The electron beam (blue) generated in the electron column (1) interacts with an X-ray target (here: needle-shaped target), (2) leading to X-ray emission (orange) used for imaging.X-rays are attenuated by the specimen mounted to a sample holder (3) and have to pass a Be window (4) and travel through air before hitting the X-ray detector(5).An EDS detector allows for chemical analysis for material characterization(6).
Figure 1 .
Figure 1.Illustration of XRM-II nano-CT:The electron beam (blue) generated in the electron column (1) interacts with an X-ray target (here: needle-shaped target), (2) leading to X-ray emission (orange) used for imaging.X-rays are attenuated by the specimen mounted to a sample holder (3) and have to pass a Be window (4) and travel through air before hitting the X-ray detector(5).An EDS detector allows for chemical analysis for material characterization(6).
Figure 2 .
Figure 2. (a) A 30 keV electron beam current depending on PC value for different apertures; (b) electron beam current depending on acceleration voltage for different apertures (PC 18).
Figure 3 .
Figure 3. Principle for measuring electron beam diameter: Scanning over an object (here: needle) leads to signal change in absorbed current.Beam diameter can be calculated with the known object width.
Figure 3 .
Figure 3. Principle for measuring electron beam diameter: Scanning over an object (here: needle) leads to signal change in absorbed current.Beam diameter can be calculated with the known object width.
Figure 5 .
Figure 5. Cross-section of electron-target interaction volume: (a) bulk target; (b) 250 nm foil target; (c) needle target with 100 nm in diameter.Simulation is performed with DTSA with W defined as target material and an incoming electron beam size of 50 nm under an angle of 45°.α is defined as target tilt angle.Yellow arrows indicate incoming electron beam and object surface is drawn in orange.Green electron trajectories represent the traveled distance within the target and black trajectories represent backscattered or transmitted electrons.Bulk and foil targets are depicted in side view and needle target in front view.
Figure 5 .
Figure 5. Cross-section of electron-target interaction volume: (a) bulk target; (b) 250 nm foil target; (c) needle target with 100 nm in diameter.Simulation is performed with DTSA with W defined as target material and an incoming electron beam size of 50 nm under an angle of 45 • .α is defined as target tilt angle.Yellow arrows indicate incoming electron beam and object surface is drawn in orange.Green electron trajectories represent the traveled distance within the target and black trajectories represent backscattered or transmitted electrons.Bulk and foil targets are depicted in side view and needle target in front view.
Figure 6 .
Figure 6.(a) Simulated X-ray spectra excited with 30 keV electrons for various target materials showing different characteristic X-ray line energies and massive variations in X-ray intensity; (b) detection efficiency depending on photon energy: transmission of X-ray photons passing through 250 µm Be, 236 mm air and 500 nm Al before interacting with the sensor material.Absorption efficiency of the sensor material made of 1000 µm CdTe is nearly 100% for the entire energy range; (c) detected X-ray spectra calculated with Lambert-Beer law showing absorption due to the beam path in XRM-II.
Figure 6 .
Figure 6.(a) Simulated X-ray spectra excited with 30 keV electrons for various target materials showing different characteristic X-ray line energies and massive variations in X-ray intensity; (b) detection efficiency depending on photon energy: transmission of X-ray photons passing through 250 µm Be, 236 mm air and 500 nm Al before interacting with the sensor material.Absorption efficiency of the sensor material made of 1000 µm CdTe is nearly 100% for the entire energy range; (c) detected X-ray spectra calculated with Lambert-Beer law showing absorption due to the beam path in XRM-II.
Figure 7 .
Figure 7. Experimental measurement of the X-ray intensity depending on target tilt angle for various materials.Figure 7. Experimental measurement of the X-ray intensity depending on target tilt angle for various materials.
Figure 7 .
Figure 7. Experimental measurement of the X-ray intensity depending on target tilt angle for various materials.Figure 7. Experimental measurement of the X-ray intensity depending on target tilt angle for various materials.
Figure 8 .
Figure 8.(a) Normalized X-ray intensity of a W bulk target depending on target tilt angle showing a shift in the maximum intensity to lower tilt angles due to PCXD energy threshold.Evaluated are different conditions: (I) emitted photons in range of 0-30 keV, (II) emitted photons in range of 5-30 keV and (III) detected photons; (b) median of X-ray intensity and BSC of W depending on target tilt angle; (c) scheme of interaction volume for low and high target tilt angles showing the number of backscattered electrons indicated by the arrow size and the distance that photons need to travel to leave the target.
Figure 8 .
Figure 8.(a) Normalized X-ray intensity of a W bulk target depending on target tilt angle showing a shift in the maximum intensity to lower tilt angles due to PCXD energy threshold.Evaluated are different conditions: (I) emitted photons in range of 0-30 keV, (II) emitted photons in range of 5-30 keV and (III) detected photons; (b) median of X-ray intensity and BSC of W depending on target tilt angle; (c) scheme of interaction volume for low and high target tilt angles showing the number of backscattered electrons indicated by the arrow size and the distance that photons need to travel to leave the target.
Figure 9 .
Figure 9. Radiographic image of a test object made of pure iron, aluminum and carbon to determine CNR.The marked areas in yellow are used for evaluation and the values correspond to element thickness.The aluminum ball is fixed to the carbon surface using carbon glue as indicated with the red arrow.
9 .
Radiographic image of a test object made of pure iron, aluminum and carbon to determine CNR.The marked areas in yellow are used for evaluation and the values correspond to element thickness.The aluminum ball is fixed to the carbon surface using carbon glue as indicated with the red arrow.
Figure 10 .
Figure 10.Radiographic images of a Siemens star acquired with a 25° tilted bulk target showing differences in spatial resolution as well as contrast (yellow arcs indicate the defined pattern size.The arc labeled with * corresponds to a pattern size of 50 nm): (a) image acquired with a pixel size of 76 nm using a Cr target; (b) image acquired with a pixel size of 32 nm using a Mo target; (c) image acquired with a pixel size of 45 nm using a Pt target.
Figure 10 .
Figure 10.Radiographic images of a Siemens star acquired with a 25 • tilted bulk target showing differences in spatial resolution as well as contrast (yellow arcs indicate the defined pattern size.The arc labeled with * corresponds to a pattern size of 50 nm): (a) image acquired with a pixel size of 76 nm using a Cr target; (b) image acquired with a pixel size of 32 nm using a Mo target; (c) image acquired with a pixel size of 45 nm using a Pt target.
JFigure 11 .
Figure 11.(a) SE image of a W needle target with a tip size of less than 100 nm and a scheme of electron beam spot size focused on a target; (b) detected X-ray intensity of a W needle target depending on electron beam positioning on different needle diameter.
Figure 11 .
Figure 11.(a) SE image of a W needle target with a tip size of less than 100 nm and a scheme of electron beam spot size focused on a target; (b) detected X-ray intensity of a W needle target depending on electron beam positioning on different needle diameter.
Table 1 .
Diameters of objective lens apertures and NC.
Table 1 .
Diameters of objective lens apertures and NC.
Table 4 .
Calculated CNR for different materials and targets.Exposure time was varied to reach a similar mean flat-field intensity. | 18,513.6 | 2023-08-01T00:00:00.000 | [
"Physics"
] |
Topological quantum weak antilocalization limit and anomalous Hall effect in semimagnetic Bi2−xCrxSe3/Bi2Se3−yTey heterostructure
Magnetic topological insulator hosts both a non-trivial surface band and quantum anomalous Hall effect when tuning the time reversal symmetry by various chemical doping into the system. Using molecular beam epitaxy, we have synthesized Bi2−xCrxSe3/Bi2Se3−yTey heterostructure, which was scrutinized via in situ angle resolved photoemission spectroscopy, ex situ x-ray diffraction, Raman, and x-ray photoemission characteristics analyzing techniques. We used the Hikami-Larkin-Nagaoka formula for quantifying the weak antilocalization limit in Bi2−xCrxSe3 and Bi2Se3−yTey non-magnetic layer coupling near the ferromagnetic transition temperature (∼7 K). A higher concentration of Cr (1.5%) in Bi site leads to bulk carrier density for entire cooling temperature with a bandgap of 85 meV. The proximity effect of Te doped Bi2Se3 with magnetic topological insulator degrades the ferromagnetic response for this heterostructure. Our study suggests that the manipulation of disorder free magnetic top layer and fine tuning of spin–orbit coupling strength in the bottom topological insulator can be helpful to understand the high temperature quantum anomalous Hall effect towards relativistic quantum electronics applications.
Introduction
Magnetically coupled spin-orbit strength in non-magnetic topological insulators (MTI) pertain a major role for the rise of quantum anomalous Hall effect (QAHE) which is crucial for many potential applications [1][2][3][4][5][6][7][8][9][10][11][12]. To amend MTI, adding magnetic dopants into topological insulators (TIs) or putting TI adjacent to a magnetic layer have been adopted recently [13][14][15][16][17][18][19]. The latter one is called the proximity induced magnetism, which was demonstrated in Bi 2 Se 3 /YIG [20], Bi 2−x Cr x Se 3 /YIG [21], and Bi 2 Se 3 /EuS [22]. All of them have enhanced magnetism in TI layers. But on the other hand, can proximity effect also lower the magnetism of an MTI if it was put adjacent to a non-magnetic layer? To answer this question, we have prepared Bi 2−x Cr x Se 3 /Bi 2 Se 3−y Te y heterostructure using molecular beam epitaxy (MBE) technique. As without the bottom layer, Bi 2−x Cr x Se 3 shows ferromagnetic transistion temperature around 30 K [23]. In in situ angle resolved photoemission spectra (ARPES), the bandgap does not exhibit a Dirac surface state (DSS) due to broken time reversal symmtery (TRS). We also studied the magnetotransport and x-ray photoemission spectroscopy (XPS) to understand the proximity effect on magnetism in the top layer by a charge transfer mechanism for the deposited heterostructure. But, beyond the limitation of atmospheric gas adsorption and bulk ferromagnetic matrix at Dirac point (DP), the protection of robust surface state and QAHE in magnetic topological insulators are considered mandatory [24][25][26]. Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence.
Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Methods
All the existing layers intended for MTI/TI van der Waals heterostructure including a buffer layer fabrication were performed by SSI-Molecular Beam Epitaxy (MBE -Smart 100) equipped with an ultrahigh vacuum level system (base pressure ∼10 −10 torr). High purity Bi (99.9999%), Se (99.999%), Te (99.999%), and Cr (99.999%) were thermally evaporated through Knudsen cells. First, GaSe buffer layer was formed by annealing GaAs(111)B substrate in Se atmosphere (−3.4% lattice mismatch in Bi 2 Se 3 epitaxial growth) at 480°C for 10 min in terms of removing dangling bonds to form GaSe on As terminated GaAs(111)B surface [27]. Then, 5 quintuple layers (QLs) of Bi 2 Se 3−y Te y, a second nonmagnetic interface layer were imprinted by the co-evaporation of Bi, Se, and Te at 525°C, 103°C, and 255°C for 10 min. Eventually, the formation of thick 10 QLs of Bi 2−x Cr x Se 3 on 5 QLs of Bi 2 Se 3−y Te y /GaSe were formed when Bi, Cr, Se sources were evaporated at 525°C, 1030°C, and 115°C. Optimal Bi to Se flux ratio was kept at 1:10 for each layer-by-layer epifilm growth. The growth rate of 0.5 QL/ min remains almost constant for Bi 2−x Cr x Se 3 and Bi 2 Se 3−y Te y , which were determined from tuning the reflection high energy electron diffraction (RHEED) intensity oscillations. During the entire growth, the substrate temperature was maintained at 155°C under Se rich condition to avoid the chalcogens deficiency in the deposited heterostructure film. After the growth, the film was annealed at 200°C to eliminate the internal stress and improve the homogeneity of MTI/TI film.
The epitaxial growth optimized high quality MBE grown film were characterized by x-ray diffraction (Bruker D8 Discover single crystal diffractometer, λ=1.5406 Å), Raman (Horiba Jobin-Yvon HR800, Ar + laser 514 nm excitation), High-performance low noise atomic force microscopy (AFM) surface analysis (NT-MDT, NTEGRA SPECTRA II, Probe aperture size ∼100 nm), ARPES spectra (SPECS PHOIBOS 150 hemisphere analyzer with 3D Micro-Mott detector and 21.2 eV UVS 300 helium lamp), XPS (PHI 5000 VersaProbe, FWHM1.0 eV, Al Kα), and Magnetotransport analysis through six Hall probe geometry (Oxford 14 T magnet, Keithley 6221 AC/DC current source and SR830 lock-in amplifier system down to 1.6 K) experiments. ) c growth axis perpendicular to the substrate has been observed from an x-ray diffraction pattern as shown in figure 1(c). Except the three (003) and six-fold (006) periodicity peaks pertaining in-plane rotations observed upto the (0021) plane shows no additional peak, which reveals the substantial crystalline nature and epitaxial growth of Bi 2−x Cr x Se 3 heterostructures. Here, the diffraction peak (0021) at 37.5°is quite profound as observed in case of Bi 2 Se 3 epitaxial films [28]. The full width at half maximum (FWHM) of (003) and (006) major intensity peaks are 4.5°and 6.7°respectively. After comparison, the overlapping of Bi 2 Se 3 (009) diffraction curve into the main peak of GaAs(111)B (2θ ∼27°) was inviolably not considered for remaining analysis. From the observed (00 l) diffraction peak of Bi 2−x Cr x Se 3, the c-axis lattice parameter was calculated to be 28.55±0.02 Å for x=0.17 concentration. Two dimensional RHEED pattern was used to estimate a-axis lattice constant from its d-spacing value (a=3.98±0.01 Å). In our case, the decreasing of both 'a' and 'c' values are due to the substitutional effect rather than the insertion of Cr between the QLs of Bi 2 Se 3 can be affirmed from the d-spacing acquired from the RHEED images. This thin film heterostructure has a moderate root mean squared (RMS) roughness of 0.55 nm as observed from the AFM images (see figure 1(e)).
Results and discussion
We further investigated the local structure from the vibrational modes of Bi 2−x Cr x Se 3 /Bi 2 Se 3−y Te y heterostructure recorded using Raman spectra as shown in figure 1(d). Three active main peaks A , apart from IR Raman active modes have been observed at 71 cm −1 , 130 cm −1 and 172 cm −1 , which are close to Se rich Bi 2 Se 3 QLs stretching and bending vibrational modes respectively [29,30]. The ex situ x-ray photoemission spectra for 10 QLs Bi 2−x Cr x Se 3 /5 QLs Bi 2 Se 3−y Te y is depicted in figures 2(a)-(c). The constituent elements of Bi 2−x Cr x Se 3 top layer core levels peaks are Bi (4 f), Cr (2p), and Se (3d). The longer exfoliated sample exhibits other surface doping mechanism as a signature of native oxidation content (O 1s −531 eV) and hydrocarbon (C 1s −284 eV) mixture from the ambient air adsorption effect (see figure 2(d)). The core level peaks of Bi (4 f) is well fitted with corresponding Bi (4 f) energy levels of about 162.1 eV and 156.7 eV [31]. Similarly, chromium 3 + oxidations peaks can also be fitted for the 2p binding energy levels of 585.2 eV and 575.3 eV. 32 We see that the atomic concentration values of Cr ∼1.5% for Bi 2−x Cr x Se 3 and Te ∼10.5% in Bi 2 Se 3−y Te y at the initial peaks of Bi (4 f), Cr (2p), Se (3d), Te (3d) binding energy levels. From Bi 2 Se 3 core level XPS spectra, the top layered Bi (4 f) and Se (3d) peaks are strongly red shifts (more than 1.3 eV) to 156.7 eV and 51.8 eV. The binding energies that correspond to 3 + (2p) oxidation state of bulk Cr are 584 eV and 575 eV, which is smaller than our observed Cr (2p) values. The spin-orbit coupling strength gradually varied from Bi (4 f) to Cr (2p) as a consequence of the blue shift between them as well as large splitting of binding energy levels (∼9 eV) rather than Cr-Se bond contribution was observed [14]. From XPS, the electron doping effect on the surface caused by a residual atmospheric gas molecule absorption plays the major role for inducing charge transfer effect [32].
Dependence of in-plane longitudinal (R xx ) and transverse resistance (R xy ) on temperature and magnetic field measurements was performed and the response is as shown in figures 3(a)-(d). Non-monotonic change of resistance till 2 K can be clearly seen in figure 3(a). Below 10 K, there have been metallic to insulator transition denoted as T c , similar to the mechanism found in diluated magnetic semiconductors (DMS) and other relative magnetic topological insulators [33]. It is noted that the observed T c is still lower than Bi 2−x Cr x Se 3 [23]. A carrier dependent metallic phase with the reduction of temperature has been observed in bulk Bi 2 Se 3 and Bi 2 Se 3-x Te x (x=0.9) topological insulators [34,35]. A sharp upward R xx versus T insulating curve near T c was observed as shown in the inset of figure 3(a). In figure 3(b), a signature of weak antilocalization (WAL) behavior in low magnetic field range at above 3 K exhibits 2D surface conducting states, which is immune to defects scattering. Further, we observe a crossover to weak antilocalization (WAL) from weak localization (WL) characteristic as seen from the upward (Λ) to downward (V) R xx versus B curve shifts with respect to temperature. Cr doping in the place of Bi for more than x∼0.10 on Bi 2 Se 3 and Bi 2 Te 3 at low temperature was ascribed by WL with classical parabolic dominated magnetoresistance (MR) as reported previously [33,36]. A notable finding of WAL from the overlapping of magnetoconductance pattern is still absent even in low magnetic fields and wider cooling temperature up to 30 K from 1.6 K for Bi 2−x Cr x Se 3 /Bi 2 Se 3−y Te y heterostructure. To understand the 2D quantum correction over WAL and WL surface magnetoconductance (Δσ xx ) channels, we used the Hikami-Larkin-Nagaoka (HLN) relation [37]; 3(c)). This indicates that our top layered MTI has a single Dirac surface state with the moderate contribution of bulk carrier-dominated typical WL behavior at low temperature.
The anomalous Hall effect (AHE) of this semimagnetic 10 QLs Bi 2−x Cr x Se 3 /5 QLs Bi 2 Se 3−y Te y were examined from R xy versus B plot at various temperatures and are shown in figure 3(d). Both in the positive and negative magnetic field sweeping directions, the Hall resistance yeilds a negative slope for several cooling temperatures which in turn is points out electron as the intrinsic majority carriers. An indistinct ferromagnetic hysteresis is clearly seen up to 7 K and thereafter hysteresis quite vanishes as depicted in figure 3(d) inset. A ferromagnetic Curie temperature (T c ) would be verified in two different ways for the identification of AHE; i) a disappearance of H c in between below and above T c expected region from R xy isothermal hysteresis curves (see figure 3(f)), and ii) the execution of Arrot plots (R 2 xy versus B/R xy in figure 3(e)), in which R 2 xy axis zero intercept towards near ferromagnetic T c constituency [3]. From these analyses, we conclude that T c is represented as ∼7 K for this MTI heterostructure. To understand the bulk carrier influences in AHE signal other than Dirac electron on a 2D surface, we extract the Hall coefficient of the sample using R H =1/en, where e is the electron charge and n is defined as carrier concentration. We have also estimated the effective mobility using the Drude model [μ(T)=R H (T)/ρ xx (T)]. Bulk electron carrier density was found to be ∼10 14 cm −2 with relatively less mobility (∼18 cm 2 V −1 .s) at lower temperatures. This is comparatively two orders higher than bulk and epitaxial films of Bi 2 Se 3 and Bi 2 Te 3 [36]. Furthermore, higher concentration of Cr (x>0.20) supplies more free charge carriers in the form of holes in Bi 2 Te 3, Sb 2 Te 3 and electrons in Bi 2 Se 3 as a result of impregnable chemical potential disorder around the Dirac point [36,38]. We conclude that this scenario would be a descriptive of non-localized bulk carrier scattering rather than localized Dirac surface electrons in the vicinity of T c for this heterostructure.
Discrimination of bandgap size of 10 QLs Bi 2−x Cr x Se 3 with a relative change of chemical potential on Fermi level by residual bulk carrier concentration can be explained using ARPES surface band structure carried out at room temperature as shown in figure 1(f). Obviously, more than 1.5% of Cr bulk doping in the place of Bi would destruct the Dirac surface state. In our case, the size of the energy gap was 85 meV, which is comparable to other Cr and V doped Bi 2 Se 3 magnetic topological insulator for x>0.10 concentration [31,32]. This energy gap resides ∼130 meV from a Fermi level and the momentum broadening of the surface state is also seen by a bulk doping effect. Besides, a residual gas condensation on the topmost layer is also playing an important role in the resemblance of the bulk charge carrier effect. So, the bulk band gap size more moderately reduces rather than preserving the Dirac surface state for this magnetic topological insulator. Therefore, either controlling of bulk carrier density around DP by charge transfer coupling between the interface of MTI/TI or designing of less deficient heterostructure engineering have a significant response to the competitive realistic device applications.
Conclusion
In summary, MBE grown 10 QLs of Bi 2−x Cr x Se 3 on Bi 2 Se 3−y Te y topological insulator was extensively studied by various structural and ARPES surface band diagrammatic analyses. We found T c is lower than 10 K, which reveal that non-magnetic Te doped Bi 2 Se 3 reduces the ferromagnetic order of Bi 2−x Cr x Se 3 top layer through a proximity effect. In the ferromagnetic transition region, a weak antilocalization limit exists in the Dirac surface state coupling with non-localized bulk carrier WL interaction accompanying promptly in our magnetotransport experiments. Ex-situ XPS results indicate that an additional electron surface bonding by residual gas molecule further weakens the TI surface state and anomalous Hall effect. In another term, this MTI/TI heterostructure does mimic the same TRS broken bulk insulating state as similar to a trivial insulator system. | 3,540.6 | 2019-11-25T00:00:00.000 | [
"Physics"
] |
Third-party transfers in WLCG using HTTP
Since its earliest days, the Worldwide LHC Computational Grid (WLCG) has relied on GridFTP to transfer data between sites. The announcement that Globus is dropping support of its open source Globus Toolkit (GT), which forms the basis for several FTP client and servers, has created an opportunity to reevaluate the use of FTP. HTTP-TPC, an extension to HTTP compatible with WebDAV, has arisen as a strong contender for an alternative approach. In this paper, we describe the HTTP-TPC protocol itself, along with the current status of its support in different implementations, and the interoperability testing done within the WLCG DOMA working group's TPC activity. This protocol also provides the first real use-case for token-based authorisation for this community. We will demonstrate the benefits of such authorisation by showing how it allows HTTP-TPC to support new technologies (such as OAuth, OpenID Connect, Macaroons and SciTokens) without changing the protocol. We will also discuss the next steps for HTTP-TPC and the plans to use the protocol for WLCG transfers.
Introduction
The primary driver for wide-area data movement for all LHC experiments is bulk data movement between storage services. This bulk data movement serves to pre-stage data to be processed by production systems or to increase data replication to make it more available for analysis. The technique to perform these transfers is third-party copy (TPC); in TPC, a central entity (the 'third party') contacts a source and destination storage endpoint to facilitate a transfer from the source to the destination. This provides for central management and coordination of transfers but allows for data to move directly between the storage systems. The high-level concept is illustrated in Figure 1.
In 2017, Globus announced the retirement of the Globus Toolkit, which served as the reference implementation for GridFTP protocol [3,19]; this has increased interest into a number of alternatives such as HTTP [10]. HTTP is protocol that underpins the World Wide Web, making it one of the most common protocols on the planet -meaning there is a large storage.site1.com storage.site2.com Third-party Step 2: Data is transferred directly between sites on the data channel.
Step 1: Third-party contacts both sides to coordinate transfer on the control channel Figure 1. The basic concept behind TPC: an entity independent of either storage service coordinates direct transfer of data between the source and destination.
community of experts and many mature implementations. Unlike GridFTP, utilizing HTTP does not expose the WLCG community to the risks of relying on a specialized protocol.
The initial work to adopt HTTP as a third party transfer protocol within the WLCG community was outlined in [9], following activities descripted in [14]; since then, the protocol has evolved and matured into what we term "HTTP Third Party Copy" (HTTP-TPC), as described in Section 2. Further, the WLCG community has formed a working group around Third Party Copy as part of the Data Organization, Management, and Access (DOMA) initiatives; as outlined in Section 3, this working group is testing and developing both HTTP-TPC and third party copy extensions for XRootD [15]. This has allowed for the continued growth of the activity -both in maturing implementations so they can be used in production and further evolving the protocol as described in Section 4.
Background
In preparation for the Run 1 of the LHC, a number of transfer protocols were considered by the LHC community before it eventually settled on GridFTP with a reference implementation provided by the Globus Toolkit [4] (at least one other production-quality implementation has been written by the dCache project [13]). This transfer protocol was augmented with the Storage Resource Management (SRM) protocol [23] which helped manage load-balancing between servers and the storage end-points. Overall, the GridFTP protocol has served the community faithfully for nearly 15 years.
By the end of Run 2, several events transpired that motivated the community to re-evaluate its use of GridFTP as a TPC protocol. First, many sites began to retire their SRM endpoints as unique space management features of SRM were largely never used, GridFTP could be used directly and native load-balancing solutions were introduced. Second, the Globus organization's retirement of the Globus Toolkit [11] meant the implementation of GridFTP in use by several of the storage systems had no original developer support. This led to the formation of the WLCG DOMA TPC working group during the WLCG DOMA face-to-face at CHEP 2018, charged with examining alternative options and growing nascent ecosystems.
For the HTTP-TPC, as explained in [9], the key concept is the use of the WebDAV COPY verb. The client sends an HTTP request using COPY to the active endpoint of the transfer Figure 2. The basic mechanism of HTTP-TPC, reproduced from [9]. Here, we illustrate "pull mode", where the third party copy client contacts the destination site and issues a request that the destination downloads (pulls) the data from the source over HTTP. along with another URL in an HTTP header. For pull mode, the active endpoint is given a Source header; for push mode, the active endpoint is given a Destination header. The active endpoint then downloads or uploads, respectively, from the passive endpoint. Features of note for the HTTP-TPC protocol include: • Separation between the "framing" and the "transfer" protocol. The URL sent to the active endpoint does not have to use https://; for instance, the dCache implementation has shown that the HTTP-TPC active endpoint can be given a GridFTP URL to move the data over GridFTP (potentially useful for enabling a transition from GridFTP).
• The active endpoint sends continuous performance markers back to the TPC client, allowing the client to monitor progress (cancelling the transfer as necessary).
• Ability for transfers to be load-balanced using HTTP's built-in redirection response.
• When in 'pull' mode (the active endpoint is the destination), multiple pipelined GET requests can be load-balanced across multiple parallel TCP streams, allowing a single transfer to proceed faster compared to when a single TCP stream is used.
As of March 2020, there are four independent implementations of the HTTP-TPC protocol, in the dCache, DPM, StoRM [1], and XRootD software products. Further, as XRootD often forms the basis of other storage services in the WLCG community, services like EOS also have HTTP-TPC support without needing a separate implementation.
Building the HTTP-TPC Community
We found that the key to maturing the use of HTTP-TPC beyond initial specification and implementation is to build a user community. The primary mechanism thus far has been the WLCG DOMA TPC working group; in the 18 months that followed the work done in [9], this group (with co-leads Bockelman and Forti) has coordinated the development and finalization of HTTP-TPC, helped deploy a test-bed for HTTP-TPC, and organized a testing infrastructure. Within the testing infrastructure, the working group operates three types of tests: nightly, integration, and full scale.
The nightly "smoke tests" [22], demonstrate compliance and functionality with the HTTP-TPC protocol by performing a small transfer against a known working endpoint in Figure 3. Example graph from the CERN monitoring system displaying the number and volume of HTTP-TPC transfers in the scale test-bed per 6 hour interval. Note that maintaining 20TB of data movement every 6 hours for a week adds up to 560TB of data transferred. addition to simple tests for acquiring a transfer token from the endpoint (Section 4). These tests are driven by a simple script and are meant to be easily reproducible by developers or administrators.
As of March 2020, the HTTP-TPC 'smoke test' test-bed has 49 participating endpoints. To catch bugs and issues as early in the development process as possible, we include endpoints from across the full software development life cycle: from endpoints on production sites to integration test-beds to developer instances. This approach allows us to test across multiple versions (e.g., stable releases transferring against the latest nightly builds) in addition to across multiple implementation.
The continuous integration tests the 'transfer matrix' between any two protocol endpoints in the system at a small-scale (2-3 GB per hour), allowing the working group to understand behavior at a modest scale and to monitor for a broader set of pairwise issues. These are driven by a dedicated instance of Rucio [20], the data management solution used by the ATLAS experiment. Unlike the smoke tests, these integration tests include the full stack of WLCG transfer utilities. While failures are more difficult to reproduce in this environment, these tests are far more representative of a production transfer activity. For example, failures due to the interaction of the FTS3-based [7] TPC client and the storage service may require a full FTS3 server to reproduce.
Finally, full-scale tests are also driven by Rucio; other than scale, these are identical to the continuous integration tests. These scale tests are driven by a driver script that uploads a randomly-generated one terabyte dataset to an endpoint, then generates a rule to replicate the dataset to all other endpoints in the system. Once the dataset is completely transferred, the driver script triggers a deletion of the replicated datasets; after the deletion is complete, the same rule is installed again. With this setup, we have demonstrated the ability to transfer approximately half a petabyte of data per week. See Figure 3. This allows us to monitor for issues at scales similar to the production system.
Evolving Approach
One of the strongest motivations for using HTTP as the base protocol in TPC is that it allows for a number of authorization schemes. The WLCG has historically used the Grid Security Infrastructure (GSI) [12] with various extensions; at its core, GSI is based on X.509 PKI. As GSI is based on X.509, its infrastructure can be used to associate a TLS session for HTTPS and authenticate a client and the TLS level. Previously in [21], work was done to delegate a grid proxy to the HTTPS-based active endpoint; with the client's delegated identity, the active endpoint could then authenticate its transfer request to the passive endpoint.
However, this use of GSI suffers from the same issue as GridFTP: the retired Globus Toolkit was the reference implementation. The WLCG DOMA TPC working group has instead settled on using bearer tokens to authenticate transfers. A number of bearer token based schemes have been proposed, including SciTokens [24] and the WLCG Common JWT profile [5]; these will eventually allow a transfer to be performed completely without the use of a X.509 client credential. As a transition mechanism, we have defined a way for clients to request a token from the storage endpoint, provided the request is made over a HTTPS connection that is GSI-authenticated.
First, the client must perform OAuth2 metadata discovery [18] against the storage endpoint to determine the associated (storage-specific) token endpoint. Then, an access token request is made against the token endpoint using the client credentials flow [16]. Unlike a typical client credential flow, when the client authenticates via an HTTP header (such as the Authorization header), this request must be done over the GSI-authenticated HTTPS channel. Based on this token request and the client's authorization, the storage endpoint will issue an access token that can be used as part of the HTTP-TPC infrastructure. Although the token format is considered opaque (implementations have been done both based on JWT [17] and Macaroons [8]), the client must ask for one or more of an agreed-upon set of scopes of the form $ACTIVITY:$PATH. If permitted, the returned token will permit the bearer to perform the specified activity ($ACTIVITY) for any resource inside the normalized $PATH. The defined authorizations (based upon the work done in dCache for its initial Macaroon support [6]) are: • UPLOAD: Authorization to create new and upload contents, provided that existing data at the endpoint is not altered.
• DOWNLOAD: Authorization to read data.
• DELETE: Authorization to delete resources from the endpoint.
• MANAGE: Change file metadata at the storage endpoint and perform operations that may overwrite existing data.
• LIST: List the contents of a directory resource at the storage endpoint.
Beyond authorization, the experience gained in the WLCG scale tests has shown that, while HTTP-TPC can be run in either push or pull mode, pull mode has become preferred. Pull mode allows the active endpoint to download with the HTTP GET requests; as GET is idempotent, the endpoint can issue numerous requests in a pipeline or partition them over a number of TCP streams to improve overall throughput. Further, the active endpoint is the most natural entity in the system to manage a queue of transfer requests. Given the ability to write to disk is considered a more scarce resource than reading, sites have preferred the pull mode.
Conclusions
Not only is third party copy an essential technique in the WLCG infrastructure, it is how the majority of the LHC data is transferred. Over the past several years, the HTTP-TPC protocol has emerged as viable replacement for the venerable GridFTP protocol. While the broad outlines of the protocol have been used by the community for years, the protocol has evolved based on operational experience and the community evolution away from using X.509 client credentials. These activities have been led by the WLCG DOMA TPC working group which organizes several types of test-beds. A key activity in the coming months will be to evaluate fully token-based data transfers, with authorization following the rules defined by the WLCG JWT profile [5] and leveraging the integration with the WLCG IAM token issuer [2].
As we go beyond tests, LHC experiments are beginning to consider the use HTTP-TPC in their production infrastructure; we expect to see significant, at-scale tests in the lead-up to Run 3. | 3,171 | 2019-11-05T00:00:00.000 | [
"Computer Science"
] |