Upload 5887 files

Med-MDPI/biomedinformatics/biomedinformatics-01-01-00001.txt

@@ -0,0 +1 @@
+With this volume, the peer-reviewed open access journal Biomedinformatics published online on the website https://www.mdpi.com/journal/biomedinformatics, and bearing the current International Standard Serial Number ISSN 2673-7426 enters the scientific community.At the beginning of the 3rd decade of the 21st century, this new journal is dedicated to research reports in the field of biomedical informatics. Biomedinformatics appears at a time when computational methods have reached clinical practice and the transformation to digital medicine is accelerating. Both digitized healthcare and bioinformatics-based research is producing and benefiting from increasingly complex data. This requires the development of tools and methods to extract information from these data and translate it into new knowledge. While biomedical research continues to require clinical and experimental data collection, digital healthcare research has clearly evolved from a collection of supporting methods to an equivalent scientific approach, enabling a paradigm shift from almost exclusively hypothesis-driven approaches to increasingly data-driven biomedical research. Indeed, computational science is a rapidly growing multidisciplinary field that uses advanced computational capabilities to understand and solve complex problems by applying new methods of computational intelligence, machine learning, and advanced statistics [1].Biomedical research is currently confronted with the emergence of a plethora of new journals, of which only those with peer-review and as soon as possible PubMed listing deserve serious attention. This leads to the dilution of the sources of scientific information as evidenced by a scientometric study that found that the relationship between impact factor and citations has weakened since 1990 [2]. A greater proportion of the 5% and 10% most cited papers were published outside of journals with top 5% and top 10% impact factors, respectively. As research activity increases and the number of papers that can be published in a journal is limited, either the rejection rate must increase or the number of journals must increase, which is what is happening. As a result, new journals are increasingly publishing highly cited reports.In particular, the proportion of bioinformatics-related topics among biomedical research papers is steadily increasing (Figure 1). A search of the PubMed database at https://pubmed.ncbi.nlm.nih.gov on 12 December 2020, using the search string “((((bioinformatics) OR (biomedinformatics) OR (medical informatics) OR (biomedical informatics) OR (data science[TIAB]) OR (data-science[TIAB]))) NOT (review[PT]))” and the R library “RISmed” (https://cran.r-project.org/package=RISmed [3]) returned 35 hits of papers published in 1950 and 57,368 hits of papers published in 2019. The total number of publications listed in PubMed that were queried using only the year as the search string (YEAR[EDAT]: YEAR[EDAT]) was 85,787 in 1950 and 1,229,451 in 2020. These figures show an absolute increase in the number of bioinformatics-related publications; however, their relative share of all the publications listed in PubMed also increased from 0.041% in 1950 to 4.7% in 2019. Although this editorial is written at a time when everyone is looking at the exponential growth of medical indicators, the evolution of biomedical publications, including the absolute and relative proportions of bioinformatics-related papers, is better described by bilinear growth when trying sliding breakpoints and judging goodness-of-fit by the Akaike information criterion [4]. The relative share of bioinformatics-related publications in all publications already shows an accelerated increase from 1990 onwards, and from 1998 onwards this topic also shows an absolute acceleration of publication activities.Over the last 70 years, the above search provided 77 countries of origin for biomedinformatics-related papers in the PubMed database (Figure 2); however, only the affiliation of the first author was considered, which may underestimate collaborative contributions from other countries, with the United States of America being both the country with the most publications and the top country for collaborative publications [7]. The US occupies indeed the largest spot in a cartogram of the world’s bioinformatics-related publications (Figure 3a), i.e., in a thematic map, in which distortion is used to convey information, for example, by distorting the outline polygons of all countries in such a way that the areas are proportional to the number of publications [8].Indeed, the raw number of bioinformatics publications is significant for the impact on the world’s knowledge. But the figure also shows the trend toward globalization. When the number of publications per year is standardized to the respective country’s population in that year according to the United Nations Department of Economic and Social Affairs, Population Division (World Population Prospects 2019, Online Edition. Rev. 1, downloaded on December 14 from https://population.un.org/wpp/Download/Standard/Population/), the weight shifts toward Western Europe, with Ireland and the Benelux countries having the largest share of biomedical informatics-related publications per capita (Figure 3b). Using only the last five full years does not yet significantly change this picture, but it is likely that the weights will continue to shift. This is reflected in the diverse editorial board with broad expertise in biomedical informatics and computational biology and medicine from many countries on three continents so far, including North America, Europe, and Asia, and a completion of this list should be a matter of only a short term.The Biomedinformatics journal has just started its online presence with a first article and the next goal will be to establish the journal as a visible publication platform, implying its inclusion in the PubMed database as the primary collection of references and abstracts on life science and biomedical topics. The journal aims to produce truly multidisciplinary publications that meet a high international standard, based on a competitive peer review process that has been established from the beginning. The editors are committed to making it a success. The next immediate goal is inclusion in the PubMed database as the standard collection for scientific papers in the biomedical research field.This research received no external funding.The author declares no conflict of interest.Publications listed in the PubMed database that are annotated with bioinformatics-related topics, based on the year of publication and the total number of biomedical publications listed. (a): stacked bar chart of all publications listed in PubMed per year with special focus on the publications found with the search string “((((bioinformatics) OR (biomedinformatics) OR (medical informatics) OR (biomedical informatics) OR (data science[TIAB]) OR (data-science[TIAB]))). NOT (review[PT]))”. (b): scatterplot of publications on biomedical topics. The line shows the bilinear trend in the number of publications with acceleration from the late 1990s. (c): scatterplot of the relative number of publications on biomedical topics from all PubMed publications, also showing a bilinear trend. The figure was created using the R software package (version 4.0.3 for Linux; http://CRAN.R-project.org/ [5]) and the library “ggplot2” (https://cran.r-project.org/package=ggplot2 [6]).Publications listed in the PubMed database and annotated with bioinformatics-related topics, based on the year of publication and the country of work of the first author. The matrix plot shows the number of publications color-coded after zero-invariant log transformation. A darker color indicates more publications. On the right, the sum of publications per country over the last 70 years is shown as a bar chart. The figure was created using the R software package (version 4.0.3 for Linux; http://CRAN.R-project.org/ [5]), and the R library “Complex Heatmap” (https://bioconductor.org/packages/release/bioc/html/ComplexHeatmap.html [9]).Cartograms [8] showing publication activity on biomedical informatics-related topics listed in the PubMed database between 1950 and 2019. (a): publication activity per county plotted with Gaussian blur as described in [8] spatial plots, with the boundaries of the regions transformed to be proportional to the publication numbers. (b): the same type of mapping, but plotting publications per million inhabitants. The figure was created using the R software package (version 4.0.3 for Linux; http://CRAN.R-project.org/ (R Development Core Team, 2008)) and the libraries “ggplot2” (https://cran.r-project.org/package=ggplot2 [6]) and “Rcartogram” (https://github.com/omegahat/Rcartogram [10]).Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-01-00002.txt

@@ -0,0 +1 @@
+Introduction: Potential benefits of implementing an electronic health record (EHR) to increase the efficiency of health services and improve the quality of health care are often obstructed by the unwillingness of the users themselves to accept and use the available systems. Aim: The aim of this study was to identify factors that influence the acceptance of the use of an EHR by physicians in the daily practice of hospital health care. Material and Methods: The cross-sectional study was conducted among physicians in the General Hospital Pancevo, Serbia. An anonymous questionnaire, developed according to the technology acceptance model (TAM), was used for the assessment of EHR acceptance. The response rate was 91%. Internal consistency was assessed by Cronbach’s alpha coefficient. A logistic regression analysis was used to identify the factors influencing the acceptance of the use of EHR. Results: The study population included 156 physicians. The mean age was 46.4 ± 10.4 years, 58.8% participants were female. Half of the respondents (50.1%) supported the use of EHR in comparison to paper patient records. In multivariate logistic regression modeling of social and technical factors, ease of use, usefulness, and attitudes towards use of EHR as determinants of the EHR acceptance, the following predictors were identified: use of a computer outside of the office for reading daily newspapers (p = 0.005), EHR providing a greater amount of valuable information (p = 0.007), improvement in the productivity by EHR use (p < 0.001), and a statement that using EHR is a good idea (p = 0.014). Overall the percentage of correct classifications in the model was 83.9%. Conclusion: In this research, determinants of the EHR acceptance were assessed in accordance with the TAM, providing an overall good model fit. Future research should attempt to add other constructs to the TAM in order to fully identify all determinants of physician acceptance of EHR in the complex environment of different health systems.The Serbian law on medical records and records in the field of health defines an electronic health record (EHR) as a set of data on one person that combines all health information relevant to one’s long-term health status [1]. The EHR would, according to necessity, authorize the future provision of health care so that the patient has a greater chance of being successfully treated [1]. The EHR is an electronic health card, which is available on all hospital computers, from the first provider of healthcare to all hospital departments [2]. Assuming that the EHR can improve the quality and effectiveness of health care, many health organizations in developed countries have invested in the development and dissemination of such systems [3,4].Implementation of an EHR increases the efficiency of health services, improves the quality of health care, and positively affects the level of user satisfaction [5,6]. However, the potential benefits of using computer performance are often obstructed by the unwillingness of the users themselves to accept and use the available systems [7]. To date, different methods have been developed in an attempt to understand how end users ultimately decide whether or not to use new technology. The theory of reasoned action (TRA) [8] and the theory of planned behavior (TPB) [9] originate from behavioral sciences, while the technology acceptance model (TAM) [10] and the unified theory of acceptance and use of technology (UTAUT) [11] come from the field of information technology. Adjustments of original methods were also done, with the intention of determining the success or failure of the implementation of new information technologies in the health care environment [12,13].The technology acceptance model (TAM) is one of the methods for identifying factors determining whether health professionals will use health information technology, and it can partly explain the success or failure of any attempted application process [14]. The TAM model was first developed by Fred Davis in 1986 [10]. This model appears to be particularly applicable in the health information technology field because it focuses on specific variables believed to influence the use of information technology. The TAM distinguishes ease of use, perceived usefulness, and attitudes towards using as the factors which most influence the adoption of new technologies [10].According to Davis [10,15], the perceived ease of use, which indicates how difficult the person believes the proposed system would be to use, has a direct influence on perceived usefulness, which is defined as a person’s belief that using a system increases their efficiency. Lastly, these two factors impact the users’ overall attitude towards using the given system, as a major determinant of whether or not the user actually uses it. Additionally, system design characteristics are also thought to have an effect on perceived usefulness and perceived ease of use [16]Van der Meijden et al. [17] identified user resistance as one of the primary factors leading to unsuccessful EHR implementation. Despite its potential advantages, implementation of the EHR system can face resistance if the users themselves are not satisfied with the system [18]. According to the results of previous studies, the slow adoption of the EHR is affected by the strong resistance among physicians as the main users [19]. Regardless of whether they support the use of EHR or not, physicians will always have a major impact on the other users, such as nurses and administrative staff. Therefore, it is essential by the physicians themselves to understand the potential barriers to implementation of EHR [5]. The aim of this research was to identify factors influencing the acceptance of EHR by physicians in the daily practice of hospital health care.The cross-sectional study was conducted among physicians in the General Hospital Pancevo in November 2015. Physicians who, at the time of the study, were out of the hospital due to specialization and those who did not still work with the EHR due to the territorial separation of the hospital departments were excluded. As the instrument for this study, a tailored questionnaire developed according to the TAM was used for the assessment of EHR acceptance among the physicians. Additional statements were added for the purpose of this research, based on a literature search of the validated questionnaires with similar content [14,15]. Out of 172 questionnaires distributed to physicians, 156 questionnaires were submitted, giving a high response rate of 91%. The questionnaire was divided into several parts, which proffered information on various aspects of the use of EHR: (1) general data (age, sex, years of employment, previous experience, specialization, and hospital department), (2) use of computers (in daily life, for reading professional literature, social networking, reading newspapers, information gathering, previous computer training, and need for additional training), (3) technical performance satisfaction (interface and browsing the patient records, terminology, and EHR capability), (4) EHR ease of use, (5) EHR usefulness, (6) attitudes towards use, and (7) acceptance of the EHR (Figure 1). Questions related to the use of computers had “yes” or “no” answers, for statements regarding daily use of computers and five levels of agreement were used (Likert scale), otherwise 7 possible levels of agreement were offered, from level 1 (very strongly disagree), 4 (neutral position) to 7 (very strongly agree). The following statements were used for assessing the technical performance of the EHR: Interface and browsing the patient records—”letters on the screen are easy to read”, “commands and functions are understandable”, “personal files are transparent”, “accessibility of the patients files is better than in paper patient records”, “time required to switch from page-to-page of the personal file is satisfactory”, “not too many steps are required for certain tasks”, and “steps are in logical order”. Terminology—”the professional terminology is adequate”, “computer terminology is clear”, “error messages are useful”, and “clarification of the commands and functions provided by the computer are helpful”. EHR capability—”speed of the EHR system is satisfactory”; “the EHR system is always reliable (the system rarely fails)”; “EHR always alerts me to potential problems”; “mistakes made in personal files can be easily corrected”; “the system allows the cancellation of a given operation”; and “when technical problems arise, the problem is quickly resolved”. Statements used in accordance with the TAM were as follows: Ease of use—”I find it easy to get the EHR to do what I want it to do”; “interacting with the EHR requires a lot of my mental effort”; “I find that it was easier to work with a paper patient record”; and “overall, I find that improvements are needed to simplify the use of the EHR. Usefulness—”using EHR improves the quality of the work I do”, “using EHR enhances my effectiveness on the job”, “using EHR reduces errors in prescribing medications”, “using the EHR reduces errors in patients’ identification”, “using EHR reduces errors in diagnosis coding”, “using the EHR provides a greater amount of valuable patient information”, “using EHR shortens the time spent on administrative tasks, “using EHR increases my productivity”, “using EHR improves communication with colleagues and superiors”, “using EHR allows me to stay up-to-date with my work”, and “overall, I find the EHR useful in my job”. Attitudes towards use—”using EHR is a good idea”, “using EHR is a wise idea”, “I like the idea of using EHR, and “using EHR would be pleasant”. Acceptance of the EHR was assessed through the following statement: “the EHR is better than the paper patient record”.Data are expressed as mean values with standard deviations or as medians with ranges. Categorical data are presented by absolute numbers with percentages and analyzed using a Chi-square test and Fisher’s exact test. For continuous variables, Student’s t test or the Mann–Whitney U-test (according to data distribution) was used. A logistic regression analysis was used to identify the factors influencing the acceptance of the use of the EHR. Categories were grouped as “disagreement” for levels 1 to 4 (very strongly disagree, strongly disagree, disagree, and neutral attitude) and “agreement” for levels 5 to 7 (agree, strongly agree, and very strongly agree). The dependent variable was the physicians’ acceptance of the EHR, assessed by the agreement that the current version of the EHR was better than the paper patient records. When deciding whether physicians prefer the EHR over traditional paper records answers from level 1 to 4 were considered negative, while answers from level 5 to 7 were affirmative. All significant variables from modeling social and technical factors, ease of use, usefulness, and attitudes towards use of the EHR were entered in the final modeling of the determinants of EHR acceptance by forward wald stepwise procedure. To assess the internal consistency of questionnaire, the Cronbach’s alpha coefficients were calculated (ranges from 0–1, the latter meaning perfect reliability). Differences were considered significant at a p value of <0.05. Statistical analysis was performed using SPSS statistical software (SPSS for Windows, release 21.0, SPSS, Chicago, IL, USA).The mean age in the studied population was 46.4 ± 10.4 years, 58.8% participants were female. The years of employment ranged from 1 to 39 years. Most physicians (83.3%) were specialists, while general practitioners accounted for 16.7%. Thirty four percent were from the surgical department; 25.6% from the internal medicine department; 11.5% from anesthesia; 8.3% from pediatric department; X-ray, CT, and laboratory diagnostics (5.1%); emergency and emergency services (1.9%); and other services (4.5%). Most physicians (72.5%) used computers every day, 22.2% several times a week, while 3.3% did not use computers out of the office. In daily life, physicians used computers to review professional literature (94.6%), for obtaining various types of information (82.4%), reading daily newspapers (93.3%), and for social networking (56.0%). Most of the respondents (68.2%) received training in EHR use at the General hospital Pancevo, and a smaller number (27.3%) were certified at formal computer training. The majority (61.2%) expressed satisfaction with their computer skills, 24.3% were neutral, and 14.5% were not satisfied. One quarter of the respondents stated they need additional training for working with the EHR (24.6%). Most physicians (88.6%) had been working up to 2 years with the EHR.Statements regarding interface and browsing the patient records (the readability, clarity of the commands and functions, transparency of patients files, and better accessibility of the patients files in comparison to paper documentation) were mostly marked with 5 (agree) on the Likert scale (Figure S1). Neutral attitude was often expressed regarding time needed to switch from page-to-page, the number of steps needed for performing certain tasks, and the logical order of the steps (Figure S1). All of the terminology statements were mostly marked with 5 (agree) on the Likert scale (Figure S2). Concerning EHR capability—system alerts, ease of correction, procedure cancelation, and quick resolution—statements were usually graded with 4 (a neutral attitude) (Figure S3). Grade 3 (disagree) was the average response to satisfaction with the speed of the EHR system (Figure S3). The reliability of the EHR system was graded the lowest (1—very strongly disagree) (Figure S3). Two EHR ease-of-use statements, “I find it easy to get the EHR to do what I want it to do” and “I find that it was easier to work with a paper patient record”, were mostly marked with 4 (neutral) (Figure S4). Physicians most often graded with 3 (disagree) the statement “interacting with the EHR requires a lot of my mental effort” (Figure S4). Physicians supported the statement that “improvements are needed to simplify the use of the EHR” with grade 7 (very strongly agree) (Figure S4). Five out of eleven statements, regarding EHR usefulness, were mostly marked with 5 (agree) on the Likert scale—”using EHR improves the quality of the work I do”, “using EHR reduces errors in patients’ identification”, “using EHR reduces errors in diagnosis coding”, “using EHR provides a greater amount of valuable patient information”, and “overall, I find the EHR useful in my job” (Figure S5). Attitudes towards using the EHR were expressed through 4 statements. “Using EHR would be pleasant” was mostly marked with 4 (neutral), whereas the physicians showed agreement (5—agree) with the rest of the three statements (Figure S6). Half of the respondents (50.1%) supported the use of EHR in comparison to the paper patients records (Figure 2). Reliability of the scales for technical performance, usefulness of the EHR, and attitudes towards use of EHR, all exceeded 0.85, and demonstrated high reliability of the scales. The domain “ease of use” presented the weakest Cronbach’s alfa presenting lower reliability of this scale.The socio-demographic variables and technical characteristics of EHR were used for modeling of EHR acceptance by physicians in the daily practice of hospital health care. In the group of socio-demographic factors—age, duration of employment, and use of the computer outside of the office for reading daily newspapers—were identified as significant predictors for the acceptance of the EHR (Table 1). In multivariate analysis, use of the computer outside of the office for reading daily newspapers was identified as the most significant predictor for the acceptance of EHR among social variables (Table 1). Among the statements regarding technical characteristics, which were all shown to be significant in univariate analysis, better accessibility of the patients files and clear computer terminology were identified as significant predictors, in multivariate analysis (p < 0.001 and p = 0.009, respectively) (Table 2).Among ease of use determinants, a significant predictor of the acceptance of EHR was the easiness to get the EHR to do what one wanted (p < 0.001) (Table 3). While all of EHR usefulness determinants were significant in univariate analysis, improvements in productivity, quality of the work using EHR, and providing a greater amount of valuable information were identified as a significant predictors of the acceptance of the EHR in multivariate analysis (p = 0.004, p = 0.032, and p = 0.002, respectively) (Table 4).Among attitude determinants, which were all significant in univariate analysis, statements that “using EHR is a good idea” and “it would be pleasant to use it” were significant predictors in multivariate analysis (p = 0.001 and p < 0.001, respectively) (Table 5).In multivariate logistic regression modeling of social and technical factors, ease of use, and usefulness and attitudes towards use of EHR as determinants of the EHR acceptance, the following predictors were identified: “use of computer outside of the office for reading daily newspapers” (p = 0.005), “EHR providing a greater amount of valuable information” (p = 0.007), “improvement of the productivity by EHR use” (p < 0.001), and “using EHR is a good idea” (p = 0.014) (Table 6). The overall percentage of correct classification in the model was 83.9%.This study aimed to bridge the gap between information technology (IT) departments and the growing demands of expanding technology use in health care and physician IT skills. It is broadly accepted that EHR records are inevitable for future healthcare, yet it is a challenge to make the change and convince leaders to put their organizations through the transformation process and make an effort for EHR implementation [20]. This study was conducted for the first time in the hospital setting of the Western Balkans using TAM. The acceptance of the EHR with physicians was associated with its technical characteristics, ease of use, usefulness, and attitudes towards use of the EHR.According to the literature, the largest barrier to accepting new technologies for the physician was a lack of knowledge of the technology and its complexity [21]. Physicians are not familiar with IT products and believe that implementation of new technology would be overly complicated, or would change the routine of the medical practice. This leads to stress and anxiety, resulting in physicians’ hesitation to use new technology [22]. Other barriers responsible for inhibiting the development and application of IT in the health care system have been identified. Variables related to these barriers include changes in the efficiency of physicians, an inadequate legal framework, a deficiency of explicit standards, interoperability, system implementation problems, breaches in privacy and confidentiality, and insufficient research in this area [23].Satisfaction with the technical performance of the EHR in this study was analyzed using three areas—interfacing and browsing patient records, terminology, and EHR capability. Physicians have expressed greater satisfaction with the interface, but less satisfaction with browsing through patient records. The greatest satisfaction expressed by physicians was easy readability of the screen (74.8%) and the clarity of appointed commands and functions (70.5%). A study by Ludwick and Doucette [24], which analyzed implementation of the EHR in seven countries, showed that quality of the system design and graphical user interface can have an effect on the outcome of implementation. Another great improvement of implementing EHR is better accessibility of the patient files, which can significantly enhance the coordination of care and efficiency of hospital care practice. Regarding the terminology in the EHR, physicians being surveyed were largely satisfied with the professional medical, as well as the computer terminology, which is an integral part of the use of EHR. On the other hand, fewer than half of those surveyed were satisfied with EHR capability. The greatest disagreement was related to reliability of the EHR system, resolution of technical and functional problems, and EHR system speed satisfaction. This was supported by the results from Bloom [25] where physicians were greatly dissatisfied by the time required for documentation management within the EHR. In research regarding the impact of EHR in our region, 55% of respondents answered that the expansion of work due to EHR use has shortened the time spent talking to patients [26].Concerning the ease of use of EHR, almost half of physicians agree with the statement that it is easy to get the EHR to do what one wants it to do (44.4%) and more than half stated that improvements are needed in order to simplify the use of EHR (64%). The majority of physicians in our study expressed satisfaction with their computer skills and most of them use computers daily. However, as with every new technology, there is an inevitable period of adaptation as one quarter of physicians in our study think that they still need additional education to work with the EHR. It is important to keep in mind that the majority of health workers have used a computer only superficially as a workplace tool, and that it cannot be expected that full productivity will be achieved without additional training [27]. Training is therefore the right setting for introduction to the principles of security policy and the rules of using the system. In this sense, well planned and timed training of health care workers to use informational technology in everyday practice can be considered as an operational goal [28]. The importance of the training and its effect on acceptance of technology was recognized in other studies as well, where over 90% of surveyed physicians expressed positive attitudes towards training and reported that it helped them use EHR more efficiently [29]. In addition to the process of implementation, it has been shown that quality of the system is equally important, which means EHR must be flexible, user-friendly, and functional in order to be acceptable [30].Results showing that perceived usefulness provides a reliable prediction for system acceptance were first shown in a study by Schultz and Slevin [31], and later confirmed by Robey [32]. This is consistent with our study results, where a considerable effect of perceived usefulness on physicians’ technology use acceptance was confirmed. Furthermore, Bandura [33] considers that both perceived usefulness and ease of use are important when it comes to predicting user behavior. Yang and Yoo [34] proposed that attitude needs to be an essential part of the TAM because of its crucial effect on system use and acceptance. This was confirmed by our study results, as well as study results conducted in Iran [35], where perceived ease of use and usefulness were shown to have significant value in attitudes toward users’ system acceptance.When it comes to technology acceptance, numerous methods have been developed in order to predict the drawbacks of the implementation and tailor it to users’ needs. One of the models that set the foundations in the field of technology was the theory of reasoned action, developed by Fishbein and Ajzen (1975) [36]. According to this model one’s specific behavior is determined by behavioral intention, which is influenced by two main concepts—prior attitude and the subjective norm regarding this behavior. In a study comparing the TRA and the TAM, while both showed notable results in predicting intention to use a system, the TRA presented limited correlation between subjective norm and behavioral intention variables [10]. As an extension to the TRA, the TPB has the similar approach to this question, but with one additional construct [9]. This improvement has made a more complete model, so in a research by Mathieson, the TPB came up as superior compared to the TAM in some areas such as delivering specific information [37]. Looking at the overall preferences the TAM has remained more attractive because it is simpler, easier to apply, and has modest empirical advantage.A fresher model developed in 2003 [11], the unified theory of acceptance and use of technology, is based on eight different technology acceptance models, with the TAM and TRA among them. It considers four variables, providing more information, which leads to a longer and more complex questionnaire. Questionnaires that are too long exponentially increase the time for respondents and therefore they are more susceptible to careless answers and a lower response rate [38].Despite the fact that the TAM have been used in numerous studies and has been widely accepted as a tool of predicting whether users will accept or decline the system, studies in health care show mixed results. In order to adapt the model, researchers added different variables, as well as revised the original model instruments which shows that the original TAM is not perfect when it comes to predicting technology acceptance in health care [39]. The main criticism is that the TAM focuses mostly on the technology component (ease of use, usefulness) as the crucial factor for acceptance. On the other hand external aspects such as social, cultural, and emotional and also different groups of users are overlooked [40]. Another drawback is that all of the models were designed outside health care settings—the TAM was developed in studies linked to e-mail and word processing systems [39]. This utilization encounters difficulties when it’s transferred to a complex environment such as health care. The systems TAM was developed on involved voluntary use, while implementing technology in health care means they are mandatory, which can induce aversion among users. Answer to this could be incorporating appropriate variables to extend model design in order to adapt it to health care settings. Analyzing all previously mentioned characteristics among these models resulted in the making of the TAM, which was our choice as a foundation for developing a model that would consider personal viewpoint on technology use as well as external factors and, thus create a straightforward survey.Other than the unquestioned benefits of its use, IT can also provoke unwanted effects and resistance from employees [41]. Since IT creates new working environments, it has influenced changes in work culture and systems of employee values so that these changes are perceived and accepted to varying degrees [42]. Research conducted on our physicians has shown that computerization has another side to it. The implementation of IT in the selected health care center, did not proceed without issue, especially in its initial phase. Simultaneous changes in methods of working and renunciation of established routines to acquire new ones is a complex mental process further complicated by the nature of the healthcare organization [26]. The EHR is considered to be the backbone supporting the combination of different information tools that could advance the uptake of evidence into clinical decisions. It also facilitates decision-making and knowledge-exchange among physicians by granting them significant, up-to-date, and timely information [43]. If physicians’ attitudes reflect the notion that they are more eager to use and to spend more time learning the EHR system, it seems more likely for them to adapt EHR technology [20].The TAM is useful in identifying factors influencing physicians from different backgrounds in the use of EHR, but it only partly explained physicians’ EHR adoption. In multivariate logistic regression modeling of social and technical factors, ease of use, usefulness, and attitudes towards use of EHR as determinants of the EHR acceptance, the following predictors were identified: use of computers outside of the office for reading daily newspapers (p = 0.005), the EHR providing a greater amount of valuable information (p = 0.007), improvements in productivity by EHR use (p < 0.001), and the statement that using EHR is a good idea (p = 0.014). Future technology acceptance research should attempt to add other constructs or integrate other theories with the TAM in order to fully identify all determinants of physicians’ acceptance of the EHR in the complex environment of different health systems.The following are available online at https://www.mdpi.com/article/10.3390/biomedinformatics1010002/s1, Figure S1: Interface and browsing the patient records, Figure S2: Terminology, Figure S3: EHR capability, Figure S4: Ease of use, Figure S5: EHR usefulness, Figure S6: Attitudes towards using the EHR.Conceptualization, A.P., N.R., J.P.S., D.A., M.U., M.P., V.P., D.S. (Dragan Spaic), S.M., D.S. (Dejana Stanisavljevic) and N.M.; Data curation, A.P., N.R., J.P.S., D.A., M.U., M.P., V.P., S.R., D.S. (Dragan Spaic), S.M., D.S. (Dejana Stanisavljevic), and N.M.; formal analysis, A.P., N.R., D.A., M.P., V.P., S.R., D.S. (Dragan Spaic), D.S. (Dejana Stanisavljevic), and N.M.; funding acquisition, N.M.; investigation, A.P., J.P.S., M.U., M.P., D.S. (Dragan Spaic), S.M., D.S. (Dejana Stanisavljevic), and N.M.; methodology, A.P., J.P.S., D.A., V.P., S.R., D.S. (Dragan Spaic), D.S. (Dejana Stanisavljevic), and N.M.; project administration, A.P., J.P.S., M.P., S.M., D.S. (Dejana Stanisavljevic), and N.M.; resources, A.P., J.P.S., M.U., M.P., D.S. (Dragan Spaic), S.M., and N.M.; software, A.P., J.P.S., M.U., S.R., S.M., and N.M.; supervision, A.P., J.P.S., V.P., D.S. (Dejana Stanisavljevic), and N.M.; validation, A.P., N.R., J.P.S., D.A., V.P., S.R., S.M., D.S. (Dejana Stanisavljevic), and N.M.; visualization, A.P., N.R., D.A., M.U., V.P., S.R., D.S. (Dejana Stanisavljevic), and N.M.; writing—original draft, A.P., N.R., J.P.S., D.A., M.U., M.P., V.P., S.R., D.S. (Dragan Spaic), S.M., D.S. (Dejana Stanisavljevic), and N.M.; writing–review and editing, A.P., N.R., J.P.S., D.A., M.U., M.P., V.P., S.R., D.S. (Dragan Spaic), S.M., D.S. (Dejana Stanisavljevic), and N.M. All authors have read and agreed to the published version of the manuscript.This research received no external funding.The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Institutional Review Board of General Hospital Pancevo (01-6663/15).Informed consent was obtained from all subjects involved in the study.The data presented in this study are available in article or Supplementary Materials.The authors declare no conflict of interest.Modified structure of the technology acceptance model (TAM).Acceptance of the EHR by physicians in daily practice.Univariate and multivariate logistic regression analysis for socio-demographic factors predicting acceptance of the electronic health record (EHR) in General hospital Pancevo.Univariate and multivariate logistic regression analysis for technical factors predicting acceptance of EHR in General hospital Pancevo.Univariate logistic regression analysis for EHR ease of use, predicting acceptance of the EHR in General hospital Pancevo.Univariate and multivariate logistic regression analysis for EHR usefulness predicting acceptance of EHR in General hospital Pancevo.Univariate and multivariate logistic regression analysis for attitudes towards use of EHR predicting acceptance of EHR in General hospital Pancevo.Final multivariate logistic regression model for predicting acceptance of EHR in General hospital Pancevo.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-01-00003.txt

@@ -0,0 +1 @@
+High expression of the anti-apoptotic TNFAIP8 is associated with poor survival of the patients with diffuse large B-cell lymphoma (DLBCL), and one of the functions of TNFAIP8 is to inhibit the pro-apoptosis Caspase-8. We aimed to analyze the immunohistochemical expression of Caspase-8 (active subunit p18; CASP8) in a series of 97 cases of DLBCL from Tokai University Hospital, and to correlate with other Caspase-8 pathway-related markers, including cleaved Caspase-3, cleaved PARP, BCL2, TP53, MDM2, MYC, Ki67, E2F1, CDK6, MYB and LMO2. After digital image quantification, the correlation with several clinicopathological characteristics of the patients showed that high protein expression of Caspase-8 was associated with a favorable overall and progression-free survival (Hazard Risks = 0.3; p = 0.005 and 0.03, respectively). Caspase-8 also positively correlated with cCASP3, MDM2, E2F1, TNFAIP8, BCL2 and Ki67. Next, the Caspase-8 protein expression was modeled using predictive analytics, and a high overall predictive accuracy (>80%) was obtained with CHAID decision tree, Bayesian network, discriminant analysis, C5 tree, logistic regression, and Artificial Intelligence Neural Network methods (both Multilayer perceptron and Radial basis function); the most relevant markers were cCASP3, E2F1, TP53, cPARP, MDM2, BCL2 and TNFAIP8. Finally, the CASP8 gene expression was also successfully modeled in an independent DLBCL series of 414 cases from the Lymphoma/Leukemia Molecular Profiling Project (LLMPP). In conclusion, high protein expression of Caspase-8 is associated with a favorable prognosis of DLBCL. Predictive modeling is a feasible analytic strategy that results in a solution that can be understood (i.e., explainable artificial intelligence, “white-box” algorithms).Diffuse Large B-cell Lymphoma (DLBCL) is one of the most frequent non-Hodgkin lymphomas (NHLs) in western countries. DLBCL accounts for an approximate 25% of NHLs and is characterized by being heterogeneous from a clinicopathological point of view, including histological morphological features, genetic changes and biological characteristics [1,2,3]. Within the category of DLBCL there are several distinct subtypes that are separated, such as the T cell histiocyte rich large B cell lymphoma, the primary DLBCL of mediastinum, the intravascular lymphoma and the lymphomatoid granulomatosis [2]. The prognosis of DLBCL is variable, and with current treatment the disease is curable in 50% of the cases [2,4]. As DLBCL is heterogeneous, it is necessary to identify biomarkers with prognostic value.The prognosis of DLBCL correlates with the International Prognostic Index (IPI) score, which includes the factors of the age, the serum lactate dehydrogenase, Eastern Cooperative Oncology Group (ECOG) performance status, the clinical stage and the number of extranodal disease sites [5,6,7,8]. A variation of the original IPI that incorporates more detailed information about these used clinical variables is the National Comprehensive Cancer Network (NCCN)-IPI [9]. In this research both IPIs will be used.The molecular genetics has also managed to stratify the patients according to their prognosis. The gene expression profiling identified three groups according to the postulated cell-of-origin: the germinal center B-cell type (GCB), the activated B-cell type (ABC), and the unclassified. The Hans’ algorithm also identifies the GCB and the non-GCB (ABC) groups, but is based on a stepwise progression of 3 immunohistochemical markers of CD10, BCL6 and MUM1 (IRF4) [10]. Other prognostic markers are the cytogenetic abnormalities of the MYC, BCL2 and BCL6 oncogenes [11,12,13,14,15,16,17,18,19], M2-like tumor-associated macrophages (M2-like TAMs) [20,21] and RGS1 (among others) [22].In comparison to the GCB, the ABC subtype is characterized by a more aggressive clinical evolution and constitutive activation of the anti-apoptotic nuclear factor kappa B (NF-kB) pathway [23,24,25,26]. We have recently described the prognostic value of a negative mediator of apoptosis in DLBCL, the tumor necrosis factor alpha-induced protein 8 (TNFAIP8) [27,28]. In this research, we had used artificial intelligence—the multilayer perceptron neural network—to analyze the gene expression of the DLBCL series of the Lymphoma/Leukemia Molecular Profiling Project (LLMPP) and to identify the genes that were associated with the overall survival of the patients. The TNFAIP8 was identified within the top 20 most relevant genes of the LLMPP series. Then, we validated the importance of TNFAIP8 by immunohistochemistry and by digital quantification using a machine-learning Weka-based segmentation method in a series of DLBCL from Tokai University Hospital, and we confirmed that high TNFAIP8 was associated with a poor overall survival of the patients [28]. TNFAIP8 acts as a negative mediator of apoptosis and may play a role in tumor progression. TNFAIP8 suppresses the TNF-mediated apoptosis by inhibiting Caspase-8 activity but not the processing of procaspase-8, subsequently resulting in inhibition of BID cleavage and Caspase-3 activation [29,30,31].In our previous research, we quantified the protein expression of TNFAIP8 in a series from Tokai University Hospital and we also correlated with two markers related to the proliferation cycle, the Ki67 and MYC. We found that through immunohistochemistry, the expression of TNFAIP8 was associated with a poor survival of the patients and also positively correlated with Ki67 and MYC in a moderate manner. Nevertheless, in our previous work we had the limitation of not knowing how in DLBCL the TNFAIP8 expression correlated with the apoptosis pathway (Caspase-8, Caspase-3, PARP), which is the main function of TNFAIP8. In Figure 1 the protein–protein interactions of TNFAIP8 are shown. These interactions highlight the apoptosis (including Caspase-8), cell cycle and the p53 signaling pathways. In addition, in our previous research our correlations included only a linear analysis, and more complex nonlinear analyses (that may fit better in the biological processes) had not been performed. Statistics and machine learning differ in their aim: statistical models infer relationship between variables. Conversely, machine learning is designed to make the most accurate predictions.The purpose of this research was to analyze the expression of Caspase-8 (CASP8) in DLBCL. A series of DLBCL from Tokai University was immunostained with Caspase-8 and the protein expression was quantified by digital image analysis, and other markers of the Caspase-8 pathway including BCL2, cCASP3, CDK6, E2F1, LMO2, MDM2, MKI67, MYB, MYC, cPARP and TP53 were analyzed as well. We performed statistics and machine learning analyses to investigate the correlations between them and with the clinicopathological characteristics of the samples. Then, we also used the multilayer perceptron neural network analysis to identify other genes related to CASP8 using the LLMPP dataset. We found that high expression of Caspase-8 was associated with a good prognosis of the patients.The DLBCL series of Tokai University Hospital is comprised of 97 cases, collected from the years 2006 to 2011. The clinicopathological characteristics are shown in Table 1. In summary, the male/female ratio was 54/43 (1.3) and the age ranged from 14 to 97 years, with a median of 67 and a mean of 64.2 ± 14.5. According to the International Prognostic Index (IPI), 38.3% of the patients were low, 30.9% were low-intermediate, 17.3% were high-intermediate and 13.6% were high. Serum IL2R was high in 77% of the cases and B symptoms were present in 24% of the cases. The location was nodal (including the spleen) in 55% of the cases. The treatment was RCHOP or RCHOP-like in 93.4% of the cases. Clinical response was achieved in 74% of the patients. The pathological characteristics showed that the cell-of-origin was non-GCB in 67% of the cases, and the immune phenotype was CD5+ in 15%, CD10 in 30%, MUM1+ in 79%, BCL2+ in 79% and BCL6 in 67% of the cases. The immunohistochemical expression of Regulator of G-protein signaling 1 (RGS1), which is a marker associated with the chemotaxis of B-lymphocytes, with the germinal centers formation and with a poor prognosis of DLBCL [22,32], was high in 54% of the cases. The clinicopathological variables associated with the overall survival of the patients are shown in Table 1. Relevant variables were the IPI, sIL2R, Epstein-Barr virus infection and the cell-of-origin molecular classification according to the Hans’ classifier [10].We used the series of the LLMPP for gene expression analysis [33,34]. This series, the GSE10846, is a robust and well annotated series of 414 cases of DLBCL from Western countries that is publicly archived and available for downloading at the webpage https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE10846 (accessed on 16 April 2021). This series was last updated on 25 March 2019 (contact: Prof. Louis M. Staudt, Center for Cancer Research, National Cancer Institute, Building 10, Room 5A02, Bethesda, MD 20892, USA).The clinicopathological features of this series are shown in detail in the Table 2. In summary, the male/female ratio was 224/172 (1.3) and the age ranged from 14 to 92 years, with a median of 62.5 and a mean of 61 ± 15.5. The 5 and 10-years overall survival of the patients was 57% and 47%, respectively. The variables with prognostic value for the overall survival included, among others, the National Comprehensive Cancer Network International Prognostic Index (the enhanced NCCN IPI) and the cell-of-origin molecular subtypes of germinal center B-cell (GCB), activated B-cell (ABC) and unclassified types (Table 2). This series is comparable to the one from Tokai University Hospital.The immunohistochemical procedures were performed using formalin-fixed paraffin-embedded tissue sections of the lymphoma samples. The immunostaining was performed in a fully automated stainer for immunohistochemistry and in-situ hybridization (Leica Biosystems Bond-Max, Leica K.K., Tokyo, Japan), including the manufacturer’s ancillary reagents and consumables such as the Dewax solution (AR9222), Wash solution (AR9590), Bond epitope retrieval solution 1 and 2 (AR9961 and AR9640) and Polymer refine detection (DS9800). The staining process included the following steps: dewaxing, antigen retrieval, peroxide block, post-primary, polymer, DAB and hematoxylin. The mounting was performed in a Leica CV5030 coverslipper. The slides were visualized in an Olympus BX53 upright microscope, with a DP74 digital camera and cellSens imaging software (Olympus LifeScience, Olympus K.K., Tokyo, Japan). The whole slides were also digitalized using a Hamamatsu digital slide scanner, the NanoZoomer S360, and visualized with the NDP.view2 Viewing software (Hamamatsu Photonics K.K., Hamamatsu, Japan). The representative areas of each marker were stored as a jpeg image for futher digital image quantification using the Fiji (ImageJ) image processing package, in a RGB and threshold strategy as we have recently described [28].The primary antibodies that were used in the immunophenotype were the following: CD3e [1:200, clone LN10, Novocastra (NV), Leica K.K., Tokyo, Japan], CD5 (1:400, 4C7, NV), CD20 (1:200, L26, NV), CD10 (1:100, 56C6, NV), MUM1 (1:100, IRF4, EAU32, NV), BCL2 (1:400, bcl2/100/D5, NV), BCL6 (1:100, LN22, NV) and RGS1 (1:100, Rabbit polyclonal, Thermo Fisher Scientific K.K., Yokohama, Japan). More than 30% expression of the tumoral B-lymphocytes of the DLBCL was assessed as positive. The presence of Epstein–Barr virus was also tested (EBER in-situ hybridization #PB0589, Leica K.K.) and the molecular characterization included the gene translocation status of BCL2 and MYC by FISH (split probes, #Y5407 and #Y5410, Dako/Agilent), and the MYD88 (L265P) mutation assessment [22,35,36].The target markers of this research were Caspase-8, BCL2, cleaved Caspase-3, CDK6, E2F1, LMO2, MDM2, Ki67, A-B-C MYB, MYC, cleaved PARP, TP53 and TNFAIP8.The primary antibodies and the staining conditions for the target markers were the following: Caspase-8 (active subunit p18, found in caspases 8a, 8b and 8h)(1:30, 11B6, NCL-CASP-8, NV), BCL2 (1:400, mouse monoclonal, bcl2/100/D5, NV), cleaved Caspase-3 (Asp175) [1:300, rabbit polyclonal, #9661, Cell Signalling (CST)], CDK6 [1:5, mouse monoclonal, 98D, Monoclonal Antibodies Unit, Spanish National Cancer Research Center (CNIO), Madrid, Spain], E2F1 (1:14, rat monoclonal, Agro368V, CNIO), LMO2 (1:10, mouse monoclonal, 299B, CNIO), MDM2 (1:50, mouse monoclonal, IF2, Invitrogen K.K., Tokyo, Japan), Ki67 (RTU, mouse monoclonal, MM1, NV), A-B-C MYB (1:10, rat monoclonal, DANI51, CNIO), MYC (1:50, rabbit monoclonal, Y69, Abcam K.K., Tokyo, Japan), cleaved PARP (Asp214) (1:100, rabbit monoclonal, D64E10, CST), TP53 (1:100, DO-7, NV) and TNFAIP8 (1:30,000, mouse monoclonal, #14559-MM01, Sino Biological, Beijing, China).Comparisons between groups was performed when needed using non-parametric tests, with the Mann–Whitney U test or the Kruskal–Wallis test, and with crosstabulations that included the Pearson Chi-Square, The Fisher’s Exact test, and the Likelihood Ratio test. Correlations between two quantitative variables were performed with Pearson and Spearman correlations. Binary logistic regression was performed to calculate Odds Ratios and to correlate the expression of Caspase-8 (as dichotomic variable) and the rest of clinicopathological variables (also as dichotomic variables).The definition of overall and progression-free survivals were the standards as described by Cheson BD et al. [37,38]. The overall survival was calculated from the time of diagnosis to the time of the death or the last follow-up. The Kaplan–Meier analysis with the Log rank test was used to calculate survival times, as well as for group comparisons; and the analysis included the Breslow and Tarone–Ware tests when necessary. Survival analysis was also performed with the Cox regression (enter method). The significance threshold was set a priori at p < 0.05.Several software were used in this research according to the manufacturer’s instructions: R software for statistical computing version 3.6.3 (https://www.r-project.org/ (accessed on 29 February 2020)) and the integrated development environment R Studio (version 1.3.959; https://www.rstudio.com/products/rstudio/#rstudio-desktop (accessed on 16 April 2021)), the Gene set enrichment analysis software (GSEA 4.1.0, build: 27, Broad Institute, Cambridge, MA, USA; https://www.gsea-msigdb.org/gsea/index.jsp (accessed on 16 April 2021)), IBM SPSS statistics (IBM Corp. Released 2019. IBM SPSS Statistics for Windows, Version 26.0. Armonk, NY, USA: IBM Corp; https://www.ibm.com/jp-ja/analytics/spss-statistics-software (accessed on 16 April 2021)), IBM data mining and predictive analytics (Modeler version 18), Xlstat (version 2018.1, Addinsoft, USA; https://www.xlstat.com/ja/solutions/premium (accessed on 16 April 2021)), Excel (version 16.0.13127.21062, Microsoft, Redmond, WA, USA; https://www.microsoft.com/ja-jp/microsoft-365/excel (accessed on 16 April 2021)) and EditPad Lite (version 8.1.2 x64, Just Great Software Co. Ltd., Rawai Phuket, Thailand; https://www.editpadlite.com/ (accessed on 16 April 2021)). The IBM SPSS Statistics documentation can be found in the following link: https://www.ibm.com/support/knowledgecenter/en/SSLVMB_26.0.0/statistics_mainhelp_ddita/spss/base/overvw_auto_0.html (accessed on 16 April 2021). The statistics algorithms are found at https://www.ibm.com/support/pages/node/874712#en (accessed on 16 April 2021) and ftp://public.dhe.ibm.com/software/analytics/spss/documentation/statistics/26.0/en/client/Manuals/IBM_SPSS_Statistics_Algorithms.pdf (accessed on 16 April 2021). The IBM Modeler can be accessed in the following link: http://127.0.0.1:57379/help/index.jsp?topic=/com.ibm.spss.modeler.help/clementine/clem_intro.htm (accessed on 16 April 2021). A package for survival analysis in R can be accessed at https://cran.r-project.org/web/packages/survival/vignettes/survival.pdf (accessed on 16 April 2021). The Multilayer Perceptron (Figure 2a–d) and Radial Basis Function analysis using the immunohistochemical data was performed following the manufacturer’s instructions, and as we have thoroughly described in our previous publications [39,40]. In this neural network analysis, the prediction of Caspase-8 by the other related markers of the pathway was performed using the immunohistochemical data of Caspase-8 as a dichotomic variable (high vs. low, with the same cut-off of the overall survival).The LLMPP dataset was downloaded from the Gene Expression Omnibus (GEO) repository located on the National Center for Biotechnology Information (NCBI) webpage. The gene expression data of the GSE10846 was normalized and log2 transformed. The probes were collapsed according to the maximum probe values. Therefore, each gene had one expression value and the final series was comprised of a total 20,684 genes and 414 cases. Using an Artificial Intelligence approach, we aimed to predict the gene expression of Caspase-8 (CASP8) by the rest of the genes of the array (20,683 genes), using the series of 414 cases of DLBCL from the LLMPP. We used the multilayer perceptron (MLP) procedure, which produced a predictive model for CASP8 (dependent, target variable) based on the values of the predictor variables. Therefore, the dependent variable was the CASP8 and the covariates were the 20,683 genes. In this analysis, the dependent variable was treated as a scale (continuous) because the values represent ordered categories with a meaningful metric, so that distance comparisons between values are appropriate. Of note, this differs from our previous publications in which the dependent (target) variables were dichotomic (high vs. low, or dead vs. alive) [27,28]. Another difference from our previous publications [27,28] is that we are using the values of the collapsed probes. In the setup, CASP8 was the dependent variable, while for the rest of the genes the covariates and the rescaling of the covariates were standardized. As partitions, 70% of the cases corresponded with the training set, while 30% corresponded with the testing set (the holdout was 0%). In the partition dataset the cases were randomly assigned based on the relative number of cases. The architecture had a series of parameters. The hidden layers setup included the number of hidden layers (one or two), the activation function (hyperbolic tangent or sigmoid), and the number of units (automatically computed or custom). The output layer setup included the activation function (identity, softmax, hyperbolic tangent or sigmoid), and the rescaling of scale dependent variables (standardized, normalized, adjusted normalized or none). The type of training could be batch, online or mini-batch; and the optimization algorithm included the scaled conjugate gradient or gradient descent. In the training options the initial lambda value was 0.0000005, the initial sigma 0.00005, the interval center 0, and the interval offset ±0.5. The output displayed the network structure (description, diagram, and the synaptic weights), and the network performance (model summary, predicted by observed chart and residual by predicted chart). In addition, the output also showed the case processing summary and the independent variable importance analysis. The predicted value or category for the dependent variable was saved as a new variable. The synaptic weight estimates were also exported as an xml file. The setup also included the user-missing values and the stopping rules.The immunohistochemical protein expression of Caspase-8 in the series of 97 cases of DLBCL from Tokai University Hospital showed a histological localization in the cytoplasm of the cells (compatible with B-lymphocytes), that had a morphology of middle or large sized centroblasts, or immunoblasts in some cases. In some cases, with high Caspase-8 expression the localization was perinuclear including some extension into the nucleus. After digital image quantification, the Caspase-8 expression ranged from 0.0% to 40.2%, with a median of 3.1% and a mean of 6.7% ± 8.3. In Figure 3, the immunohistochemical expression of Caspase-8 is shown, with a characteristic low and high expression pictures. In addition, the immunohistochemistry of the other markers is also shown in the Figure 3, Figure 4 and Figure 5.The protein expression of Caspase-8 as a quantitative variable correlated with the overall survival of the patients (Figure 6). The Cox regression analysis showed a trend of correlation with the overall survival, with high values associated with better survival, Beta = −0.045, p value = 0.071, Hazard risk = 0.956 (95% CI 0.911–1.004). A cut-off was searched and at 8.7% two groups of patients were identified, with different overall survival. The group of high Caspase-8 expression (>8.8%, n = 27/97, 27.8%) was characterized by a more favorable overall survival than the group of low expression (<8.8%, n = 70/97, 72.2%): Beta = −1.3, p value = 0.009, Hazard risk = 0.3 (95% CI, 0.1–0.7). This means, on average, a 70% lower risk of death, and a 233% increase in survival time. When the overall survival was compared using the Kaplan–Meier and the Log rank test, the group of high Caspase-8 expression was characterized by a favorable prognosis, with a 3, 5 and 10-year overall survival of 85%, 85% and 75%. Conversely, the group of low expression had an unfavorable prognosis, with a 3, 5 and 10-year survival of 56%, 52% and 40% (p value = 0.005).The protein expression of Caspase-8 was also correlated with the progression-free survival of the patients. As a quantitative variable, Caspase-8 did not correlate with the progression-free survival (p value = 0.251). Using the same cut-off as the overall survival (8.8%), high Caspase-8 expression was associated with a more favorable progression-free survival of the patients (Beta = −0.952, p value = 0.036, Hazard risk = 0.386 (95% CI, 0.2–0.9).Using the same cut-off of the survival analysis (8.8%), a correlation was performed with several clinicopathological characteristics of the series. Nevertheless, no significant correlations were found (Table 3). Therefore, other factors that are not the conventionally tested in DLBCL may be related to the Caspase-8 expression.The immunohistochemical protein expression of Caspase-8-related markers was also analyzed in the series of 97 cases of DLBCL from Tokai University Hospital. In the Table 4 the distribution of the markers, including cleaved Caspase-3, cleaved PARP, MDM2, BCL2, TP53, MYC, Ki67, E2F1, CDK6, MYB, LMO2 and TNFAIP8 is shown in detail. The expression of Caspase-8 correlated with these markers, and positive correlation was found for cleaved Caspase-3 (correlation coefficient 0.435), MDM2 (0.389), E2F1 (0.324), TNFAIP8 (0.248), BCL2 (0.217), and Ki67 (0.204) (Table 5). Of note, cleaved Caspase-3 positively correlated with cleaved PARP (correlation coefficient = 0.679, p = 0.003).The correlation with overall survival and the progression-free survival of these markers is also shown in the Table 6 and Table 7.Predictive analytics was performed to model the immunohistochemical expression of Caspase-8 as a dichotomic variable (high vs. low, using the same 8.7% cut-off) with all the other Caspase-8-related markers, which were used as quantitative variables.Twelve different models were executed, including the algorithms of C5.0 node that builds a decision tree or a rule set, logistic regression, Bayesian Network, discriminant analysis, k-Nearest Neighbor (KNN), Support Vector Machine (SVM), Tree-AS decision tree, Chi-squared Automatic Interaction Detection (CHAID) decision tree, Classification and Regression (C&R) Tree and Neural Network.Results of the analysis showed that 9 models predicted the Caspase-8 expression. When ranked according to overall accuracy, they were as follows: CHAID (92%, 4 variables), Bayesian Network (88%, 12 variables), SVM (87%, 12 variables), Discriminant (86%, 12 variables), C5 (85%, 2 variables), Logistic regression (83%, 12 variables), Neural network (80%, 12 variables), C&R Tree (72%, 12 variables), and KNN Algorithm (69%, 12 variables).The CHAID node graph is shown in Figure 7; this decision tree predicted the Caspase-8 expression using cCASP3, BCL2, LMO2 and cPARP. The CHAID classification method builds decision trees by using chi-square statistics to identify optimal cut-offs (splits). Unlike the C&R Tree and the QUEST nodes, the CHAID method can generate non-binary trees. Therefore, the splits can be of more than 2, and the trees are wider.The bayesian network model is shown in Figure 8. A Bayesian network is a graphical model that shows variables (i.e., nodes) in a dataset and the probabilistic, or conditional, independencies between them. Causal relationships between nodes may be represented, but the links in the network (i.e., arcs) do not necessarily represent direct cause and effect. The basic view contains a network graph of nodes that displays the relationship between the target (dependent) variable and the predictor variables, and the relationship between the predictors. The distribution view shows the conditional probabilities for each node in the network as a mini graph, the corresponding tables for cleaved Caspase-3 and E2F1 are shown below.The discriminant analysis had 6 excluded cases due to having at least one missing discriminant variable, so the valid cases were 91 of 97 (93.8%). The number of discriminant functions was 1, with an eigen value (discriminating ability) of 0.612 (p = 0.83 × 10−4). The standardized canonical discriminant function coefficients for the different markers were Ki67 (−0.178), LMO2 (0.125), MYC (0.148), MDM2 (−0.329), CDK6 (0.035), E2F1 (0.569), BCL2 (0.190), MYB (0.109), TP53 (−0.328), cPARP (−0.345), cCASP3 (1.118) and TNFAIP8 (0.180).The C5.0 algorithm builds a decision tree by splitting the sample based on the field that provides the maximum information gain. The C5.0 node can predict only a categorical target. In this model, the Caspase-8 expression (high vs. low) was predicted by cleaved Caspase-3 and E2F1 variables as shown in the Figure 9.The logistic regression (i.e., nominal regression) classifies records based on values of input fields. It is comparable to the linear regression, but the target variable is categorical instead of a numeric one. The logistic regression equation for the High Caspase-8 expression was the following: −0.02362*Ki67 + 0.01278*LMO2 + 0.1576*MYC + −0.2012*MDM2 + 0.009816*CDK6 + 0.9908*E2F1 + 0.9499*BCL2 + 0.05347*MYB + −0.2845*TP53 + −1.631*cPARP + 3.21*cCASP3 + −0.004535*TNFAIP8 + −2.65. The predictor importance, from most to less was the following: cCASP3, E2F1, BCL2, TP53, MYC, MYOB, CDK6, LMO2, TNFAIP8, Ki69, cPARP and MDM2. As shown in Table 8, in this model the significant variables were cCASP3, cPARP, MDM2 and E2F1.When the logistic regression was repeated using the backward method, the predictor importance rank was cCASP3 (most important), E2F1, TP53, BCL2, MYC, cPARP and MDM2 (less). In this case, the equation for High Caspase-8 expression was 0.1527*MYC + −0.1942*MDM2 + 0.8855*E2F1 + 0.09246*BCL2 + −0.3098*TP53 + −1.666*cPARP + 3.114*cCASP3 + −2.697.The Neural Network analysis predicted the Caspase-8 expression as a dichotomic variable (high vs. low) with an overall accuracy of 80.4%, using the quantitative values of the Caspase-8-related markers as predictors. This analysis was repeated with two consecutive but independent multilayer perceptron (MLP) and radial basis function (RBF) artificial neural network (ANN) analyses. The details of the neural networks and the results are shown in Table 9 and Figure 10. In summary, the MLP was characterized by a better “performance” because of a lower percent of incorrect predictions both in the training (9.4% vs. 15.7%) and the testing (7.4% vs. 23.8), better overall % of correct classification of training (90.6% vs. 84.3%) and testing (92.6% vs. 76.2%), and a slightly better area under the curve (0.891 vs. 0.880). According to the MLP analysis, the most relevant markers for predicting the Caspase-8 expression as a dichotomic variable were cleaved Caspase-3 (100%), E2F1 (93%), CDK6 (58.8%), TP53 (46.8%), MYC (42.5%), MYB (30.2%), Ki69 (30%), cleaved PARP (12.7%), BCL2 (9%), TNFAIP8 (8.2%) and LMO2 (1.5%).The results of several tests were integrated to calculate the percentage of importance for the association to Caspase-8. The most relevant markers were cCASP3, E2F1, TP53, MDM2, BCL2 and TNFAIP8 (Table 10).The LLMPP DLBCL dataset that is comprised of 20,684 genes was used to identify in an unsupervised manner which genes are associated with the CASP8 expression. A multilayer perceptron analysis was performed, with CASP8 as dependent variable (quantitative data) and the rest of 20,863 as predictors (also as quantitative variables). As a result of the artificial neutral network, the genes were ranked according to their normalized importance for prediction of the CASP8 expression. The neural network moderately managed to predict the CASP8 expression. According to their normalized importance, the top most relevant genes were: MED29 (1st), PRH1, YIPF3, PLEKHH1, PRB4, IKZF1, CYSRT1, ACTC1, FAM160B1, TBC1D10C, TMEM176B, ADAMTS10, CTSV, CEP20, AZGP1, ZNF557, SDCCAG8, CSKMT, BGLAP and SRP54 (20th).To understand the relationship between CASP8 expression and the top 20 genes, the expression of CASP8 was modeled using the top 20 genes. The analysis included the following model types: regression, generalized liner, linear-AS, LSVM, random trees, Tree-AS, linear, CHAID, C&R tree and neural network. The most relevant models were the following: CHAID (correlation 0.806), neural network (0.712), regression (0.668), generalized linear (0.668), linear (0.667) and C&R tree (0.647).A visualization of the CHAID and neural network is shown in Figure 11. The regression output was the following: CASP8 = MED29*0.1483 + PRH1* − 0.1032 + YIPF3* − 0.1555 + PLEKHH1*0.1117 + PRB4* − 0.001069 + IKZF1*0.1014 + CYSRT1*0.008583 + ACTC1*0.04482 + FAM160B1*0.2315 + TBC1D10C*0.2088 + TMEM176B*0.1449 + ADAMTS10*0.1131 + CTSV* − 0.0005433 + CEP20*0.1234 + AZGP1*0.06398 + ZNF5571*0.08978 + SDCCAG8* − 0.04932 + CSKMT*0.05439 + BGLAP*0.08571 + SRP54*0.3457 − 6.131.Further analysis was performed focusing on CASP8 as a dichotomic variable in the DLBCL GEO GSE10846. Using a ROC curve analysis, the best cut-off of CASP8 for the overall survival phenotype (dead/alive) was searched, and the value was 10.3805. Among the 414 cases of the series, CASP8 was high in 180 (48.3%) and low in 234 (69.2%). We confirmed the association of most of the previously identified 20 top genes of the neural network analysis with a high CASP8 expression. The Gene Set Enrichment Analysis (GSEA) is a biostatistical method that confirms if a defined set of genes correlates between two biological states (e.g., phenotypes). We used GSEA to correlate the phenotype CASP8 high vs. low with several set of genes (pathways). The whole collection of the MSigDB gene sets were used (23,677 genes sets in total, MSigDB database v7.3 updated March 2021), which include 9 major collections: H (hallmark genes), C1 (positional), C2 (curated), C3 (regulatory target), C4 (computational), C5 (ontology), C6 (oncogenic signature), C7 (immunologic signature), and C8 (cell type signature). From the 23,677 tested genes sets, 843 gene sets were significantly enriched at nominal p value < 5%, either towards high or low CASP8. For example, significantly enriched pathways of the oncogenic signature that associated to high CASP8 were ALK, KRAS, PGF, P53 and CYCLIND1. Other correlations included sets of the immunologic signature such as macrophages (Genes up-regulated in bone marrow-derived macrophages treated with IL4, GSE25088). The complete results are available on request from the corresponding author (Carreras J).This research focused on the analysis of Caspase-8 in DLBCL from Tokai University Hospital. The protein expression of Caspase-8 was evaluated by immunohistochemistry, followed by marker quantification by digital image analysis. We found that high Caspase-8 protein expression was associated with a favorable prognosis of the patients, including a favorable overall and progression-free survival.Apoptosis is a term to designate programmed cell death. The mechanism of cell death has multiple roles, including a function in the pathogenesis, homeostasis, and control of several types of infection, as well as in cancer [41]. Excessive cell damage results in passive necrosis. On the other hand, the mechanism of cell death can be triggered by several molecular programs including cellular stress, oncogenic changes that involve tumor suppressor genes and oncogenes, several pathogens, and other immune mechanisms. Apoptosis is one of the most known and studied types of programmed cell death [41]; other types of programmed cell death are necroptosis, pyroptosis, ferroptosis, mitotic catastrophy and autophagic cell death, among others [41]. The pathway of apoptosis includes an extrinsic (controlled death receptors of the TNFR superfamily) and an intrinsic (mitochondrial) pathway. Interestingly, ligation of these death receptors induces both activation of extrinsic apoptosis and necroptosis, and the balance between these two pathways determines whether the cell lives. Caspase-8 has a role in initiating of extrinsic apoptosis and inhibiting necroptosis [41]. Caspase-8 activates Caspase-3 by proteolytic cleavage, and then Caspase-3 cleaves other vital cellular proteins or other caspases, which result in activation of cPARP, which eventually leads to apoptosis [42,43,44].In DLBCL, the mechanisms of cell survival are dysregulated [45]. Dysregulation of an inhibitor of apoptosis proteins (IAPs) has been described in DLBCL [45]. For example, overexpression of XIAP (an apoptosis inhibitor) was associated with a worse outcome in DLBCL [46]. Another inhibitor, the Survivin, was also found overexpressed in DLBCL [47] and in ABC molecular type DLBCL the overexpression was also associated to a poor prognosis [47]. Besides, we recently described that high expression of another apoptosis inhibitor (TNFAIP8) was associated with a poor prognosis of DLBCL [40]. In this project the protein expression of Caspase-8 was analyzed in a series of Tokai University’s, and we found that high expression was associated with a favorable survival of the patients. Therefore, while anti-apoptosis seems to be associated to a poor prognosis of DLBCL, the pro-apoptosis Caspase-8 associates to a favorable outcome of the patients.In DLBCL there is also dysregulation of TP53 [45], which includes not only mutations or deletions of TP53, but also alterations of TP53 pathways-related markers of BCL6, MDM2, CDKN2A, etc. In this research some of these markers were analyzed by immunohistochemistry in the Tokai series, and the relationship between them as well as with Caspase-8 was explored as shown in Figure 1. In addition, using several modeling analyses, we showed how these markers correlated with the Caspase-8 expression, either as positive or negative correlation, so a pathogenic model can be postulated. For example, the Caspase-8 expression could be calculated as 0.2*MYC + −0.2*MDM2 + 0.9*E2F1 + 0.1*BCL2 + −0.3*TP53 + −1.7*cPARP + 3.1*cCASP3 − 2.697.This research focused on the analysis of Caspase-8 in a series of Tokai University’s and we found that high protein expression of Caspase-8 correlated with a favorable outcome of the patients, both the overall survival and the progression-free survival. As shown in the Figure 6, the 30% of the patients with high Caspase-8 expression had a favorable overall survival. At the 10-years’ time, around 80% of the patients with high Caspase-8 expression were still alive. Conversely, at that time only 40% were alive in the low expression group. This finding was important and to the best of our knowledge, to date, this association has not been reported in DLBCL. Nevertheless, the Caspase-8 did not correlate with the conventional clinicopathological variables that are usually associated with the prognosis of DLBCL such as the cell-of-origin molecular classifications (Hans’ algorithm) and the International Prognostic Index (IPI) that integrates the clinical variables of age, performance status, LDH, extranodal sites and stage. Therefore, a functional network association analysis was performed, markers associated to Caspase-8 were identified (Figure 1), and finally several types of predictive modeling were tested.Predictive analytics was performed to model the immunohistochemical expression of Caspase-8 as a dichotomic variable (high vs. low, using the same 8.7% cut-off for the overall survival analysis) with the other Caspase-8-related markers, which were used as quantitative variables.Twelve different models were executed, including the algorithms of C5.0 node that builds a decision tree or a rule set, logistic regression, Bayesian Network, discriminant analysis, k-Nearest Neighbor (KNN), Support Vector Machine (SVM), Tree-AS decision tree, Chi-squared Automatic Interaction Detection (CHAID) decision tree, Classification and Regression (C&R) Tree and Neural Network. All these models of data mining are tools that enable to develop predictive models using the research experimental data. This data mining process allowed better results and data interpretation, and integrated methods of machine learning, artificial intelligence, and statistics. Of note, each method had certain strengths and was best suited for particular types of problems. Among the 12 different models that were executed, 9 models predicted the Caspase-8 protein expression as a dichotomic variable (high vs. low). When ranked according to their overall accuracy for Caspase-8 prediction, the results were as follows: CHAID tree (92%, 4 variables), Bayesian Network (88%, 12 variables), C5 tree (85%, 2 variables), Logistic regression (83%, 12 variables) and Neural network (80%, 12 variables). The results of all these types of analysis were compatible between them, and each model provided insights into the relationship between Caspase-8 and the rest of the markers. Nevertheless, as previously stated, each method had strengths and weaknesses. For example, the decision trees had an overall accuracy that ranged from 92% for the CHAID tree to 85% of the C5 tree. This means that prediction of Caspase-8 was successful, although variable. Nevertheless, in these models not all the markers were used in the final model, so the relevance of some of the markers cannot be properly assessed. The Bayesian Network built a probabilistic model and made use of all the markers. Bayesian Networks are very robust where information is missing and make the best possible prediction using whatever information is present. Causal relationships between nodes may be represented but the links in the network (i.e., arcs) do not necessarily represent direct cause and effect. The logistic regression (i.e., nominal regression) classifies records based on values of input fields. It is comparable to the linear regression, but the target variable is categorical instead of numeric. This method had the strength of allowing us to know which were the most relevant markers for the prediction of Caspase-8, with information of the direction of the association (increase or decrease) and the strength of that association. Neural networks are simple models of the way the nervous system operates. The basic units are neurons, which are typically organized into layers. There are three parts in a neural network: the input, the hidden and the output layers. The network learns thorough training. Since the output is known, as the training progresses the network becomes increasingly accurate in replicating the known outcomes. Since the deep neural networks have a multilayer non-linear structure (i.e., black box model), neural networks are criticized to be non-transparent because their predictions are not traceable by humans. In our analysis we could rank the markers according to their normalized importance for Caspase-8 prediction, but the reason for this association was elusive because the synaptic weights are only sort of meaningful. In summary, we used a series of algorithms to create classification models. Each model used the values of the input fields (our markers) to predict the value of one output or target field (Caspase-8 as a dichotomic variable, high vs. low), and the integration of all the information made the results more understandable (explainable). As shown in Table 10, the most relevant markers associated with Caspase-8 were the following: cCASP3, E2F1, TP53, cPARP, MDM2, BCL2 and TNFAIP8. Caspase 3, PARP, BCL2 are known markers closely related to apoptosis. Therefore, it makes sense that they were highly associated with Caspase-8. Nevertheless, some of the markers are also associated with other pathways. MDM2 is a ligase that inhibits the p53 and p73-mediated cell cycle arrest and apoptosis [31]. The p53 protein is a tumor suppressor that also controls the cell cycle and induces apoptosis. MYC proto-oncogene is a transcription factor that activates the transcription of growth-related genes and promotes angiogenesis. Ki67 has a role in chromatin organization and it is a widely used marker of cell proliferation. E2F1 is also involved in the cell cycle. CDK6 is a kinase that also controls the G1/S cell cycle transition and the cell differentiation [31]. MYB also controls the cell cycle and cell differentiation. LMO2 is a nuclear marker of normal B lymphocytes of the germinal centers, and DLBCL is supposed to be developed from these lymphocytes. Finally, TNFAIP8 is a negative regulator of apoptosis and plays a role in tumor progression [31]. In summary, the most relevant markers that we have highlighted belonged to the apoptosis and the control of cell cycle.Finally, the Capase-8 gene expression as a quantitative variable was also analyzed in an independent series of DLBCL of the LLMPP, as the relationship with other genes could also be successfully explored. The most relevant gene was MED29, a component of the Mediator complex that is involved in the regulation of transcription [31]. MED29 has been related to prostate cancer [48].Future research directions should include analyzing the same markers in larger series of DLBCL to validate our findings. In addition, in-vitro or in-vivo analyses may also help to clarify the pathological function of Caspase-8 in DLBCL.In conclusion, high immunohistochemical protein expression of Caspase-8 is associated with a favorable overall survival and progression-free survival of the patients in a series of DLBCL from Tokai University Hospital. The relationship of Caspase-8 with other related markers could also be confirmed by predictive analytics including decision trees, Bayesian network, logistic regression and artificial neural networks. Therefore, the immunohistochemical analysis of Caspase-8 could be implemented in the routine diagnosis of DLBCL as a prognostic marker.Conceptualization, J.C.; methodology, J.C.; software, J.C.; validation, R.H.; formal analysis, J.C.; investigation, Y.Y.K., M.M., S.H., S.T., H.I., Y.K., A.I., G.R., S.S., K.A.; resources, N.N., K.A., R.H., G.R.; writing—original draft preparation, J.C.; writing—review and editing, J.C.; supervision, N.N.; project administration, J.C.; funding acquisition, J.C. and R.H. All authors have read and agreed to the published version of the manuscript.This research was funded by THE MINISTRY OF EDUCATION, CULTURE, SPORTS, SCIENCE AND TECHNOLOGY (MEXT) and THE JAPAN SOCIETY FOR THE PROMOTION OF SCIENCE, grant number KAKEN 18K15100 to Joaquim Carreras. Rifat Hamoudi was funded by AL-JALILA FOUNDATION (grant number AJF201741), THE SHARJAH RESEARCH ACADEMY (grant number MED001) and UNIVERSITY OF SHARJAH (grant number 1901090258).The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Institutional Review Board and the Ethics Committee of Tokai University, School of Medicine (protocol code IRB14R-080 and IRB-156).Informed consent was obtained from all subjects involved in the study.The gene expression data of DLBCL (GEO dataset GSE10846) was obtained from the publicly available database of the NCBI resources webpage, located at https://www.ncbi.nlm.nih.gov/gds (accessed on 16 April 2021). The data from Tokai University presented in this study are available on request from the corresponding author. The data are not publicly available due to data protection policy.We would like to thank and acknowledge all the members of the Lymphoma/Leukemia Molecular Profiling Project (LLMPP) who participated in the generation of the GSE10846 dataset, including LM Staudt, E Campo, ES Jaffe, WC Chan, WH Wilson, TA Lister, RD Gascoyne, JM Conners, G Wright, SS Dave, LM Rimsza, A Ronsenwald, D Wrench, H-K Muller-Hermelink, G Ott and E Hartman (among others).The authors declare no conflict of interest.Interactions between the Caspase-8 and the Caspase-8-related proteins. The aim of this research is to analyze the role of Caspase-8 in Diffuse large B-cell lymphoma, focusing in the investigation of the possible pathological mechanism, the correlations with Caspase-8-related markers and the clinicopathological correlations. This network summarizes the predicted associations of Caspase-8 with the group of pathway-related proteins. The nodes are the proteins and the edges represent the predicted functional associations: action types (activation, binding, inhibition, etc.) and effects types (positive, negative, and unspecified). The basic network only has the markers (nodes) of this project (left), the extended network (right) includes additional nodes for better action types and action effects information.(a) General architecture for the multilayer perceptron artificial neural network. (b) Activation functions for the multilayer perceptron artificial neural network. (c) Error functions for the multilayer perceptron artificial neural network. (d) Notation for the multilayer perceptron artificial neural network.Immunohistochemical expression in the DLBCL samples of Caspase-8, cleaved Caspase-3, cleaved PARP, MDM2 and BCL2 (Tokai series). Caspase-8 protein is a protease with a key role in the programmed cell death (extrinsic apoptosis). Once activated, Caspase-8 cleaves and activates other effector caspases including Caspase-3 and PARP1. It also regulates necroptosis and innate immunity. MDM2 is a ligase that inhibits the p53 and p73-mediated cell cycle arrest and apoptosis. BCL2 is an apoptosis inhibitor, controlling the mitochondrial membrane activity and inhibiting caspase activity [31]. By immunohistochemistry, Caspase-8 protein expression was cytoplasmic and perinuclear, with some staining in the nucleus when the protein expression was high. Cleaved Caspase-3, cleaved PARP and MDM2 staining was nuclear. BCL2 expression was mainly cytoplasmic and perinuclear.Immunohistochemical expression in the DLBCL samples of TP53, MYC, Ki67, E2F1 and CDK6 (Tokai series). P53 is a tumor suppressor that controls the cell cycle and induces apoptosis. MYC proto-oncogene is a transcription factor that binds the DNA and activates the transcription of growth-related genes, promotes angiogenesis and regulates somatic reprogramming. Ki67 plays a key role in cell proliferation, with a role in chromatin organization maintaining the mitotic chromosomes dispersed. E2F1 is a transcription factor involved in cell cycle regulation (progression from G1 to S phase) and DNA replication. E2F1 binds RB1 and can mediate both cell proliferation and p53 apoptosis. CDK6 is a kinase involved in the control of cell cycle (G1/S transition) and cell differentiation [31]. By immunohistochemistry all the markers show nuclear staining. CDK6 also shown cytoplasmic localization.Immunohistochemical expression of MYB, LMO2 and TNFAIP8 (Tokai series). MYB is a transcriptional activator that binds the DNA and plays a role in the control of cell proliferation and differentiation. LMO2 is a nuclear marker expressed by normal B lymphocytes in the germinal centers. It also regulates hematopoietic stem cell differentiation. TNFAIP8 is a negative regulator of apoptosis and play a role in tumor progression. It inhibits Caspase-8, subsequently resulting in inhibiting the activation of Caspase-3 [31]. We have recently described that high expression of TNFAIP8 correlates with poor survival of DLBCL patients [28]. MYB and LMO2 protein expression is nuclear, TNFAIP8 is in the cytoplasm and perinuclear.Overall and progression-free survival according to the Caspase-8 expression by immunohistochemistry (Tokai series, immunohistochemical data). High percentages of Caspase-8 associated with a favorable prognosis of the patients with DLBCL, including both the overall survival and the progression-free survival.CHAID node decision tree analysis (Tokai series, immunohistochemical data). The Chi-squared automatic interaction detection (CHAID) is a classification method for building decision trees that identify optimal splits by using chi-square statistics. CHAID examines the crosstabulations between each input field and the outcome, and tests for significance. CHAID can generate nonbinary trees (splits of more than two branches). In this analysis we aimed to predict the Caspase-8 expression as low (1) versus high (2), which is the same cut-off used for the survival analysis. The Caspase-8 expression could be predicted using cleaved Caspase-3, BCL2, cleaved and PARP. This decision tree is highlighting the Caspase-8, cCaspase-3, cPARP apoptosis pathway.Bayesian Network (Tokai series, immunohistochemical data). The Bayesian network allows to build a probabilistic model combining observed and recorded evidence with “common-sense” real-world knowledge to establish the likelihood of occurrences by using seemingly unlinked attributes. Therefore, Bayesian networks are used for making predictions. Each of the nodes is one of the markers that have been analyzed by immunohistochemistry in the Tokai series of DLBCL. In this analysis we aimed to predict the Caspase-8 expression (target) by the rest of the markers (predictors). Bayesian networks are very robust where information is missing and make the best possible prediction using whatever information is present. In this figure, the conditional probabilities of cCaspase-3 and E2F1 are also shown.C5.0 node decision tree analysis (Tokai series, immunohistochemical data). The C5.0 algorithm was used to predict the Caspase-8 expression as a categorical target (low versus high, same cut-off for the survival analysis) by the rest of the markers (predictors). C5.0 models are quite robust when missing data is present and there are large numbers of input fields. C5.0 models tend to be easier to understand. In this analysis we found that Caspase-8 expression could be predicted by cCaspase-3 and E2F1, highlighting the apoptosis pathway.Artificial Neural Network analysis for the prediction of Caspase-8 by the Caspase-8-related markers (Tokai series, immunohistochemical data). The neural network model determines how the network connects the predictors (our series of 12 markers, input layer) to the targets (the Caspase-8, output layer, as a dichotomic variable high versus low, same cutoff used for the survival analysis) through the hidden layers. The multilayer perceptron (MLP) allows for more complex relationships. Conversely, the radial basis function (RBF) is generally faster and has only one hidden layer, but at the cost of reduced predictive power. The hidden layer(s) contains unobservable units. The value of each hidden unit is some function of the predictors. In this figure, the relevance of each marker for prediction of Caspase-8 is shown by the width of the node and by the value of the normalized importance for prediction. The performance of the network can be checked by the area under the curve ROC curve, of which the higher it is, the better the prediction of Caspase-8 expression. The synaptic weights from the output of the network are available on request from the corresponding author (Carreras J).Prediction of CASP8 by 20,683 genes of the LLMPP series and modeling using the top 20 most relevant genes (gene expression data). The DLBCL gene expression data of the GEO dataset GSE10846 of the Lymphoma/Leukemia Molecular Profiling Project (LLMPP) was used to predict the expression of the CASP8 as a quantitative target variable. In this analysis, the predictors were the 20,863 genes of the gene expression array. Conversely to the analysis of the Tokai cases, in the LLMPP data analyses the CASP8 is predicted as a quantitative variable, which we have not performed in our previous publications (thus the novelty). In neural networks, the predicted by observed chart is used for continuous targets and displays a binned scatterplot of the predicted values on the vertical axis by the observed values on the horizontal axis. The importance of each predictor in making the prediction is shown in the independent variable importance figure. The synaptic weights from the output of the network and the normalized importance chart are available on request from the corresponding author (Carreras J). Typically, the modelling will focus on the predictor fields that matter most and those that matter least will be dropped or ignored. Therefore, the neural network was repeated only with the top 20 genes. In addition to the neural network analysis, this figure also shown the result of the CHAID decision tree.Clinicopathological characteristics of the DLBCL series of Tokai University Hospital.DLBCL, Diffuse Large B-cell Lymphoma; IPI, International Prognostic Index; CR, clinical response; PD, persistent disease; PR, partial response; GCB, germinal center B-cell type.Clinicopathological characteristics of the DLBCL series of the LLMPP.ECOG, Eastern Cooperative Oncology Group; LDH, Lactate dehydrogenase; NCCN IPI, NCCN, National Comprehensive Cancer Network; IPI, International Prognostic Index (IPI); RCHOP, rituximab, cyclophosphamide, hydroxydaunorubicin, oncovin, prednisone/prednisolone; GCB, germinal center B-cell type; ABC, activated B-cell type. Note: The GSE10846 dataset represents previously published data of the LLMPP, which is not the authors’ own work.Correlation between the clinicopathological characteristics of the DLBCL cases and high immunohistochemical expression of Caspase-8 (Tokai series).Immunohistochemical expression of Caspase-8-related markers in DLBCL (Tokai series).Correlation between Caspase-8 and the Caspase-8-related markers (Tokai series).Spearman’s rho non-parametric correlation.Overall survival of the Caspase-8-related markers in DLBCL (Tokai series).*1 Kaplan-Meier with Breslow (Generalized Wilcoxon) test, p = 0.047. *2 Kaplan-Meier with Breslow (Generalized Wilcoxon) test, p = 0.045.Progression-free survival (PFS) of the Caspase-8-related markers in DLBCL (Tokai series).*1 Kaplan-Meier with Breslow (Generalized Wilcoxon) test, p = 0.045. *2 Kaplan-Meier with Log Rank (Mantel-Cox) test, p = 0.046. *3 Kaplan-Meier with Breslow (Generalized Wilcoxon) test, p = 0.035.Logistic regression of Caspase-8 by the Caspase-8-related markers (Tokai series).Artificial Neural Network analysis for Caspase-8 prediction by the Caspase-8-related markers (Tokai series).*1 Error computations are based on the testing sample. *2 Determined by the testing data criterion: The “best” number of hidden units is the one that yields the smallest error in the testing data.Integrated analysis, ranking of markers according to relevance of Caspase-8 association.1, highlighted in the model; 0, not highlighted. MLP, multilayer perceptron; RBF, radial basis function; ANN, artificial neural network.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-02-00004.txt

@@ -0,0 +1,5 @@
+Sample size calculation for adequate power analysis is critical in optimizing RNA-seq experimental design. However, the complexity increases for directly estimating sample size when taking into consideration confounding covariates. Although a number of approaches for sample size calculation have been proposed for RNA-seq data, most ignore any potential heterogeneity. In this study, we implemented a simulation-based and confounder-adjusted method to provide sample size recommendations for RNA-seq differential expression analysis. The data was generated using Monte Carlo simulation, given an underlined distribution of confounding covariates and parameters for a negative binomial distribution. The relationship between the sample size with the power and parameters, such as dispersion, fold change and mean read counts, can be visualized. We demonstrate that the adjusted sample size for a desired power and type one error rate of α is usually larger when taking confounding covariates into account. More importantly, our simulation study reveals that sample size may be underestimated by existing methods if a confounding covariate exists in RNA-seq data. Consequently, this underestimate could affect the detection power for the differential expression analysis. Therefore, we introduce confounding covariates for sample size estimation for heterogeneous RNA-seq data.Sample size and power are important factors for planning a biological experiment using high-throughput sequencing technologies for differential gene expression (RNA-seq). Larger sample sizes typically provide a more accurate estimate of the differential gene expression with high confidence. However, since RNA-seq techniques are costly, a large sample size is sometimes not feasible when limited research budgets are considered. Therefore, an optimized sample size is desired to achieve a specific power for detecting gene expression changes within realistic budget constraints. Moreover, since read depths often vary significantly between runs, this particular technical variation also needs to be taken into consideration in any sample size estimation.With the rapid growth of RNA-seq studies, a number of sample size estimation methods and software tools have been proposed [1,2,3,4,5,6,7,8,9]. However, these methods have their limitations and assumptions. Since RNA-seq data are short read counts, Fang and Cui (2011) used a Poisson distribution to derive a sample size calculation formula combined with a Wald-like Z-statistic test on a single gene [1]. Li et al. (2013) extended sample size calculation methods using a Wald test, a score test and a likelihood ratio test (LRT) based on testing a single gene or multiple genes [2]. However, the studies [10,11] found that a Poisson distribution may not be appropriate to model gene read counts in RNA-seq data due to over-dispersion as a result of natural biological variation. To address this issue, a negative binomial (NB) distribution combined with an exact test and/or likelihood ratio test (LRT) was proposed to model RNA-seq data in differential gene expression analysis [10,11,12]. Subsequently, other sample size calculation methods were proposed using an NB distribution [3,4,5,7,8]. Hart et al. (2013) proposed a sample size calculation method using a score test based on a single gene [7] and Liu et al. (2014) further proposed sample size calculations using an exact test implemented in edgeR [8]. Later, Li et al. (2013) developed the RnaSeqSampleSize R package based on TCGA data [13]. Similarly, Ching et al. [6] and Wu et al. [14] performed a power analysis implemented in DESeq2 and/or edgeR while controlling false discovery rate (FDR). These methods employed the common analysis approaches with the aid of DESeq2 or edgeR. However, these studies have reported the actual FDR resulting from NB-based methods such as DESeq2 and edgeR was inflated in many cases [15,16,17,18,19,20]. To address this issue, Yu et al. [9] proposed a power analysis based mainly on simulation studies for a given desired type I error rate. In addition, several sample size calculations were developed using a Wald test, a log-transformed Wald test and an analytical method using a log LRT test based on a single gene or multiple genes with controlling FDR [4]. Recently, we proposed a method for sample size calculation using a generalized linear model (GLM) with an NB distribution where the dispersion was estimated on the basis of a variance–covariance matrix between two groups [5]. However, the sample sizes estimated in all these studies may only be appropriate for homogeneous data with tightly controlled conditions. With a GLM, it is very important to identify independent covariates and confounding covariates in an experimental design. The difference in covariates is that the independent covariates can be controlled by experimental design, while the confounding covariates cannot be controlled. These confounding covariates from heterogeneous data commonly exist in clinical RNA-seq studies such as cancer and other disease-associated datasets. For example, age and sex are common confounding factors in RNA-seq, as are more complex variables such as diet, exercise, and environmental influences. Existing methods for determining sample size are suitable for cell lines or animal studies where other variables can be tightly controlled. However, when a confounding covariate exists in an experiment, such as with nearly all human studies, these methods may underestimate the sample size, eventually affecting the statistical power of the experiment.To address this issue, we introduce confounder-adjusted sample size calculation using a simulation-based empirical approach. These simulated data are based on a NB regression model with the aid of the rnbinom and glm.nb functions of the MASS R package. The confounding covariate of the simulated study is defined as a continuous variable (i.e., age) or a categorical variable (i.e., sex). We illustrate how to calculate age and sex-adjusted sample size and power using the public colon adenocarcinoma (COAD) data downloaded from Broad GDAC Firehouse. The method described here can provide an additional option for clinical researchers to determine sample size in designing complex RNA-seq experiments.A Generalized liner model (GLM) has been widely applied in scientific fields [21]. For a single gene in RNA-seq data, the independent random sample (Yij) for the sample j j=1, …,ni in condition i (i=0, 1) is assumed to have an identical NB distribution, such as Yij~NBµij,ϕ. Thus, the probability mass function of the observation yij is defined as: (1)P(yij)=Γϕ−1+yijΓϕ−1yij!ϕµij1+ϕµijyij11+ϕµijϕ−1,
+        where µij=sijγi, sij is the size factor for normalizing read depth, γi is the true expression of the gene and is unknown, and µij is the expected mean expression.For the purpose of power and sample size calculation within the framework of a GLM, we defined the expected mean read counts (µij) for yij by a log link function as:(2)log uij=log sij+ψ0+ψ1Zi +bLi+λXi,
+        where the covariate Zi, a treatment group indicator, takes value Z0 =0 if i=0 for the control group and Z1 =1 if i=1 for the treatment group. The multiple covariates Li and Xi, confounding variables, are assumed to be a continuous variable and/or categorical variable, respectively, and the quantity log sij denotes an offset. The true expression of γi is analyzed directly by the GLM and can also be expressed as:(3)log γi=ψ0+ψ1Zi +bLi+λXi ,Thus, the true expression γ0 and γ1 from Equation (3) can be obtained as:(4)γ0=eψ0+bL0+λX0 , γ1=eψ0+ψ1+bL1+λX1  and γ1γ0=eψ1+bL1−L0+λX1−X0Replacing uij=sijeψ0+ψ1+bLi+λXi =sijγi in Equation (1), the log-likelihood function is expressed as:(5)l=∑i=01∑j=1nilog Γϕ−1+yijΓϕ−1yij!+yijlog ϕsijeψ0+ψ1Zi+bLi +λXi −yij+1ϕlog1+ϕsijeψ0+ψ1Zi+bLi +λXi The covariant-adjusted coefficient ψ1 is expected to be different for the log count of the gene between the treatment and the control groups. In this study, the p-value along with the coefficient ψ1 obtained from the glm.nb function is used to determine if the gene read counts in the treatment group is statistically significant from the control. The RNA-seq data relies on parameters to be simulated. The sample size and actual power are determined by the DEG analysis for the different parameter settings. In this study, we considered the presence of both single and dual confounding variables.Simulation of single confounding factor data: For a single confounding variable, we have two sets of linear predictors in the form log γ0=ψ0+λX0 and log γ1=ψ0+ψ1+λX1  for the control and treatment group, respectively. For dual confounding variables, they are log γ0=ψ0+bL0+λX0 and log γ1=ψ0+ψ1+bL1+λX1 . Three scenarios in single confounding variables are described as follows. In the first scenario, we consider the confounding covariate X0, given Z0=0, and X1, given Z1=1, follows a normal distribution with equal and/or unequal means and variance resulting in four settings: N0(0, 1) and N1(0, 1.52), N0(0, 1) and N1(1.5,1), N0(40, 52) and N1(42, 52), N0(40, 22) and N1(42, 52) and N0(30, 32) and N1(40,32) for the control and treatment group, respectively. In the second scenario, we consider the covariate X0 and X1 follows a Poisson distribution with equal and/or unequal means resulting in three settings: Pois0(10) and Pois1(12), Pois0(10) and Pois1(15) and Pois0(25) and Pois1(20). Two settings are a mixture of normal and Poisson distributions: Pois0(10) and N1(12, 1) and Pois0(10) and N1(12, 10) for the control and treatment groups, respectively. This is assumed in a rare situation. In the last scenario, we consider the confounding covariate X0 and X1 as categorical variables, each taking the binary value 0 or 1. There are six different settings, including I(0.25, 0.25, 0.25, 0.25), II(0.2, 0.3, 0.3, 0.2), III(0.3, 0.2, 0.2, 0.3), IV(0.1, 0.4, 0.4, 0.1) and V(0.4, 0.1, 0.1, 0.4) and VI(0.1, 0.3, 0.4, 0.2). Each of the six settings corresponds to the different proportion of the single covariate in two groups (0,1) such as sex (male, female). Three of them were originally proposed by Self Steven for a GLM with a Bernoulli distribution [22]. I(0.25, 0.25, 0.25, 0.25) is assumed from a homogeneous confounding covariate, and VI(0.1, 0.3, 0.4, 0.2) is assumed completely unequal proportion in control and treatment groups Simulation of dual confounding variables: For the dual confounding variables, one confounding covariate (L0 and L1) is set to follow a normal distribution with equal and/or unequal mean and variance resulting in two settings: N0(0, 1) and N1(1.5, 1), and N0(40, 22) and N1(42, 52). The second confounding covariate (X0 and X1 ) is set as a categorical variable with four different settings: I(0.2, 0.3, 0.3, 0.2), II (0.1, 0.4, 0.4, 0.1), III (0.1, 0.3, 0.4, 0.2) and V(0.15, 0.35, 0.35, 0.15).Parameter estimation: In the simulation, the alternative hypothesis test on a single gene is that the gene is considered differentially expressed when ψ1≠0. In this study, the fold change (ρ) is set to be 0.5,  1.5,  2.0 or 3.0, corresponding to ψ1≠0. The minimum mean read count of the DEGs in the control group, µ0, is set to be 5 and 10. The ratio of mean size factors, w=s¯1s¯0, is set to be 1 for normalized RNA-seq data and w≠1 for unnormalized RNA-seq data; the constant dispersion parameter ϕ is set to be 0.1,  0.2 or 0.5; the ratio of sample sizes is set to be k=1 for a balanced design.Simulation of RNA-seq data: For a fixed sample size of n given the designed parameter setting, two groups of NB random samples are generated. For the control group, the random samples are generated given the parameters n0, µ0 and ϕ. For the treatment group, the random samples are generated given the parameters n1=kn0, µ1=ρw, µ0 and ϕ. k ≠1 indicates an imbalanced design. Given a fixed n and a specified covariate distribution for the control and treatment groups, the two datasets are randomly and independently generated. Sample size and power estimation: For a given model and covariate distribution, sample sizes are estimated by testing the hypothesis: H0: ψ1=0 vs. H1: ψ1≠0 with significance level α and power 0.80. Each Monte Carlo estimate of power associated with a fixed sample size is imputed under different scenarios and settings through 1000 independently generated datasets.The procedure for sample size and power estimation can be briefly summarized as the following steps:Obtain the pre-specified parameters, such as fold change (ρ), the ratio of size factors w and the ratio of sample sizes k between two-sample groups.Specify a desired statistical power (i.e., 0.80) and significance level α 0.05.Simulate control and treatment groups RNA-seq data given the mean counts in the control group (u0) and common dispersion (ϕ) for a fixed n using an NB distribution with the aid of the rnbinom function in R.Simulate a confounding covariate under different scenarios given a fixed n and distribution with the aid of the rnorm function for a normal distribution, the rpois function for a Poisson distribution and rnbinom for a binomial distribution for a categorical confounder.Fit the GLM with a NB distribution using the R glm.nb function.Obtain the coefficient ψ1 along with the standard error, z-score and p-value for statistical test on ψ1 from the simulated data set. For a two-sided test, record whether a p-value ≤ α/2 in testing a single gene or p-value ≤ α*/2 in testing multiple genes.Repeat steps 3–6 for 1000 times and impute the statistical power for the fixed sample size.Repeat steps 3–7 by increment of sample size by one (n=n+1) if the power is smaller than 0.7999 ≈0.80. Stop when a desired statistical power is obtained and then record the sample size n and the actual power. The R source codes are provided for the illustration of estimating the empirical power and sample size n (Supplementary File S2). In this study, the size α for a single gene has been adjusted for testingmultiple genes, which has been implemented in recent studies [3,5]. A similar approach to the previous studies was used to calculate the new size α by incorporating FDR [2,3,4]. Briefly, given a nominal FDR at a specified level f of 0.05, the adjusted significance level α* for the expected number of true rejection t1 is defined
+          (6)α*=t1ft01−f ,
+          where t0 is the number of true null hypotheses. Replacing the size α 0.05 in testing a single gene with a smaller α* in simulation study from steps 1 to 8, the expected sample sizes and estimates of power corrected by FDR at level f are then obtained. The sample size n (biological replicates) and actual power are calculated given a significance level alpha of 0.05 and a desired 80% power for a single gene or multiple genes at a controlling FDR of 0.05. Monte Carlo estimates are based on empirical data generated for different parameter settings. We performed a GLM with an NB distribution incorporating potential confounding covariates denoted either as a categorical or continuous variable at a normal and Poisson distribution. The Wald-like z test in glm.nb with a log link function is used for testing the significance of the coefficient ψ1 with the inverse of dispersion 1/ϕ. ψ1 is the coefficient of the treatment group as an independent variable in a GLM. ϕ is the dispersion parameter in an NB distribution. The variance of NB distribution is a function of its mean and additional overdispersion of ϕ. The genes between the two groups are considered significantly different when the p-value is ≤ α/2 for a two-sided test. The procedures are repeated 1000 times, and the power is calculated as the percentage of the number of times that the null hypothesis H0 is rejected. Table 1 summarizes the results under different scenarios with a variety of parameter settings, as illustrated in Figure 1, Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8. The bar graphs in Figure 1, Figure 2 and Figure 3 illustrate the sample size n versus the fold change ρ with fixed ϕ and µ0 adjusted by different covariates for testing a single gene. The actual power calculated from the simulation is ≥ 0.8. The color-coded bars in Figure 1 represent confounding covariates in a normal distribution with equal/unequal mean and variance between two groups. The height of the bars illustrated in the figures represents the number of the sample size. We first observe that the n decreases as the fold change ρ increases from 1.5 to 3, where the ρ of 0.5 indicates a 2-fold down-regulated gene. The figures show that a much larger n is required for the gene that has a fold change of 1.5 compared to a fold change of 2 or 3, which is expected. Given a fixed µ0 and ρ, n increases as ϕ increases from 0.1 to 0.5 (Figure 1a–f). This is also expected, which indicates that a larger n is required for a higher variation of samples. We also observed that n decreases as the read count µ0 increases from 5 to 10 (Figure 1a–f), given a fixed ρ and ϕ or vice versa. Since µ0 represents the abundance of gene expression, this suggests that a larger n for a lowly expressed gene is required in order to achieve an empirical power close to 0.80 in the DEG analysis. Furthermore, we need to point out that the values of n are not similar for a fold change of 0.5 and 2 because there is no symmetry between laws in H0 and H1. Given different confounding covariates, similar changes of n for the parameter settings (ρ, µ0 and ϕ) are observed in Figure 2 and Figure 3.Next, we examined the change of n between the confounding covariates in Figure 1. We observed that the confounder-adjusted n is generally larger than that for a non-adjusted one. The colored-coded bars show that the adjusted n obtained from the confounding covariate with the difference of the mean in two groups, such as N0(0, 1) and N1(1.5, 1) in green, N0(40, 52) and N1(42, 52) in cyan and N0(30, 32) and N1(40, 32) in magenta, are much larger than a non-adjustment in orange. However, the n obtained from the equal mean confounders N0(0, 1) and N1(0, 1.52) in yellow, and N0(40, 22) and N1(40, 52) in azure either has no effect or a small effect compared to the non-adjustment. The confounding covariates corresponding to the adjusted n from largest to smallest are: N0(30, 32) and N1(40, 32) > N0(0, 1) and N1(1.5, 1) > N0(40, 52) and N1(42, 52) ≥ N0(0, 1) and N1(0, 1.52), and N0(40, 22) and N1(40, 52). This indicates that a larger n is required in highly heterogeneous data to achieve a desired power compared to homogeneous data (no-confounder present), which is expected.Figure 2 displays the n obtained when the confounding covariate follows a Poisson or normal distribution with different settings between two groups. The color-coded bars illustrate that a larger n is required for all of the confounding covariates relative to when there was no confounder present. The n changes with the characteristics of confounding covariate heterogeneity (Figure 1 and Figure 2) are similar. We observed that the greater the mean difference of the confounders between the two groups, the larger n is required to achieve a desired power. The confounding covariates corresponding to the adjusted n from largest to smallest are: Pois0(10) and Pois1(15) in azure > Pois0(25) and Pois1(20) in magenta > Pois0(10) and N1(12,1) in yellow > Pois0(10) and Pois1(12) in cyan > Pois0(10) and N1(12, 102) in green. It is interesting to observe that a different distribution of confounders between the two groups, such as a Poisson distribution Pois0(10) with a mean and variance of 10 and normal distribution N1(12,1) with a mean of 12 and variance of 1, requires a larger n compared with the same distribution of the covariate (Pois0(10) and Pois1(12)) or different distribution of the covariate with same variance Pois0(10) and N1(12, 102). This suggests that high variances in confounding covariates can affect sample size.Figure 3 lists the n calculated from the simulated data when the covariate is a categorical confounder. In this scenario, the confounder covariate X0 and X1 take the binary value 0 and 1 for the control and treatment group, respectively. Six different settings are denoted as I(0.25, 0.25, 0.25, 0.25), II(0.2, 0.3, 0.3, 0.2), III(0.3, 0.2, 0.2, 0.3), IV(0.1, 0.4, 0.4, 0.1), V(0.4, 0.1, 0.1, 0.4) and VI(0.1, 0.3, 0.4, 0.2). Each of the six settings corresponds to the different proportion of the single confounder in two groups (0,1) such as sex (male, female). For example, the IV(0.1, 0.4, 0.4, 0.1) has high disproportion between control and treatment groups, which stands for 10% female and 40% male in the control group, and 40% female and 10% male in the treatment group. Compared with no confounders or an equal proportion between the two groups (I (0.25, 0.25, 0.25, 0.25)), we observed a larger n at high proportion between the two groups is required. The categorical confounding covariates corresponding to the adjusted n from largest to smallest are: IV(0.1, 0.4, 0.4, 0.1) in cyan and V(0.4, 0.1, 0.1, 0.4) in purple > II(0.2, 0.3, 0.3, 0.2) in green, III(0.3, 0.2, 0.2, 0.3) in light green and VI0.1,0.3,0.4, ,0.2 in magenta ≥ I(0.25, 0.25, 0.25, 0.25) in yellow. In summary, the greater the heterogeneity of the confounding covariate, the larger n is required to achieve a desired power 0.80 with a significance level alpha of 0.05 compared to the homogeneous covariates such as N0(0, 1) and N1(0, 1.52), N0(40, 22) and N1(40, 52) and I(0.25, 0.25, 0.25, 0.25).Figure 4 illustrates the n adjusted by two confounders. In the upper panel A (a–f), a larger n in cyan is observed for the confounding variable N0(0, 1) and N1(1.5, 1) combined with the high disproportion covariate of II(0.1, 0.4, 0.4, 0.1). Similarly, a larger n in cyan (the bottom panel B: g–l) is observed for N0(40, 22) and N1(40, 52) combined with the high disproportion covariate of II(0.15, 0.35, 0.35, 0.15) compared to I(0.2, 0.3, 0.3, 0.2) and III0.1,0.3,0.4,0.2 with low disproportion. The two confounding covariates corresponding to the adjusted n from largest to smallest are: {N0(0, 1) and N1(1.5,1), II(0.1, 0.4, 0.4, 0.1) > {N0(0, 1) and N1(1.5, 1), II(0.2, 0.3, 0.3, 0.2) or III(0.1, 0.3, 0.4, 0.2)}[5,9] > {N0(40, 22) and N1(40, 52), II(0.15, 0.35, 0.35, 0.15)} > {N0(40, 22) and N1(40, 52), II(0.2, 0.3, 0.3, 0.2) or III(0.1, 0.3, 0.4, 0.2)}. While compared to the results from the single confounding variable in Figure 1, Figure 2 and Figure 3, we observed a larger n is required for adjusting two confounding variables such as N0(0, 1) and N1(1.5, 1) combined with I(0.15, 0.35, 0.35, 0.15) or II(0.1, 0.4, 0.4, 0.1). However, there is no significant difference in the n for the covariate with equal mean (e.g., (N0(40, 22) and N1(40, 52)) combined with a categorical covariate at smaller disproportion (II(0.2, 0.3, 0.3, 0.2) and III(0.1, 0.3, 0.4, 0.2)). The objective is to calculate the n and actual power for testing multiple genes via rejecting at least one null hypothesis when given a set of genes. In this simulation, the total number of genes per sample T is set to be 10000, true positive genes (DEGs) T1 is set to be 200. Thus, we have the number of true negative t0=T−T1, which is the number of genes that are not differentially expressed under H0. The expected number of true DEGs for a desired power 0.80 is t1=160. The rest of the parameters, including µ0, w, ρ and ϕ, remain the same as in testing a single gene. Thus, a significance level α* in the Equation (6) is calculated as 0.000859, given a nominal FDR f=0.05. For each combination of these parameter settings, the n is calculated when the observed power is close to the nominal power of 0.80. The gene between two treatment groups for the multiple corrections is considered to be significantly different only when a p-value is ≤ α*2=0.000043 using a two-sided test. The actual power is imputed as the percentage of the number of times that the null hypothesis is rejected at the significance level α*/2 in the 1000 simulated dataset. Results for each combination of the desired parameters are described below.Figure 5, Figure 6, Figure 7 and Figure 8 list the n for testing multiple genes in combinational settings corresponding to Figure 1, Figure 2, Figure 3 and Figure 4 for testing a single gene, respectively. The confounding covariates in Figure 5 and Figure 6 are continuous variables following either a normal or Poisson distribution. The confounder in Figure 7 is a categorical covariate. The sample sizes in Figure 8 are obtained by adjusting two confounding variables. We observed that the pattern of the n changed from different combinations of µ0, ϕ and ρ in Figure 5, Figure 6, Figure 7 and Figure 8 is similar to the one observed in Figure 1, Figure 2, Figure 3 and Figure 4. However, given the similar setting, a much larger n for each group is required to achieve the desired power of 0.80 with α* when testing multiple genes compared to a single gene.We used a colon adenocarcinoma (COAD) data set to illustrate how to calculate sample sizes that are adjusted by age, sex or both in the case of testing multiple genes. The mapped raw reads with 20,531 genes and 500 samples from the file (COAD.mRNAseq_raw_counts.txt) and corresponding clinical matrix data with 459 samples and 3222 covariates from the file (COAD.clin.merged.txt) were downloaded from the Broad GDAC Firehouse on 22 January 2020 (https://gdac.broadinstitute.org). The COAD data file was used in this study is provided (Supplementary File S1). With the aid of R scripts, we extracted 359 COAD and 41 uninvolved tissue samples that were adjacent to the COAD primary tumors called the normal group in this study. The age and sex for these samples are matched using the COAD.clin.merge.txt file. The genes with more than 60% zero counts across all the samples in both groups and the mean counts across the sample fewer than five were filtered out. A total of 16682 genes remained for downstream analysis. We used the edgeR package to perform the analysis [12]. Briefly, the raw read counts with 500 samples containing 16682 genes were loaded into edgeR for estimating common dispersion and normalization factors (size factor). The TMM (trimmed-mean M value) normalization method from edgeR was used to estimate the size factor. The ratio of the size factor (w) between the normal and COAD groups is 1.05. The estimated common dispersion ϕ is approximately 0.53 [11].For the confounding covariate age, we estimated the sample mean and variance using the TMM normalized data for the normal and COAD groups. The mean age in the normal and COAD groups is 70.34 and 66.88 years, respectively. The standard deviation of age in the normal and COAD groups is 13.23 and 13.1 years, respectively. Thus, we set age as N0(70, 132).) and N1(67, 132). For the categorical covariate sex, the proportion of males and females in the normal group is 0.24 and 0.26, respectively, while the proportion of males and females in the COAD is 0.26 and 0.24, respectively. Thus, we set sex as VII (0.24, 0.26, 0.26, 0.24) for sample size estimation. We assumed that the top 500 of 16682 genes are likely prognostic genes (DEGs) and have the largest FC for up or down-regulated genes. The sample size was estimated by setting the mean counts in the control group to be µ0=2, 5 and 10 for the genes in different scenarios. In this study, the nominal power is set to be 0.80, which indicates that 400 or more out of the 500 differentially expressed genes (DEGs) will be detected. Given the FDR at f=0.05 and a 0.80 nominal power, we set T=16682, T1=500, t0=T−T1 and t1=400 (the expected DEG). The FC is set to be ρg=0.5, 1.5 and 2 with ϕ ≈0.53. With these settings, the new alpha α*=0.0013 is obtained from the formula (5) at a desired t1=400. Finally, the n and actual power are estimated using α*/2 and a nominal power 0.80 (Table 2). Table 2 reports sample size n in the control and the COAD groups with and without covariate-adjusted by the age, sex and both while assuming 500 DEGs. For the 2 FC of down-regulated genes at ρ = 0.5, the minimum n for the case of non-adjustment is 107, given the minimum mean reads of the gene in the control group µ0=2. As the µ0 increases to 5 and 10, the n decreases to 71 and 59, respectively. We observed that the n adjusted by the age or sex and both variables is slightly larger than that of non-adjustment in some of the settings. However, the samples size n adjusted by both of age and sex is slightly larger than non-adjustment for all the settings. Similar results are observed for upregulated genes with ρ = 1.5 and 2. This indicates that age and sex could be the potential confounding variables in the COAD RNA-seq data.In this study, we performed both non-covariate and covariate-adjusted sample size and power calculations using simulated data as well as a real dataset. Taking the confounding covariates into consideration is an extension of our previous work [4,5]. This approach is an advancement over the current methods for sample size calculation in designing RNA-seq experiments [1,2,3,6,7,8,13,14]. Based on our knowledge, currently, there are no existing methods for calculating sample sizes by adjusting confounding covariates for buck RNA-seq experimental design. Therefore, there are no benchmark comparisons in our study. More importantly, our simulation-based method for estimating sample size and power described here is quite flexible and very useful to apply in both basic science and clinical science RNA-seq data.In performing the simulation studies, we considered different scenarios for the confounding covariates with a different data type and distribution. We found that a large sample size is required to achieve the desired 80% detection power when the heterogeneous confounding variables exist. Without consideration of cofounding covariates, the sample size obtained by the methods will likely be underestimated. Consequently, the power for detecting the DEGs will probably be below the desired power of 0.8. Similarly, we used a two-sided statistical test for the model parameter ψ1 from a standard GLM to estimate sample size and power, which are based on the DEG analysis in RNA-seq data [5]. We have incorporated a common dispersion parameter, the size factor and confounding covariates via a log link function using an NB regression model, which is extended from the previous study [23].Most importantly, in this paper, sample size calculation methods are presented under a wide range of settings for accommodating confounding covariates denoted by a continuous [24], a categorical variable [22] or both. The actual power in this study is very close to or higher than the nominal power of 0.80 for all the settings. The results indicate the required sample size is larger given additional heterogeneity in the data, which needs to be addressed in RNA-seq studies.In the simulation study, we arbitrarily chose µ0=5 as a minimum read in control group of DEGs, which is commonly used as a cutoff to filter out lowly expressed genes in RNA-seq analysis. For a low µ0, a study requires a large n to achieve a nominal power at 80% or higher, which may not be feasible in practice due to the cost. As an alternative, a higher read depth sequencing may be chosen to increase the mean read counts for each sample instead of directly increasing the sample size, as is shown in Lamarre et al. [25]. In current simulations, µ0 parameters are simply fixed as 5 and 10. For the real dataset, µ0 varies with sequencing read depth and experimental conditions. For differentially and highly expressed genes in an experiment, the µ0 could be chosen to be larger than 5 or 10 or vice versa. Moreover, when testing multiple genes in the simulation, we arbitrarily chose 10000 genes with 200 true DEGs (Figure 5, Figure 6, Figure 7 and Figure 8). In reality, the total number of detected genes could vary depending upon the read depth in each sequencing sample and experimental conditions. In this example analysis, we demonstrate that the sample size is calculated based on the number of genes and parameters that are estimated from real RNA-seq data. Due to the large sample size from COAD data, we set 500 true DEGs to estimate the sample size with a desired power. Currently, the number of DEGs identified by the common RNA-seq analysis tools is varied due to high false-positive rates [18]. Determining the true number of DEGs is usually objective by researchers because it depends on the tools and the cutoff value of fold change and adjusted p-value that are chosen. Finally, in this study, we focused on the equal read depth due to improvements in RNA-seq technology and library preparation. We also focused on a balanced experimental design for the simulation study.In summary, the methods described here illustrate how to estimate sample size when confounding variables are likely to exist in any complex RNA-seq experimental design. We observed that a larger sample size is required for the likely presence of single or multiple confounding variables in order to achieve a nominal power of 0.80. The results provide investigators with a variety of choices for the sample size that might be required for designing their experiments. Most importantly, when a confounding covariate with a known distribution exists in an experiment, one should incorporate such information into sample size calculation.The following are available online at https://www.mdpi.com/article/10.3390/biomedinformatics1020004/s1: File S1: R source codes for the simulation study with detailed explanations are provided. File S1 R codes in PDF format illustrate how to estimate sample size and power for testing a single gene for Figure 1 and multiple genes for Figure 5 given FC = 2 and other parameters in the presence of confounding covariates. File S2: Datasets used for the analysis. This zipped file folder contains COAD raw reads of 500 samples (COAd.uncv2.mRNAseq_raw_counts.txt).Conceptualization, X.L. and S.N.R.; formal analysis, X.L.; methodology, X.L.; software, X.L.; writing—original draft preparation, X.L. and T.E.O.; writing—review and editing, T.E.O., S.N.R., E.C.R. and N.G.F.C.; supervision, E.C.R. and N.G.F.C.; funding acquisition, N.G.F.C. All authors have read and agreed to the published version of the manuscript.This research was supported by the National Institutes of Health grant, P20GM103436. The R code and datasets used in this study are available (Additional File S1 and File S2), respectively.The authors gratefully thank the reviewers for the very good suggestions, advice and comments.The authors declare no conflict interests. Calculated sample size n and actual power adjusted by a confounder with a normal distribution. The color-coded bars represent covariates, and the height of the bars represents the sample size given α for testing a single gene. (a–c) shows n vs. fold change ρ given dispersion ϕ (0.1, 0.2, 0.5) and mean counts in control µ0=5. (d–f) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=10.Calculated sample size n and actual power adjusted by a confounder with a Poisson distribution or a mixture of normal and Poisson distribution. The color-coded bars represent confounding covariates, and the height of the bars represents n given α for testing a single gene. (a–c) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=5. (d–f) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=10.Calculated sample size n and actual power adjusted by a categorical confounder. The color-coded bars represent confounding covariates, and the height of the bars represents n given α for testing a single gene. (a–c) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=5. (d–f) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=10.Calculated sample size n adjusted by two confounders. The color-coded bars represent categorical confounders (Covariate 2), and the height of the bars represents n given α for testing a single gene. The confounder (Covariate 1) in panel A (a–f) has a normal distribution: N0(0, 1) and N1 (1.5, 1). The confounder (Covariate 1) in panel B (g–l) has a normal distribution: N0 (40, 52 ) and N1 (42, 52 ).Calculated sample size n adjusted by a confounder in a normal distribution. The color-coded bars represent confounding covariates, and the height of the bars represents n given α* for testing 10000 genes. (a–c) shows n vs. fold change ρ given dispersion ϕ (0.1, 0.2, 0.5) and mean counts in control µ0=5. (d–f) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=10.Calculated sample size n adjusted a confounder in a Poisson distribution or a mix of normal and Poisson distribution. The color-coded bars represent confounding covariates, and the height of the bars represents n given α* for testing 10000 genes. (a–c) shows n vs. fold change ρ given dispersion ϕ (0.1, 0.2, 0.5) and mean counts in control µ0=5. (d–f) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=10.Calculated sample size n adjusted by a categorical confounder. The color-coded bars represent confounding covariates, and the height of the bars represents n given α*. (a–c) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=5. (d–f) shows n vs. ρ given ϕ (0.1, 0.2, 0.5) and µ0=10 for testing 10000 genes.Calculated sample size n adjusted by two confounders. The color-coded bars represent categorical confounders (Covariate 2), and the height of the bar graph represents the n given α* for testing 10000 genes. The confounder (Covariate 1) in the panel A (a–f) has a normal distribution: N0(0, 1) and N1 (1.5, 1). The confounder (Covariate 1) in the panel B (g–l) has a normal distribution: N0 (40, 52 ) and N1 (42, 52 ).A summary of the simulation characteristics for the sample size calculation illustrated in Figure 1, Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8.The parameter settings are ϕ(dispersion) = (0.1, 0.2, 0.5), µ0(mean read counts of control) = (5, 10),ρ(fold change) = (0.5, 1.5, 2, 3), α (type I error rate) = 0.05, α*adjusted α=0.000859 and a nominal power at 0.8.The calculated sample size n and estimates of power from COAD data.Shown are n and actual power adjusted by the confounders of age and sex variables given nominal power of 0.8 with FDR 0.05, the ratio of size factor w=1.05, dispersion ϕ=0.53 and adjusted size α*=0.0013.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-02-00005.txt

@@ -0,0 +1 @@
+Systematic reviews and meta-analyses have been increasingly used to pool research findings from multiple studies in medical sciences. The reliability of the synthesized evidence depends highly on the methodological quality of a systematic review and meta-analysis. In recent years, several tools have been developed to guide the reporting and evidence appraisal of systematic reviews and meta-analyses, and much statistical effort has been paid to improve their methodological quality. Nevertheless, many contemporary meta-analyses continue to employ conventional statistical methods, which may be suboptimal compared with several alternative methods available in the evidence synthesis literature. Based on a recent systematic review on COVID-19 in pregnancy, this article provides an overview of select good practices for performing meta-analyses from statistical perspectives. Specifically, we suggest meta-analysts (1) providing sufficient information of included studies, (2) providing information for reproducibility of meta-analyses, (3) using appropriate terminologies, (4) double-checking presented results, (5) considering alternative estimators of between-study variance, (6) considering alternative confidence intervals, (7) reporting prediction intervals, (8) assessing small-study effects whenever possible, and (9) considering one-stage methods. We use worked examples to illustrate these good practices. Relevant statistical code is also provided. The conventional and alternative methods could produce noticeably different point and interval estimates in some meta-analyses and thus affect their conclusions. In such cases, researchers should interpret the results from conventional methods with great caution and consider using alternative methods.Systematic reviews and meta-analyses have been widely used to synthesize results from multiple studies on the same research topic in medical sciences [1,2]. The reliability of the synthesized evidence depends critically on appropriate methods used to perform meta-analyses [3,4]. However, despite the mass production of meta-analyses, it has been found that many meta-analyses need improvements in their methodological quality [5,6,7,8,9,10]. This is a particularly crucial issue in the COVID-19 pandemic because of the concerns about the expedited peer-review process [11,12,13,14].This article uses a systematic review on COVID-19, recently published in The BMJ, to illustrate some good practices for performing a meta-analysis from statistical perspectives. Many non-statistical recommendations and quality assessments for a systematic review and meta-analysis can be found in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklists [3,15,16], the GRADE (Grading of Recommendations Assessment, Development and Evaluation) approaches [17,18], the AMSTAR (A MeaSurement Tool to Assess systematic Reviews) tools, etc. [19,20]. In recent years, meta-analyses have begun to adopt these non-statistical recommendations, but there is still much room for improvement in terms of statistical analyses.For example, several papers pointed out that the well-known statistical method for the random-effects meta-analysis proposed by DerSimonian and Laird [21] is suboptimal [22,23,24]. Various methods with potentially better performance are available and can be readily implemented with various statistical software programs [25,26,27,28,29]. Nevertheless, the DerSimonian–Laird (DL) method continues to dominate contemporary meta-analyses [9]. Some popular software programs for meta-analysis (e.g., Review Manager) use the DL method as the default and perhaps the only option.In this article, based on the aforementioned systematic review on COVID-19, we aim at exploring potential issues when using statistical methods for its meta-analyses and illustrating potential better alternatives. Reproducible code for all analyses is provided. We hope these materials will help practitioners accurately use appropriate statistical methods to perform high-quality meta-analyses in the future.We use the data of meta-analyses reported by Allotey et al. [30] as our examples. This study conducted a living systematic review, which will be updated periodically to incorporate evidence from new studies. We use the version of update 1 of the original article published on 1 September 2020. The systematic review identified a total of 192 studies and performed multiple meta-analyses to investigate the prevalence, clinical manifestations, risk factors, and maternal and perinatal outcomes in pregnant and recently pregnant women (henceforth, pregnant women) with COVID-19. We select this systematic review for illustrations due to several considerations. It deals with the important research topic of COVID-19, where the appropriate use of statistical analyses is particularly crucial for timely and accurate decision-making. Also, this review covers a wide range of meta-analysis settings; the included meta-analyses had diverse outcomes, types of studies (non-comparative and comparative), numbers of studies, sample sizes, extents of heterogeneity, etc.This article uses three meta-analyses from this systematic review to illustrate several statistical advances. The first two meta-analyses synthesize comparative studies; their outcomes are fever and cough in pregnant women compared with non-pregnant women of reproductive age with COVID-19. Each meta-analysis contains 11 studies. The original meta-analysis on fever yielded a pooled odds ratio (OR) of 0.49 with 95% confidence interval (CI) (0.38, 0.63) and I2 = 40.8% suggesting moderate heterogeneity. The original meta-analysis on cough yielded a pooled OR of 0.72 with 95% CI (0.50, 1.03) and I2 = 63.6% suggesting moderately high heterogeneity. Overall, pregnant women with COVID-19 were less likely to have fever and cough than non-pregnant women with COVID-19. The association with fever was statistically significant, while that with cough was not. For illustrative purposes, Figure 1 shows the forest plot of the meta-analysis on cough.The third meta-analysis combines non-comparative data from 60 studies to obtain a pooled prevalence of COVID-19 in pregnant women; Figure 2 presents its forest plot. The original analysis gave a pooled prevalence of 7% with 95% CI (5%, 8%) and I2 = 98.0% suggesting extremely high heterogeneity.Meta-analysts should provide sufficient information of included studies so that peer reviewers and other researchers could reproduce the meta-analyses and validate the results. The PRISMA statement and its extensions give comprehensive overviews of the reporting of meta-analyses [15,16,31,32]; meta-analysts are advised to carefully follow these guidelines for general purposes. Here, we focus on the reporting from statistical perspectives; the non-statistical parts (e.g., study selection) are not discussed, while they are equally critical for validating meta-analyses. The statistical data from individual studies can be feasibly provided in meta-analyses of aggregate data. However, this practice may be challenging for meta-analyses of individual participant data (IPD), which could involve concerns about data privacy. In such situations, meta-analysts may provide detailed procedures with other researchers to apply for access to the de-identified participant-level data. In the following, we restrict the discussions to meta-analyses of aggregate data.In all three examples, the meta-analyses use aggregate data, i.e., the number of subjects with fever or cough and the sample sizes of pregnant and non-pregnant women in the comparative studies, and the number of cases of COVID-19 and the sample size of pregnant women in the prevalence data. These aggregate data are transparently provided by Allotey et al. [30], displayed in the corresponding forest plots; see, e.g., Figure 1 and Figure 2. With these data available, we can reproduce the results, such as the prevalence and OR, of each individual study. They also permit us to employ alternative meta-analysis methods (detailed later).In addition to the information of individual studies, reproducibility of meta-analyses also requires transparency in the statistical analyses, including the choice of measures for quantifying the study results, models for pooling the individual-study data, methods for assessing heterogeneity between studies and small-study effects, software program and its version used for performing the analyses, as well as subgroup analyses and sensitivity analyses (if applicable).For example, Allotey et al. [30] specified that the OR was used for pooling the comparative dichotomous data with random-effect models. If comparative continuous data are needed to be pooled with dichotomous data, the standardized mean difference was used as the effect measure of the continuous data and was transformed to the log OR using the method by Chinn [33]. For the prevalence data, the Freeman–Tukey double-arcsine transformation was applied to the proportion estimate from each study to stabilize its sample variance [34]. The authors used the I2 statistic to assess heterogeneity [35,36], but they did not assess small-study effects or publication bias. When the random-effects model is used, it is also critical to specify the estimator of the between-study variance, which is a key parameter in this model and could greatly affect the pooled results, particularly 95% CIs. The authors only specified that the DL method was used for pooling prevalence data but did not specify that for pooling comparative data. We have reproduced their meta-analyses of comparative data and found that the DL method was also used for comparative data. The original meta-analyses were all performed with Stata 16, which is widely used in the current literature of meta-analyses.Based on our knowledge, it was not uncommon that some inappropriate terminologies were used for meta-analysis methods. For example, in the systematic review by Allotey et al. [30], the prevalence was incorrectly referred to as “rate ratio” in the meta-analyses of prevalence (Figures 2 and 3 in the original article). As its name suggests, the rate ratio is a ratio of incidence rates for comparative studies, while the prevalence (or proportion) is a type of non-comparative data. The incidence rate also includes certain time elements (e.g., person-year), while the prevalence does not include such elements.Besides these minor issues in this case study, Ioannidis [37] explored the problem of massive citations in detail. Additional examples include referring to the forest plot [38] as ‘‘Forrest plots”, “honoring the nonexistent Dr. Forrest,” and the funnel plot [39,40]. for assessing small-study effects as “Beggar’s funnel plot,” “apparently copy-pasting from some original source(s) that mistyped Colin Begg’s funnel plot.” Moreover, the commonly used Q test for heterogeneity is often referred to as the “Cochrane’s Q” or “Cochran’s Q.” The former wrongly relates the Q test to the Cochrane Collaboration. The latter is used in many meta-analyses due to the paper by William G. Cochran [41], although it was not designed for testing for heterogeneity in Cochran’s original work [42].In order to use appropriate statistical methods for a meta-analysis, the first step is to specify their names correctly. When referring to certain meta-analysis methods, we suggest researchers always reading and citing the original methodological articles or tutorials that proposed, introduced, or reviewed the methods.A meta-analysis has the power to yield more precise results than individual studies, but it could also inherit potential research errors from individual studies. It may be difficult to discover and correct the errors hidden in individual studies. The potential erroneous results from suspectable studies could be removed from the meta-analysis, or sensitivity analyses could be conducted to evaluate such studies’ impact on the pooled results.Additional errors could occur when pooling the individual studies; researchers should try their best to avoid such errors when inputting the data and outputting the results. For example, a systematic review team may assign two or more researchers to independently extract individual studies’ data, perform meta-analyses, check the results, and proofread the final manuscript. With sufficient information provided, it is more likely to check for potential internal reporting discrepancies. Several examples of internal reporting discrepancies are discussed by Puljak et al. [43].Taking the systematic review by Allotey et al. [30] as an example, several discrepancies appeared. In the meta-analysis on fever comparing pregnant women with non-pregnant women with COVID-19 (Figure 5 in the original article), the OR of the study “Wei L 2020” was reported as 0.40 with 95% CI (0.11, 0.77), and the OR of another study “Wang Z 2020” was reported as 0.29 with 95% CI (0.11, 0.77). The reported CIs were identical, while the point estimates of the ORs were different, and the CIs were displayed differently in the forest plot. Based on the forest plot, the CI of “Wei L 2020” encompassed the null value 1, so the reported CI of this study was likely erroneous when copying and pasting the numeric results in the forest plot. Fortunately, because the event counts and sample sizes (8 and 17 for pregnant women and 18 and 26 for non-pregnant women) were reported for this study, we can derive the correct 95% CI as (0.11, 1.40). A similar issue occurred in the study “Zambrano LD 2020,” whose OR was reported as 0.52 with 95% CI (0.50, 0.50). The CI did not even encompass the point estimate; again, this was likely due to a typesetting error. In the meta-analysis on cough (also Figure 5 in the original article), the total sample sizes of pregnant and non-pregnant women across the 11 studies were reported as 5468 and 75,053, respectively. These total sample sizes were also apparently erroneous because they are smaller than the sample sizes in the single study “Zambrano LD 2020.” The correct total sample sizes should be 17,806 and 222,493 (Figure 1).As mentioned earlier, the example meta-analyses were performed with the random-effects model, and the well-known DL method was used to estimate the between-study variance. The DL estimator is based on the method of moments. This method is popular possibly because it is a simple, non-iterative method with a closed-form [21]. Many alternative estimators have been proposed for the between-study variance [44,45,46]. Although the DL estimator retains its usefulness in some situations (e.g., large sample sizes) [22], it could bias the estimated between-study variance, and the restricted maximum-likelihood (REML) estimator generally performs better among various frequentist methods [25,26,47]. Bayesian methods can also be good alternatives as they have the ability to incorporate prior information (e.g., from external evidence or experts’ opinions) in the final estimates [48,49]. The method used to estimate heterogeneity plays a crucial role in a meta-analysis because it could greatly affect the estimated overall effect, particularly the width of its CI and thus the statistical significance. Therefore, we suggest researchers exploring alternative options for estimating the between-study variance offered by the software programs used for their meta-analyses. In many cases, the alternative estimators may produce similar results to the DL estimator, and the DL estimator may be considered reliable. However, if these estimators yield fairly different results, researchers may consider alternative estimators.In the example meta-analysis on fever, the DL method estimated the between-study variance as 0.053, leading to an overall OR estimate of 0.488 with 95% CI (0.377, 0.632). Using the REML method, the estimate became 0.127, leading to an overall OR estimate of 0.453 with 95% CI (0.326, 0.629). In the example meta-analysis on cough, the DL method estimated the between-study variance as 0.168, leading to an overall OR estimate of 0.719 with 95% CI (0.502, 1.031). Using the REML method, the estimate became 0.239, leading to an overall OR estimate of 0.711 with 95% CI (0.476, 1.061).Conventionally, the CI of the overall estimate in a meta-analysis is produced assuming normality (e.g., for the log OR). However, this normality assumption might be questionable in some situations [50]; as such, the normality-based CI may not have the desired coverage probability (e.g., 95%). Hartung and Knapp [51,52] and Sidik and Jonkman [53,54] independently introduced a refined CI based on the t-distribution for the random-effects meta-analysis. This t-based CI has been shown to have better coverage probabilities than the standard normality-based CI by various simulation studies, particularly when a meta-analysis only contains a few studies [29,55,56,57]. Of note, this CI was designed for the random-effects meta-analysis, and it is inappropriate to apply it to the fixed-effect (also known as common-effect) meta-analysis that assumes no heterogeneity.The t-based 95% CI of the overall OR in the example meta-analysis on fever was (0.352, 0.678) and (0.316, 0.650) using the DL and REML methods, respectively, both wider than their counterparts of normality-based 95% CIs (0.377, 0.632) and (0.326, 0.629). Similarly, in the example meta-analysis on cough, the t-based 95% CI of the overall OR was (0.463, 1.118) and (0.453, 1.114) using the DL and REML methods, respectively. Also, both were wider than their counterparts of normality-based 95% CIs (0.502, 1.031) and (0.476, 1.061).Heterogeneity between studies frequently appears and is generally expected in a meta-analysis [58]. Standard meta-analysis approaches use the random-effects model to account for the heterogeneity and use the estimated between-study variance τ2 and/or the I2 statistic to quantify it. However, it is difficult to apply these metrics to clinical practice for future research. Over the last decade, much effort has been made to promote the reporting of the prediction interval (PI) in a meta-analysis, but only a small proportion of meta-analyses adopt this recommendation in the current literature [9,59,60,61,62,63]. The PI represents the expected range of the true effects in future studies, making it easier to apply meta-analysis results to clinical practice. The PI is wider than the CI due to the heterogeneity between existing studies in a meta-analysis and future studies. A meta-analysis may have a CI not encompassing the null value (thus implying a statistically significant effect), but its PI could encompass the null, indicating that a future study could have opposite results [64].Despite the attractive features of the PI, researchers should note that the PI could be subject to large uncertainties when the number of studies in a meta-analysis is relatively small (e.g., <10). In the presence of small-study effects (detailed in the following subsection), the PI could have poor coverage due to biased estimates. Therefore, the PI should be interpreted with caution in these situations. Also, the PI is designed for a random-effects meta-analysis; it is not sensible for a fixed-effect meta-analysis.In the example meta-analysis on fever, based on the REML estimator of the between-study variance, the 95% PI of the overall OR is (0.186, 1.104), encompassing the null value 1. Recall that the 95% CI of this meta-analysis is (0.326, 0.629), not encompassing 1. Therefore, although the meta-analysis concludes a statistically significant association between fever and pregnancy, this conclusion could be changed in a new study.In the example meta-analysis on cough, based on the REML estimator of the between-study variance, the 95% PI of the overall OR is (0.214, 2.356), much wider than its 95% CI (0.476, 1.061). The PI can be incorporated into the forest plot [60,65], as shown in Figure 1. Of note, the results in Figure 1 were produced using the DL method to reproduce the original results in Allotey et al. [30], so the interval estimates were different from the foregoing results based on the REML method.Small-study effects refer to the phenomenon that smaller studies containing fewer subjects have substantially different results from larger studies with more subjects. They could be caused by publication bias, when small studies with statistically significant findings or effect estimates in the desired direction are more likely published in the literature than those with non-significant findings or effect estimates in the opposite direction [66,67]. Assessing small-study effects is a crucial step for validating the synthesized evidence from a meta-analysis; if substantial small-study effects appear, the certainty of the synthesized evidence should be rated down [3,68,69]. Common approaches to assessing small-study effects include graphical tools, such as the funnel plot [39,40], and quantitative methods, such as Egger’s test, Begg’s test, and skewness [70,71,72,73,74]. The asymmetry in a funnel plot is an indicator of potential small-study effects. Additional contours that depict areas of various statistical significance levels can be further added to the usual funnel plot, referred to as the contour-enhanced funnel plot [40,75,76]. They help distinguish publication bias from other potential factors (e.g., subgroup effects) that might cause small-study effects.We assessed small-study effects in the meta-analyses on fever and cough. Figure 3 presents their contour-enhanced funnel plots, where the contours represent the commonly used statistical significance levels at 0.01, 0.05, and 0.1.In Figure 3A, the funnel plot shows that the (log) ORs from the 11 studies on fever were distributed asymmetrically. Smaller studies with larger standard errors tended to have smaller ORs away from the null value 1, indicating small-study effects. The potential missing studies at the lower right part of the funnel plot are likely located within the white area with p-values > 0.1. Therefore, this contour-enhanced funnel plot supports the existence of publication bias. Nevertheless, the p-value of Egger’s test was 0.275, suggesting that publication bias was not statistically significant. Of note, if the potential missing studies were located in areas with very small p-values, then the small-study effects may not be explained by publication bias. In such cases, meta-analysts are encouraged to explore the factors that might cause the funnel plot’s asymmetry, e.g., by performing subgroup analyses to examine whether the asymmetry was attributable to heterogeneity [40,77].Although small-study effects were not statistically significant based on Egger’s test in this case study, it did not mean that the assessment of small-study effects was unnecessary. Statistical methods for detecting small-study effects usually have low powers, particularly in meta-analyses with only a few studies. As such, the significance level for detecting small-study effects is typically set to 0.1, higher than the most popular cutoff of 0.05 [78]. Here, meta-analysts should distinguish the p-value of tests for small-study effects from the p-values of individual studies’ effect estimates. The significance levels depicted in the contour-enhanced funnel plot are intended for the latter. In addition, if a meta-analysis contains less than 10 studies, it might be inappropriate to use the funnel plot to detect small-study effects because it is hard to distinguish chance from real asymmetry [40].In Figure 3B, the funnel plot for the meta-analysis on cough does not show apparent missing studies in the white area of non-significance. Therefore, it does not support the existence of publication bias.Conventionally, meta-analyses are performed with two-stage methods; that is, within-study estimates are first obtained, and then the study-specific estimates are pooled together as an overall estimate. The two-stage methods are usually simple and intuitive; the study-specific estimates provided by them are also necessary for producing the forest plot for visualizing a meta-analysis and the funnel plot for assessing small-study effects. Nevertheless, they suffer from several limitations. First, the study-specific estimates in the two-stage methods are typically assumed to approximately follow normal distributions. For this purpose, certain transformations are applied to the original effect measures. For example, the OR is typically analyzed on the logarithmic scale, and the Freeman–Tuckey double-arcsine transformation is widely used to transform proportion estimates, as in the original analyses by Allotey et al. [30]. The transformed estimates may approximately follow normal distributions when the sample sizes are sufficiently large, while the approximation may be inaccurate for studies with small sample sizes [50]. In recent years, there are also growing concerns about the appropriateness of the Freeman–Tuckey double-arcsine transformation for meta-analyses of proportions [79,80,81]. Second, the variances of the effects from individual studies need to be estimated in the two-stage methods, and the estimated within-study variances are typically treated as fixed variables. Again, this practice may be valid for large-sample settings, but it is questionable for studies with small sample sizes [57]. For example, the (log) OR’s variance depends on the event counts, which are actually random variables instead of fixed variables. The (log) OR and variance are thus intrinsically associated, and such association could lead to non-negligible biases for small sample sizes and/or low even rates [82,83,84].With the recent development of statistical methods for meta-analysis, many software programs have commands to pool data via one-stage methods, such as generalized linear mixed models (GLMM) and Bayesian hierarchical models. The one-stage methods assume exact likelihood functions for the observed data (e.g., the binomial likelihood for the event count from a group of patients). They do not need the estimation for each individual study and thus avoid some unrealistic assumptions made by the two-stage methods. Moreover, these methods are widely applicable to many types of meta-analyses, including comparative studies, proportions, and diagnostic tests [27,85,86,87,88,89,90]. In the following, we illustrate the use of GLMMs and Bayesian hierarchical models with two example meta-analyses.In the meta-analysis on cough, recall that the overall OR was 0.719 with 95% CI (0.502, 1.031) based on the original analysis (the DL estimation) by Allotey et al. [30]; it was 0.711 with 95% CI (0.476, 1.061) using the REML estimation. We re-analyzed this dataset using the GLMM and Bayesian hierarchical models with a logit link function. For the Bayesian models, we used the vague normal prior N(0, 1002) for the overall log OR and the uniform prior U(0, 5) for the between-study standard deviation τ. We also considered the informative log-normal prior LN(−2.89, 1.912) for τ2, which was derived by Turner et al. [48] based on a large Cochrane database. The GLMM estimated the overall OR as 0.710 with 95% CI (0.493, 1.022). The Bayesian model with U(0, 5) prior for  τ produced the estimated OR of 0.701 with 95% credible interval (CrI) (0.415, 1.143), and that with LN(−2.89, 1.912) prior for τ2 gave 0.709 with 95% CrI (0.462, 1.047).In the meta-analysis of the prevalence of COVID-19 in pregnant women, we re-analyzed it using the GLMM and Bayesian model with a logit link function, in addition to the original two-stage method used by Allotey et al. [30] (i.e., the DL estimation with the Freeman–Tuckey double-arcsine transformation). Based on the original two-stage method, the overall prevalence was estimated as 6.77% with 95% CI (5.28%, 8.44%). Based on the GLMM, the estimated overall prevalence became 5.44% with 95% CI (4.09%, 7.19%). The Bayesian model with U(0, 5) for τ produced the estimated OR of 5.44% with 95% CrI (4.04%, 7.34%). The prevalence estimates by both one-stage methods were smaller than those by the two-stage method by over 1%.This article provided a summary of good practices for performing a meta-analysis from statistical perspectives. We illustrated these practices using meta-analyses published in a recent systematic review on COVID-19 in pregnancy. We hope they may help improve the methodological quality of future meta-analyses. For facilitating researchers to implement the methods reviewed in this article, the Supplemental File gives all code for our analyses.Due to the urgent need for COVID-19 research, it has been dramatically expedited to conduct and peer-review meta-analyses. Nevertheless, it is critical to safeguard the integrity of scientific evidence during this challenging period of accelerated publishing [14]. This article shows that some statistical methods used in the example meta-analyses may be suboptimal. In our re-analyses with better alternatives, some meta-estimates had noticeable changes. Also, potential small-study effects might exist. Extra attention is needed to examine whether such effects might continue to exist in the future updates of this living systematic review after including new studies.This article has several limitations because we were only able to focus on select statistical advances for meta-analysis based on a single case study on COVID-19. For example, for assessing small-study effects or publication bias, some selection models may be applied as sensitivity analyses to examine the robustness of synthesized results to potential bias [91,92]. Alternative meta-analysis methods are available to offer some benefits over the traditional fixed-effect and random-effects models under specific cases [93,94,95]. In addition, the current literature has debates on the choice of effect measures, e.g., relative risk, in meta-analyses [96,97]. This article has also not covered topics on meta-analyses of diagnostic tests [98]. All examples are meta-analyses of aggregate data, while meta-analyses of IPD may involve additional issues and require specific methods [99]. For a more comprehensive review of meta-analysis methods, one may refer to the Cochrane Handbook [100].Systematic reviews and meta-analyses are a type of transdisciplinary research. Therefore, in addition to many statistical considerations reviewed in this article, non-statistical guidance is also crucial for conducting high-quality meta-research. For example, heterogeneity between studies may be assessed beyond the statistical perspectives [101]. To aid the statistical assessment of small-study effects, researchers are suggested to search for relevant unpublished studies (e.g., on preprint servers and trial registries), include them in meta-analyses, and explore their potential differences from the published studies [100]. Of course, because the unpublished studies are not peer-reviewed, they could be subject to a high risk of bias. The risk of bias must be carefully appraised if incorporating such studies in the systematic review [102].The following are available online at https://www.mdpi.com/article/10.3390/biomedinformatics1020005/s1: code for producing the results presented in the main content.This research received no external funding.Ethical review and approval were waived for this study because we focused on statistical methods for meta-analysis and used published data. Patient consent was waived because this study used published data.The Supplementary Materials include the data used in this study.The authors declare no conflict of interest.Forest plot of the meta-analysis on cough among pregnant women compared with non-pregnant women of reproductive age with COVID-19.Forest plot of the meta-analysis of the prevalence of COVID-19 in pregnant women.Contour-enhance funnel plots for the meta-analyses on fever (A) and cough (B). The dashed lines represent the fixed-effect estimate and the corresponding 95% confidence limits, and the dotted vertical line represents the random-effects estimate.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00006.txt

@@ -0,0 +1 @@
+Background: Multiple myeloma (MM) is one of the most common cancers of the blood system. N6-methyladenosine (m6A) plays an important role in cancer progression. We aimed to investigate the prognostic relevance of the m6A score in multiple myeloma through a series of bioinformatics analyses. Methods: The microarray dataset GSE4581 and GSE57317 used in this study were downloaded from the Gene Expression Omnibus (GEO) database. The m6A score was calculated using the GSVA package. The Random forests, univariate Cox regression analysis and Lasso analyses were performed for the differentially expressed genes (DEGs). Kaplan–Meier analysis and an ROC curve were used to diagnose the effectiveness of the model. Results: The GSVA R software package was used to predict the function. A total of 21 m6A genes were obtained, and 286 DEGs were identified between high and low m6A score groups. The risk model was constructed and composed of PRX, LBR, RB1, FBXL19-AS1, ARSK, MFAP3L, SLC44A3, UNC119 and SHCBP1. Functional analysis of risk score showed that with the increase in the risk score, Activated CD4 T cells, Memory B cells and Type 2 T helper cells were highly infiltrated. Conclusions: Immune checkpoints such as HMGB1, TGFB1, CXCL9 and HAVCR2 were significantly positively correlated with the risk score. We believe that the m6A score has a certain prognostic value in multiple myeloma.Multiple myeloma (MM) is the second most common hematologic malignancy [1]. Its main characteristics are plasma cell cloning and proliferation and heterogeneous genome landscape [2]. MM has been treated with immunomodulatory drugs, protease antibodies, monoclonal antibodies and stem cell transplantation in the past few decades. However, it is still a type of incurable plasma cell malignancy [3,4]. Meanwhile, the MM has been reported, and its development is evolutionary [5]. It obviously exacerbates the difficulty of treatment. Fortunately, the development and application of microarray technology and bioinformatics analysis could help to screen and identify genetic variation in the course of the disease to a certain extent. Now, risk-adapted therapies are becoming the new standard of care. Survival analysis is essential in treatment decisions and clinical studies. Therefore, there is an urgent need to develop reliable and robust models to estimate patient survival from large amounts of data.N6-methyladenosine (m6A) is one of the post-transcriptional modifications of RNA and one of the most common internal chemical modifications in eukaryotes and nuclear replicating viruses [6]. M6A-related regulatory factors can be involved in various physiological and pathological processes by regulating RNA stability, mRNA splicing or translation [7,8,9,10]. Several studies have reported that m6A was involved in the occurrence and progression of a variety of cancers, including colorectal cancer [11], endometrial cancer [12], oral squamous cell carcinoma [13], osteosarcoma [14] and glioma [15]. However, the clinical value and potential mechanism of m6A-related genes in multiple myeloma still need to be further explored.However, no one has analyzed its prognostic role in multiple myeloma based on the m6A score. In this study, we intend to download multiple myeloma data from the GEO public database. The m6A score was calculated, and a risk model was constructed. The survival analysis and function prediction were performed. Thus, it might provide a research basis for the study of m6A in multiple myeloma.In this study, all datasets used were publicly available, and the workflow is shown (Figure 1). There are multiple myeloma data sets from Gene Expression Omnibus (GEO) (http://www.ncbi.nlm.nih.gov/geo/, accessed on 13 July 2019) to download. Datasets GSE4581 (n = 413) and GSE57317 (n = 55) from the same platform (GPL570) were used for myeloma network analysis. After chi-square test analysis (Figure 1), GSE4581 was randomly divided into a training set (207 samples) and an internal testing set (206 samples) in a 1:1 ratio. GSE57317 contains 55 samples as an external validation set. The raw data from the microarray dataset were generated by Affymetrix. The quantile normalization and background correction of the raw data were performed using the Affy package’s RMA algorithm. The R Bioconductor software package and the R software (version 3.6.1) were applied to analyze all the data.A total of 21 m6A genes were obtained. The score of m6A was calculated using the GSVA package. The Gene set variation analysis (GSVA) transforms the expression matrix of genes between different samples into the expression matrix of genes set between samples to evaluate the enrichment of different pathways. The value of each pathway in each sample is given a score. Thus, the m6A score was obtained. The minimum p-value method was used to determine the two groups with high and low m6A scores.We downloaded all the gene sets from the MSIgDB database. The ssGSEA method was used to calculate the abundance of 28 types of immune cells. We used the GSVA R software package for immune checkpoint cluster and scoring by executing the GSVA, including the GO BP (Biological Process), KEGG and Hallmark Gene sets [16]. The selection criteria of immune checkpoint cluster-related pathways were based on the corrected p value less than 0.05. The selection criteria of immune checkpoint score-related pathways were based on the correlation analysis p value less than 0.05.The differentially expressed genes (DEGs) between the low and high m6A score groups were identified by the Limma package in R [17]. The p values < 0.05 and |log FC| > 1.5 sets were used to determine the immune checkpoint subtypes of DEGs in the significance standard [18]. Then, the Random forest in the machine learning method of the Caret package was used for gene dimensionality reduction, and Univariate Cox regression analysis was applied to determine representative genes. The prognostic genes and their Lasso regression coefficients were obtained by selecting the highest lambda value (“min” lambda) through 1000 cross-validation in the training set using the Lasso method. The risk score is the sum of the expression value of genes screened by LASSO * the regression coefficient of LASSO.  * represented a multiplication symbol.The R package ggplot2 was used to visualize the data. The Benjamini–Hochberg method was applied to analyze differentially expressed genes by converting p-values to FDR to identify important genes. The receiver operating characteristic package (pROC) was used to generate the ROC curve and calculate the area under the curve (AUC) [19]. The generation and visualization of survival curves are realized by applying Kaplan–Meier analysis. The Logarithmic rank test was used to determine the statistical significance of the differences in each dataset. The R package survminer was adopted to generate all survival curves. The pheatmap was applied to generate all the heat maps. All data statistical analysis was performed in R (https://cran.r-project.org/bin/windows/base/old/3.6.1/, version 3.6.1, accessed on 5 July 2019). The normality of variables was tested by using the Shapiro–Wilk normality test. For normally distributed variables, an unpaired Student’s t-test was used to compare the differences between the two groups. The Wilcoxon test was used to compare variables that are not normally distributed. One-way ANOVA analysis was applied for the parametric method to compare the mean values between multiple groups. However, the Kruskal–Wallis test was applied for the non-parametric. The two-sided was used in all tests. p values < 0.05 were considered statistically significant.We obtained 21 m6A genes in the GSE4581 dataset, and the GSVA package was used to calculate the m6A score. A heat map visualization of the expression of 21 m6A genes with m6A score changes can be seen in Figure 2A. The 19 m6A genes, including ALKBH5, YTHDF1, FTO, WTAP and YTHDF2, were significantly correlated with the m6A score. The survival time of the high m6A score group was significantly lower than that of the group with the low score (p = 0.017) (Figure 2B). Meanwhile, the high and low m6A score groups showed the difference in the biological functions, including the Cell Cycle, DNA Damage Repair, DNA Replication and EMT2 (Figure 2C). The genetic differences between the two groups’ m6A scores were analyzed by performing the Limma package and visualized by a volcano map (Figure 2D). Therefore, it is speculated that the m6A score might have a certain diagnostic value to distinguish multiple myeloma patients.Next, we analyzed the data for 28 immune cell infiltrates. The immune cells showed significant differences between the high and low m6A scores groups, including the activated dendritic cells, CD56 bright natural killer cells, Monocytes, Plasmacytoid dendritic cells, Type 1 T helper cells, Type 2 T helper cells and other immune cells (Figure 3A). We continued to analyze the situation of immune cell infiltration from low to high m6A scores (Figure 3B). With the increase in the m6A score, the abundance of Memory B cells and Type 2 T helper cells increased, while activated dendritic cells, CD56 bright natural killer cells, Effector memory CD8 T cells and Monocytes decreased. Meanwhile, we analyzed the changes of immune checkpoints in the m6A score. With the m6A scores ranked from low to high, the expression abundance of immune checkpoints showed significant differences in multiple categories, including antigen-present, cell adhesion, co-inhibitor, co-stimulator, ligand and receptor (Figure 3C).We obtained 286 DEGs between the two groups of m6A scores using Limma analysis, then 74 genes were obtained using Random forest (Figure 4A), and 13 genes were obtained using Univariate analysis (Figure 4B). The nine genes used to construct the Risk model were screened using the LASSO method, including PRX, LBR, RB1, FBXL19-AS1, ARSK, MFAP3L, SLC44A3, UNC119 and SHCBP1 (Figure 4C). Risk score = (−0.1257) * PRX + (0.1278) * LBR + (−0.1790) * RB1 + (−0.1191) * FBXL19-AS1 + (0.2528) * ARSK + (−0.1169) * MFAP3L + (−0.1636) * SLC44A3 + (−0.0158) * UNC119 + (0.1486) * SHCBP1. * represented a multiplication symbol. The heat map was used to visualize the expression of 21 m6A genes with risk scores (Figure 4D). The levels of m6A genes such as CBLL1, FMR1, YTHDF3, hnRNPA2B1 and ZC3H13 significantly changed with the risk score change.We further evaluated the prognosis of the risk score model. In the training set (Figure 5A), the validation set (Figure 5B), the overall data set (Figure 5C) and the external validation set (Figure 5D), the survival analysis showed that patients with the high-risk score had a poor prognosis, with p values less than 0.05. Then, the effectiveness of the risk score model was evaluated in the training set (Figure 5E), the verification set (Figure 5F), the overall data set (Figure 5G) and the external verification set (Figure 5H). Among them, in overall sets, the sensitivity at 1, 2 and 3 years was 0.557, 0.557 and 0.671, respectively; the specificity was 0.823, 0.823 and 0.704, respectively; the accuracy was 0.772, 0.772 and 0.697, respectively; and the AUC was 0.768, 0.701 and 0.715, respectively. The time-dependent ROC curve analysis showed that the model is effective.Then, in the GO BP (Biological Process), KEGG and Hallmark gene concentration, we used the GSVA R package to conduct related pathway analysis for the risk score (Figure 6A). As the risk score went up, the Cell adhesion molecules cams, the JAK-STAT signaling pathway, the TNFA signaling pathway and the MAPK signaling pathway all decreased. The DNA repair, the MYC targets V1, the P53 signaling pathway, the Cell cycle DNA replication, the DNA replication, the Meiotic cell cycle, the Cell cycle, the Mitotic sister chromatid segregation, the G2M checkpoint, the Cell cycle G1 S phase transition and the Mitotic cell cycle checkpoint increased. The higher the risk score is, the worse the patient’s survival rate is. It was suggested that the above pathway is related to the poor survival of patients. The immune cell infiltration was ranked according to the risk score from low to high (Figure 6B). The Activated CD4 T cell, Memory B cell and Type 2 T helper cell significantly increased with increasing risk scores. The Activated CD8 T cell, CD56 bright natural killer cell, Effector memory CD8 T cell, Monocyte, the natural killer cell and Type 1 T helper cell decreased significantly as the risk score increased. The immunization checkpoint was ranked as the risk score goes from low to high (Figure 6C). The levels of HLA-C, HLA-B, SLAMF7, BTN3A1, CD27, BTLA, HLA-DPA1, HLA-DQB1 and TNFRSF4 decreased significantly with the increase in the risk score, while HMGB1, CXCL9 and HAVCR2 increased.This study performed survival analysis and functional analysis of multiple myeloma genomes based on the m6A score. A risk model based on the m6A score was established to explore the influence of risk score on function, immune cells and immune checkpoints.M6A is one of the most common RNA modifications. It has been reported that m6A disorder was closely related to the occurrence and development of cancer [20]. Our study found significant differences in the survival time and the biological function analysis between the two groups with high and low m6A scores. The higher the m6A score was, the shorter the survival time of patients was, which suggested that the study of multiple myeloma grouped by m6A score has certain clinical significance. The m6A score was negatively correlated with the survival rate of patients. It has been reported that m6A could regulate the biological process of cells by regulating the expression of genes [21,22]. In myeloid leukemia, promoter-bound METTL3 relies on translational control by m6A to maintain its status [23]. In liver cancer studies, KIAA1429 promotes cancer progression through post-transcriptional modification of GATA3 dependent on m6A [24]. All these suggested that m6A is involved in cancer development and confirmed the rationality of our grouping from the side.Then, the Limma analysis, Random forest analysis, Univariate Cox regression analysis and Lasso analysis were performed to screen the DEGs of the two groups with high and low m6A scores. The risk prognoses model was constructed, containing PRX, LBR, RB1, FBXL19-AS1, ARSK, MFAP3L, SLC44A3, UNC119 and SHCBP1. Among them, PRX is involved in the regulation mechanism of the P53 signal [25,26]. RB1 is an important indicator of cell cycle regulation [27]. MFAP3L and SHCBP1 are involved in the development of breast cancer [28,29]. Meanwhile, the enrichment of pathways changed with the risk scores, including the P53 signaling pathway, Cell cycle DNA replication, DNA replication, Meiotic cell cycle and cell cycle. These pathways are involved in the development of multiple myeloma and are associated with the poor prognosis of patients. Therefore, it could be speculated that genes related to the m6A score affect the prognosis of multiple myeloma by regulating the related signaling pathways. The effect of the m6A score on the prognosis of patients with multiple myeloma was also verified.The immune cells mainly influence the immune microenvironment of the tumor. We analyzed the invasion of immune cells and found that the immune cells, including the Activated CD4 T cell, the Memory B cell and the Type 2 T helper cell, were highly infiltrated in patients with multiple myeloma with poor prognosis. The CD4 T cell is also highly invasive in colorectal cancer patients [30]. The Type 2 T helper cells show contradictory effects in cancer immunity [31]. The opposite is true for immune cells with anti-tumor activity, including the activated CD8 T cell, CD56bright natural killer cell, Effector memory CD8 T cell, Monocyte, natural killer cell and Type 1 T helper cell, etc. GSVA was used to analyze the correlation between risk score and immune checkpoints such as the antigen present, cell adhesion, co-inhibitor, co-stimulator, Ligand and receptor. Among them, HMGB1, CXCL9 and HAVCR2 were all significantly positively correlated with the risk score. HMGB1 is one of the damage-associated molecular patterns (DAMP) and plays a multifunctional role in inflammation and the development of cancer (such as colorectal cancer) [32,33]. It has been reported that CXCL9 is associated with the infiltration of immune cells and affects the prognosis of breast cancer patients [34]. HAVCR2 inhibitors have been proven to have certain tumor-suppressive effects in various preclinical tumor models [35]. All of these suggest that the risk score model has good efficiency. Meanwhile, it is further verified that the m6A score has a certain diagnostic effect on the prognosis of patients with multiple myeloma.However, due to the lack of clinical data on multiple myeloma in the data set, our study was still limited. We were unable to further analyze its clinical features. Meanwhile, we need to verify their specific expression in clinical samples for m6A-related genes, and DEGs screened based on public data. As for the risk model constructed by nine DEGs, it still needs the support of a large number of sample data. Functional prediction also needs to be validated in concrete experimental models.In our study of multiple myeloma, the patients with a high m6A score had a poorer prognosis. The m6A score was closely related to immune cell infiltration and the immune checkpoint. Thus, the m6A score could be used as a potential prognostic marker in multiple myeloma.G.X., Q.Y. and W.W. were involved in the conception of this study. G.X. and Q.Y. analyzed the dataset and prepared the figures and tables. Q.Y. and W.W. checked the manuscript. All authors have read and agreed to the published version of the manuscript.This work is supported by a China National Science Foundation grant (81401227); Natural Science Foundation of Hunan Province, China grant (2019JJ20035); Natural Science Foundation of Hunan Province, China grant (2020JJ5942).Not applicable.Not applicable.The data supporting the findings of this study are from previously reported studies and datasets. The processed data are available from corresponding author upon request.The authors declare that there is no conflict of interest regarding the publication of this paper.The Study workflow. The prognostic gene characterization of m6A in multiple myeloma. GSE4581 (overall set) was randomly divided into a training set and an internal testing set. GSE57317 is an external validation set.The prognostic characteristics of m6A score. (A) The expression of the m6A gene. (B) The survival analysis of high and low m6A scores. (C) The biological function analysis of high and low m6A score groups. (D) The volcano maps showed different genes between high and low m6A scores. * indicates significant, with the p value less than 0.05, *** indicates p value less than 0.001, and **** indicates p value less than 0.0001.Immune infiltration and immune checkpoint changes with m6A score. (A) Invasion of immune cells between the high and low m6A score groups. (B) The abundance of immune cells with m6A score. (C) The levels of the immune checkpoint with m6A score. * indicates significant, with a p value less than 0.05, ** indicates p value less than 0.01, *** indicates p value less than 0.001, and **** indicates p value less than 0.0001.Screening of high and low m6A score genes and construction of the risk score. (A) The Random forest analysis. (B) The Univariate Cox analysis of DEGs. (C) The LASSO regression analysis. (D) The levels of m6A genes. * indicates significant, with a p value less than 0.05, *** indicates p value less than 0.001, and **** indicates p value less than 0.0001.Prognosis and efficacy evaluation of risk score model. (A–D) Survival analysis. (E–H) ROC curve.Functional sequencing of prognostic models. (A) Pathway analysis. (B) Immune cell invasion. (C) Expression of immune checkpoints. * indicates a correlation p value less than 0.05, ** indicates p value less than 0.01, *** indicates p value less than 0.001, and **** indicates p value less than 0.0001.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00007.txt

@@ -0,0 +1 @@
+Today Deep Learning (DL) is state-of-the-art in medical imaging segmentation tasks, including accurate localization of abdominal organs in MRI images. But segmentation still exhibits inaccuracies, which may be due to texture similarities, proximity or confusion between organs, morphology variations, acquisition conditions or other parameters. Examples include regions classified as the wrong organ, some noisy regions and inaccuracies near borders. To improve robustness, the DL output can be supplemented by more traditional image postprocessing operations that enforce simple semantic invariants. In this paper we define and apply totally automatic post-processing operations applying semantic invariants to correct segmentation mistakes. Organs are assigned relative spatial location restrictions (atlas fencing), 3D organ continuity requirements (envelop continuity), and smoothness constraints. A reclassification is done within organ envelopes to correct classification mistakes, and noise is removed (fencing, enveloping, noise removal, re-classifying and smoothing). Our experimental evaluation quantifies the improvement and compares the resulting quality with prior work on DL-based organ segmentation. Based on the experiments, we conclude post-processing improved the Jaccard index over independent test MRI sequences by a sum of 12 to 25 percentage points over the four segmented organs. This work has an important impact on research and practical application of DL because it describes how to post-process, quantifies the advantages, and can be applied to any DL approach.Magnetic Resonance Imaging (MRI) is an imaging technique based on capturing magnetic signal changes in the resonance of hydrogen protons after triggering radio-frequency pulses. Computerized processing of those signals outputs MRI images which can be used for diagnosing medical conditions. The resulting MRI scan is a sequence of slices, where a slice is a 2D image of the body part that is being scanned. A sequence of 2D images generates a clear 3D volume with details of the body part that was scanned. Deep learning-based segmentation networks can learn to segment automatically either the 2D slices or the 3D volumes based on training examples. They are state-of-the-art in segmentation of this and other medical imaging contexts.The segmentation network is itself an evolution of the classification Convolution Neural Network (CNN). While the CNN is an image classifier that inputs an image and outputs a classification for the image, the segmentation network classifies each image pixel, resulting in a complete segmentation of the image. The segmentation network has an encoder, which is a sequence of convolution stages (convolution layers together with regularization and pooling) that extract and compress features automatically from the original image, and a decoder, which is a sequence of deconvolution layers, and a final pixel classification layer. The decoder effectively converts the compressed features back to an image sized segmentation map.In this work, the targets of segmentation are a set of organs, in particular the liver, the spleen and the two kidneys. Figure 1 shows an example slice extracted from a full MRI abdomen sequence, with ground-truth segmentation shown on the left and segmentation results using CNN architecture named DeepLabV3 [1] on the right. It is possible to see some inaccuracies in the form of some wrongly classified pixels, most errors in the example being related to spilling to neighbor areas, and also a wrong classification of some areas as part of the organ.In order to prepare the segmentation networks for segmentation, it was necessary to train them with a large number of training sequences and corresponding ground-truths (correct segmentations) for the networks to learn how to segment specific organs. We used the Chaos challenge data [2] as the dataset. The dataset includes 120 DICOM sequences from MRI. In total, there are 1064 slices, 80% used for training and the remaining for test (10%) and validation (10%). This dataset is further augmented to double the size using data augmentation (random translations of up to 10 pixels, random rotations up to 10 degrees, shearing up to 10 pixels and scaling up to 10%). The convolution network learns to segment based on the back-propagation algorithm, which iteratively adjusts convolution and deconvolution filter weights based on gradient descent methods to progress into minimization of the loss metric using a learning rate. The loss metric itself is a function that quantifies the segmentation error, a measure of the difference between the segmentation output and the ground-truth.Many factors can contribute to inaccuracies in the results of deep learning-based segmentation, and specifically of segmentation of abdominal organs. We already pointed out textures similarities, proximity or confusion between organs, morphology variations, acquisition conditions and other parameters as the major reasons for those errors, resulting in regions classified as the wrong organ, some noisy regions and inaccuracies near borders. Some simple examples of errors include classification of parts of a left kidney as a right kidney and vice-versa, erroneous classification of parts of the spleen as kidney or vice-versa, some other structures classified as one of the organs, some parts of the background classified as a part of an organ and spilling the organ segments into neighboring regions.In all these examples a human can detect the errors, and the semantics that the human uses to detect the errors can be enforced automatically. Even a very good deep learning segmentation of abdominal organs might score between and 80% and 90% Jaccard index, leaving an additional 10 to 20% space for further improvement by post-processing. A solution for post-processing is to apply constraints based on the additional semantic invariants that are obvious to a human observer but not so obvious to the automated DL procedure. These invariants can, however, be coded as automated post-processing steps. One procedure reminiscent of atlas-based approaches is to obtain the expected 3D location and volume from the training images, with some tolerance added, whereby an organ is expected to be located in a specific region of the body and to have a certain volume (we call this procedure “fencing”). Additionally, this fencing also allows us to remove incorrect assignments and to do reclassification of regions inside the incorrect fence. Another constraint is continuity, whereby the segment of an organ is expected to have continuity inside of its 3D and 2D structures, and to form a solid organ envelope. This allows the algorithm to fill gaps and also connect parts of disconnected regions within an organ’s envelope. Additionally, small, isolated regions outside the 3D envelope can be considered noise and reclassified. Finally, smoothness is expected. Under smoothness, borders of organ volumes and slices should be smooth. In essence, the proposed approach applies a set of image processing operations using several techniques to improve segmentation.In this work, we first propose the post-processing operations, then we build an experimental setup to quantify the quality improvement. For the experimental setup we first compare the quality of segmentation of three well-known off-the-shelf segmentation networks to choose the best performing one. Before engaging in experimental work, we first tuned training parameters by evaluating the quality of segmentation, as measured by the metric IoU (intersect-over-the-union) on the test set as we varied learning rates (0.01, 0.005, 0.001, 0.0005, 0.0001), different learning algorithms (Adaptive Moment estimation = Adam, Root Mean Square Propagation, and RMSProp, Stochastic Gradient Descent with Momentum = SGDM), different numbers of epochs (70, 100, 300, 500,700), different minibatch numbers (8, 16, 32, 64) and momentum (1, 0.9). The top performing alternative was chosen (0.005 learning rate, SGDM, 500 epochs, 32 minibatch size, momentum 0.9) and used for the experimental runs. After choosing and tuning the network, it was trained with MRI sequences and then used to segment an independent set of test MRI sequences. Finally, we applied the post-processing operations to improve the results and assess the amount of improvement. Our assessment was based on evaluating the quality of segmentation of the liver, spleen, left kidney and right kidney, and the quality after post-processing, to understand how much the post-processing operations were able to improve the quality of the result. Using this experimentation approach, we were able to conclude that post-processing operations improved the Jaccard index of segmentation by 12 to 25 percentage points for total improvement over the four organs in our experimental setup. We also reviewed the quality achieved by related works segmenting abdominal organs, for comparison purposes. Finally, we showed how post-processing transformed a real test sequence. We conclude that the approach improves the robustness of segmentation by correcting errors, an important advantage being that it can be applied to improve the quality of the outputs of any segmentation network.Precise segmentation of abdominal organs is a relevant task for several clinical procedures, including visual aids to diagnosis, detailed analysis of abdominal organs for correct positioning of a graft prior, abdominal aortic surgery and many other tasks. Ongoing research tries to improve segmentation results and to overcome many challenges due to the highly flexible anatomical properties of the abdomen and limitations of modalities reflected in image characteristics. Previously, segmentation would be done mostly using multi-atlas techniques, with interesting results when applied to different anatomical parts. For instance, Bereciartua et al. [3] segmented the liver using a 3D liver model guided by a precomputed probability map. Le et al. [4] used histogram-based liver segmentation with a subsequent geodesic active contour refinement step, and Huynh et al. [5] used watershed transformation and active contours. In many of these approaches, organ volumes and statistical information regarding those volumes were often used as part of the segmentation process. In the last few years, deep learning-based approaches have been quickly overtaking traditional ones in the task of segmentation of medical images in general and for MRI segmentation of abdominal organs specifically.Deep learning approaches not only improve segmentation scores significantly when compared to more traditional techniques, they are also much more capable of learning and adapting automatically. However, many of the image processing concepts used in prior approaches are still relevant in the deep learning era. In the next paragraphs we review previous deep learning approaches applied to segmentation of abdominal organs in both MRI and CT sequences.Zhou et al. [6] showed that fully convolutional networks (FCN) produce excellent results in segmentation of abdominal organs from computer tomography (CT) scans. Zhou et al. [6] transformed the anatomical structure segmentation on 3D CT volumes into a majority voting of the results of 2D image segmentation on a number of 2D-slices from different image orientations, then an organ localization module was used with two major processing steps: (1) individual organ localization that decides the bounding box of a target organ based on window sliding and pattern matching in Haar-like and LBP feature spaces; (2) group-wise calibration and correction based on general Hough voting of multiple organ locations. Bobo et al. [7] applied the approach to segmentation of the whole abdomen from magnetic resonance imaging sequences (MRI). The authors show that fully convolutional neural networks (FCN) improve abdominal organ segmentation significantly when compared with multi-atlas methods. The FCN they used resulted in a dice similarity coefficient (DSC) of 0.930 for the spleen, 0.730 for the left kidney, 0.780 for the right kidney, 0.913 for the liver and 0.56 for the stomach. Larsson et al. [8] proposed SeepSeg, a method to segment abdominal organs based on three steps: (1) localization of region of interest using a multi-atlas approach; (2) pixelwise binary classification using a convolutional neural network and (3) post-processing by thresholding and removing all positive samples except the largest connected component. In this approach, the authors found a region of interest for each organ, which was done as a pre-processing step for a classification network to classify fewer voxels. The authors achieved IoU (Intersect-over the Union, a.k.a Jaccard coefficient) scores of 0.90, 0.87, 0.76 and 0.84, respectively, for liver, spleen, right and left kidneys. Groza et al. [9] proposed an ensemble of networks and voting for output in the segmentation of MRI scans. There were five different networks used for the final averaged ensemble. The first was the DualTail-Net architecture, which consisted of an encoder part and two independent decoder parts. The other four networks had very similar U-Net-based architectures, and the final result was an averaged ensemble of predictions obtained by the five networks. Conze et al. [10] also worked in the segmentation of MRI scans. They tested several segmentation network architectures, including a deeper version of UNet using VGG19 instead of VGG16, comparing one version training from scratch to one starting from pretrained nonmedical data. The authors also cascaded two networks, combining two v19pUNet networks by inputting posterior probabilities resulting from the first v19pUNet output into the second one. In another alternative, the cascaded pipeline was used as generator within a conditional Generative Adversarial Network (cGAN), a model including a discriminator whose role was to distinguish real ground-truth segmentations from those arising from the generator, to strengthen the ability of the generative part to create segmentation masks that are as realistic as possible. From the various architectures tested by Conze, the most significant improvements came from UNet with VGG19, and also to a smaller degree by the two cascaded UNets. In [11] the authors proposed a convolutional neural network (CNN)-based fully automated MR image-based multi-organ segmentation technique, namely ALAMO (Automated deep Learning-based Abdominal Multi-Organ segmentation). A multi-slice 2D neural network was developed to account for information between adjacent slices. Within the study, the authors investigated multiple approaches, including network normalization, data augmentation, and deeply supervised learning. They also introduced a novel multiview training and inference technique. As part of the work the authors compared two popular networks, PlainUnet and DenseUnet, showing that DenseUnet used fewer parameters and offered more accurate segmentation results. By adding multiple skip connections within the convolutional blocks, the network was forced to reuse its weights, thus dramatically reducing the number of parameters for the same performance. The authors also showed that normalization had insignificant performance gain and combining three different views could boost performance further.These approaches reviewed focused mostly on architectural variations, multiple views, ensembles and voting to try to improve the quality of segmentation. Not as much focus has been placed on the central idea we explore in this work, which is that relevant improvement can also be obtained by applying post-processing operations that enforce semantic invariants. The use of post-processing operations to improve the quality of segmentation also has the advantage of robustness, since whatever network architecture is chosen, the quality of the resulting images and segmentation depends on acquisition details, morphology and other factors. Details such as low contrast between the tissue near the borders of organs and surrounding tissue, calibration specifics, morphology variations along slices and between different patients, and multiple other factors, can result in better or worse segmentation quality, even when top performing architectures are used. In that context, and to improve the chances of a best-possible result regardless of the network architecture, a set of post-processing steps that apply relevant semantic concepts is a very useful add-on to make the results robust to variability. In this paper we define, apply and experiment with post-processing operations that can correct and improve the final quality of segmentation of the abdominal organs, thus complementing previous works on segmentation of MRI sequences of abdominal organs.In this section we define a sequence of post-processing operations to be applied automatically to the output of segmentation of an MRI sequence scanning the abdominal organs. The segmentation convolution neural network outputs a segmentation which is then passed through the following set of operations: (1) organs are first assigned relative 3D spatial location restrictions (fencing), based on the locations in training data; (2) a re-classification is done within organs envelops to correct classification mistakes and remove wrong classifications; (3) in-organ holes filling and 3D organ continuity filling are applied (enveloping), and finally surface smoothness operations are used to smooth the surface.The ground-truth MRI sequences are an essential element in the post-processing operations, since the expected volumes of the organs (an atlas or each organ) are obtained automatically from those sequences. Each slice is an image-sized labelmap (ground-truth labelmap), i.e., each pixel contains the pixel class label. The labels identify the structure or organ that the pixel belongs to, labeling 0 for background (no class). The output of segmentation is also a labelmap, which we denote as “segmentation labelmap”. The purpose of post-processing is to transform the segmentation labelmaps using invariants inferred from the ground-truth labelmaps. We also define the term organmap as being a labelmap containing only one of the organs, which is obtained from a labelmap by zeroing all labels except the one identifying the specific organ to keep (the organmap can also be transformed into a binary map by assigning 1 to the organ label).Besides the labelmaps, we generate a (contiguous) regionsmap for the 3D volume. While labelmaps label pixels based on the class they belong to, the regionsmap labels pixels based on the labels based on pixels connectivity in the image. This allows us to distinguish different connected regions of each organ in the segmentation output and enables further automatic reasoning regarding those regions.The operations we describe next are, in order: “fencing”, “class re-assignments and removal of noise”, “computation and filling of organs envelops” and “slice smoothing and filling”.Conceptually, the fence is a 3D volume that defines maximum organ size and/or positions for each organ. The fence is defined for each organ separately. Given the set of training sequences, the fence of one organ is the union of all 3D volumes of that organ in the training sequences. A dilation operation follows [12] (imdilate function in matlab) to add a volume tolerance (δ) over the union of volumes (δ-dilation over the union of organ volumes V; we used δ = 5 pixels).In order to compute the fence given the MRI training set, the ground-truth sequences are first aligned in the scan sequence dimension using image registration algorithm [13]. The second step involves, for each organ O present in the ground-truth sequences, isolation of O in all sequences and computation of the union of all volumes of O in all sequences, resulting in a volume V that contains all volumes of the instances of that organ in all training sequences.Algorithm 1 shows the steps involved in pseudo-code. In that code, the loop of step 2 cycles over each organ. In that loop, for each organ, we create an initial empty 3D volume with the size of the largest sequence (step 2a), then for each sequence (step 2.b) we OR its 3D volume with the current organ volume (sout{end}) (step 2.b.i). After processing all training sequences we simply apply an imdilate function [12] (step 2.d).Considering the abdominal organs liver, spleen and kidneys, Figure 2 illustrates three different views of the union of organ volumes obtained from the training ground-truths. Blue stands for the liver, yellow and purple stand for kidneys and grey stands for the spleen. The red volumes show intersections of two or more organs, i.e., a spatial volume where more than one organ can appear in the union of all training sequences.After fencing, class reassignments and noise removal are done using a set of steps.Define the largest continuous spatial region of a specific organ, O, as the main volume of that organ.For each region that is classified as another organ, O′, but is completely within the organ fence and which has a volume larger that a predefined threshold (the threshold was set to 500 pixels in our experiments), consider it as organ O (reclassify).For each region that is classified as another organ O’ but is completely within the organ fence and which has a volume smaller than the predefined threshold (the threshold was set to 500 pixels in our experiments), consider it as background (reclassify to background).Regions classified as organ O, but which are smaller than a certain threshold (i.e., “too small regions”) are considered noise and reclassified as background (the threshold was set to 500 pixels in our experiments).Regarding implementation details, Algorithm 2 shows the main steps involved. The inputs to the algorithm are the segmentation outputs, denoted as 3D arrays in a set of sequences s{[]} and the fences. In that algorithm step 2 processes each sequence separately, and step 2.b iteratively obtains the volumes of pixel classifications as each organ, so that the algorithm can process each organ separately from the segmentation output. Next the fence is applied to zero all pixels outside the fence (step 2.b.ii), keeping only organ pixels that are inside of it. The next step (steps 2.b.iii and 2.b.iv) identifies the largest region as being the specific organ being processed in the current iteration. To do that, an organ regions labelmap (bw) is created from the organ volume using bwlabel [14] (step 2.b.iii). The sizes of regions are next calculated on those bw regionmaps (number of occurrences of each label). The region label with top number of occurrences identifies the largest region, and therefore the organ extent (step 2.b.iv). Finally, step 2.b.v reclassifies regions classified as other organs inside the organ fence and having sizes larger than a threshold (500 pixels) as being part of the organ itself. Since those regions are inside the fence of the organ and are not small, they are wrongly classified as another organ and should, therefore, be reclassified. The remaining regions with volume lower than 500 are reclassified as background (step 2.b.vi). Step 2.b.vii removes regions that were classified as the organ but are actually disconnected from its main volume (have a different label in the bw labelmap) and have sizes lower than a threshold (500 pixels).Given the imperfections of segmentation near borders (including spurious pixel classifications as part of the region and sometimes thin connections to neighbour regions), we also found out that the best results were obtained by preceding the calculation of the largest region by morphological erosion [15] then calculating and isolating the largest region, subsequently applying dilation [15] with the same structuring element and size to reverse the previous erosion operation. The erosion frequently eliminates noise in the borders and some spurious connections to neighbouring regions.After filling each slice individually (imfill [12]), the next step joins disconnected regions that are inside the organ’s fence and are classified as the same organ. In this context, a discontinuity is a gap between two regions of an organ in the sequence of MRI slices. The space between the disconnected regions is filled by interpolation between the border pixels of the two extremity slices bordering the gap. Figure 3 illustrates the objective. In the figure, the space between the two regions r1 and r2 is filled by interpolation.The result after all the previous operations is a 3D envelope of each organ based on the largest volume classified as that organ, and the remaining major volumes reclassified as that organ within the fencing volume of the organ, with the space between the parts filled to create a solid.Slices smoothing is the process of improving each slice by removing small protruding pixels, filling holes and smoothing edges of each organ independently. The steps are shown in Algorithm 3. For this algorithm, slices are processed in sequence. For each organ in each slice, the first part of the algorithm (steps 1.a.i to 1.a.iii) removes small protruding pixels from the main volume. This is done using a 2D erode operator (imerode [12]) to isolate the main volume from spurious pixels, keeping only the largest region, and then applying a dilate operator (imdilate [12]) to restore the organ area to the original size. Step 1.a.iv fills holes inside the organ region using the imfill morphological operator ([12,16]). The final step 1.a.v smooths contours. This works on the binary image of the organ in the slice. The step first blurs the contours using a 3 × 3 2-D average pixel value convolution. From the output of the blurring, all pixels with value smaller than 0.5 are zeroed. The resulting nonzero pixels are the extent of the smoothed organ.Figure 4 is an illustration of transformation based on the above-described operations (fencing-enveloping-noise removal-reassignment-smoothing). It shows what happened when a specific test image was segmented using the DeepLabV3, followed by the post-processing operations that corrected the output. In the figure, the liver is orange, the kidneys are yellow and purple, and the spleen is green. Figure 4a shows the ground-truth abdominal organs. Figure 4b shows the segmentation output, where some errors can be seen (e.g., part of the spleen was classified as right kidney, another part as left kidney and only a smaller region is correctly classified as spleen; the right kidney region is infiltrated by a region with a few contiguous slices classified wrongly as liver; there are also some small noisy regions, including some larger noise outside fences, e.g., on top of the spleen). Figure 4c shows the final corrected result after applying all the post-processing transformations. Fencing removed some of the imperfections, such as the incorrect regions classifications as right kidney in the left part of the figure. Enveloping and reclassification transformed those and other incorrect regions standing inside the fence of spleen into spleen and merged the resulting parts. It also transformed the region inside the right kidney that was incorrectly classified as liver into right kidney, and finally smoothing smoothed the slices to obtain the final result shown in Figure 4c. Although still not exactly equal to the original model, the post-processed Figure 4c is much closer to the original Figure 4a than the segmented Figure 4b.In this section we first describe the architecture of the segmentation networks used in our experiments, then we describe the dataset and details of the experimental setup.The architecture of the segmentation network is a relevant factor for the quality of segmentation; therefore, a lot of prior research has focused on improving architectures. In this work we follow the strategy of choosing a best-performing network from a set of popular architectures widely used in medical imaging in general. The U-Net [17], DeepLabv3 [1] and FCN [18] are our choices of networks, and our focus is on the following sequence: (1) first compare those base networks with the dataset used for the experiments, including data augmentation and other training optimizations to pick the best-performing architecture for the experimental setup used; (2) try to maximize the quality by testing training options, in particular this led us to experiment with a successful modification of the loss function; (3) experiment with the post-processing functionality using that network. Next, we describe the base network architectures used.U-Net. U-Net is a 58-layer segmentation network using VGG-16 stages for feature extraction (encoding), followed by an intermediate section connecting encoder to decoder, and a decoding section that is symmetric to the encoding section. Figure 5a summarizes the layers of U-Net. The encoding section consists of contraction blocks applying two 3 × 3 convolution layers followed by a 2 × 2 max pooling layer. The decoding section is symmetric to the encoding section, consisting of the same number of expansion blocks as there are contraction blocks in the encoding section. As with encoding, each expansion block has two 3 × 3 convolution layers followed by a 2 × 2 up-sampling layer. But each expansion block also appends feature maps from the corresponding contraction block. The rationale is that features learnt in the contracting block of the image will be used to reconstruct it at the symmetric stage.FCN. The structure of FCN is sketched in Figure 5b. FCN also uses VGG-16 (with seven stages, corresponding to 41 layers) as encoder, plus a much smaller sequence of up-sampling layers (decoding stages) for a total network size of 51 layers. FCN also forwards feature maps (the pooled output of coding stage 4 is fused with output of the first up-sampling layer, and the pooled output of coding stage 3 is fused with the output of the second up-sampling layer). Finally, the image input is also fused with the output of the third up-sampling layer, all this followed by the final pixel classification layer.DeepLabV3. DeepLabV3 is the deepest network tested in this work, with 100 layers and a generic layout of layers shown in Figure 5c. DeepLabV3 uses Resnet-18 as feature extractor, with eight stages totaling 71 layers, the remaining stages being Atrous Spatial Pyramid Pooling (ASPP), plus the final stages. Forwarding connections are also added from encoding stages to the ASPP layers for enhanced segmentation of objects at multiple scales. The outputs of the final DCNN layer are combined with a fully connected Conditional Random Field (CRF) for improved localization of object boundaries using mechanisms from probabilistic graphical models.The magnetic resonance imaging data used in our experimentation is CHAOS Dataset, a publicly available set of scans in [2]. Table 1 lists the main dataset configurations. The data consists of 120 MRI sequences capturing abdominal organs (liver, kidneys and spleen) obtained using the T1-DUAL fat suppression protocol. Table 1 summarizes the dataset configurations. The sequences were acquired by a 1.5T Philips MRI, which produces 12-bit DICOM images with a resolution of 256 × 256. The ISDs varies between 5.5–9 mm (average 7.84 mm), x-y spacing is between 1.36–1.89 mm (average 1.61 mm) and the number of slices is between 26 and 50 (average 36). In total there are 1594 slices (532 slice per sequence) used for training and testing, with the testing sequences being chosen randomly to include 20% of all sequences in 5-fold cross-validation runs. Given the relatively limited size of the dataset, data augmentation was added after we verified that it would contribute to improved scores by increasing diversity and the size of the dataset.Data augmentation was defined based on random translations of up to 10 pixels, random rotations up to 10 degrees, shearing up to 10 pixels and scaling up to 10%. The networks were pretrained on object recognition tasks (Imagenet dataset). Network training was configured using SGDM as the learning algorithm, with an initial learning rate = 0.005 and piecewise learning rate with a drop period of 20 and a learn rate drop factor of 0.9 (the learn rate would decrease to 90% every 20 epochs). The default loss function used was cross-entropy (crossE), but we also experimented with post-processing on the segmentation output of the network trained with IoU loss. We include IoU loss because we found out that it improved the quality of segmentation of individual organs. Class balancing was applied in the pixel classification layer, training iterations were 500 epochs after we verified that convergence to a stable loss is achieved before that with minibatch size = 32, and momentum = 0.9. The training and testing were done on a machine with a GPU NVIDEA G Force GTX1070. The experiments were divided into two phases. The first phase involved choosing the best performing segmentation network. Using the chosen network we then proceeded to segment all MRI test sequences and then applied automatic post-processing to all sequences. The last step involved evaluating the quality of the results. We focused our analysis using global metrics and per-class IoU (a.k.a. Jaccard index JI), since this is a reliable and common metric for evaluation of segmentation performance.The coding for this work was done on Matlab R2019b and required the image processing toolbox, the deep learning toolbox, the computer vision toolbox and the deep learning toolbox model for the Resnet-18 network. The DeepLabv3 segmentation network was built using the deeplabv3plusLayers function. The FCN architecture was built using the fcnLayers function and the UNet used function unetLayers. The networks were trained using the trainNetworks function with training options defined using the trainingOptions object. The MRI images were handled by imageDatastore objects, and the ground-truths were handled by pixelLabelDataStore objects. The code for post-processing was all written from scratch in Matlab using multidimensional arrays to keep the training and test data and Matlab library functions to implement the necessary image processing operations. Matlab functions used as helpers to implement the steps included bwlabel, imbinarize, bwareafilt, morphological operations such as imfill, imerode, imdilate, and convolution operations using conv2. Links for code Supplementary Materials are given in the end of the work.In this section we evaluate the proposed post-processing approach. Our experiments were based on first choosing the best-performing network, then presenting the post-processing results in the form of evaluation of the amount of improvement derived from applying it. Finally, we review performance results from related works on both MRI and CT of abdominal organs to put our results in perspective in relation to the performances reported in those works.Table 2 shows the results of the first part of our experiments, finding the best-performing base network. As described in the experimental setup, these results were obtained with data augmentation. IoU (JI) and Dice of the three segmentation network architectures is reported. The best-performing network was deepLabV3 (2 percentage points (pp) better than FCN, FCN being 3 pp better than UNET). For that reason, we proceeded with the next experiments using deepLabV3.For the next experiment we took the segmented sequences after applying deeplabV3-based segmentation to the original test sequences and ran them through the sequence of post-processing operations. The experiment was divided into two main tests to reflect the use of two different loss functions, i.e., cross entropy (crossE, the default loss function) and IoU. We included the two loss functions because IoU loss improved the quality of segmentation further. That way, we were able to compare the amount of improvement with the base crossE loss to that with already improved segmentation scores using IoU.Table 3, Table 4, Table 5 and Table 6 show our results concerning the effects measured as IoU and Dice of applying the post-processing improvements in sequence. Table 3 and Table 4 report the results using IoU as the loss function, and Table 5 and Table 6 shows the results using crossE, the default loss function. Each column in the tables identifies the organ and the last column is the sum of percentage points (pp) increase in IoU or Dice over all organs. The first row shows the IoU/Dice of the base segmentation results, then each other row shows the IoU/Dice achieved after each post-processing operation.The results in the tables essentially show that the post-processing steps were quite useful. Fencing improved the quality of segmentation by 2 pp in Table 3 (IoU loss) and by 8 pp in Table 5 (crossE loss). The next operation, re-classification and noise reduction, had the most significant contribution in both cases (6 pp and 9 pp for IoU and crossE loss respectively), and the final operation (enveloping, filling and slice filling and smoothing) contributed with 5 and 8 pp increases, respectively. The contributions were larger in the experiment with crossE because the base segmentation had more errors in that case. In general, these results prove that post-processing improved the results significantly. Importantly, we applied post-processing in a specific experiment setting and with a specific dataset and network, but it can be generalized to any approach and dataset since the defined post-processing steps are not dependent on either the network, the experimental settings or datasets tested.In this section we first compare the final scores obtained after post-processing in our experimental work with some techniques based on the architectural features tested on the same dataset. Then, we review some of the best scores obtained for segmentation of abdominal organs by other authors using a variety of techniques that include improved network architectures and ensembles, many of those works segmented CT scans, others segmented MRI scans. We converted scores to IoU (Jaccard index) when necessary, since many of those works report scores using the dice metric.Table 7 compares several architectural variations from the work [10] that include the U-Net, a modified U-Net with VGG-19 instead of VGG-16 (V19UNet), a pretrained version (V19pUNet) and finally a cascade of two V19pUNet (V19pUnet1-1). The results show that our post-processing approach is better than any of these.Table 8 shows the quality of segmentation achieved by related MRI (first three cases) and CT segmentation approaches as reported by the authors in their own papers. As can be seen from the table, there are many related approaches, and the reported scores vary significantly and also depend on the dataset used. From Table 8, refs. [19,20] achieve high scores in MRI segmentation of some of the organs (only the liver in [20]). Hu et al. [21,22] obtained the best results for CT, usually above 90%. The scores we obtained after post-processing (Table 3 and Table 5) are higher than [7] and comparable to some of the best scores reported in related works segmenting MRI sequences (e.g., [19]). Most importantly, the scores on those advanced architectures would also benefit from applying our post-processing approach, since it can be applied to any segmentation technique.Figure 6 shows a 3D depiction of the sequence of slices of abdominal organs from a real test sequence. The 3D depiction is shown as a 3D model that includes all four organs (liver, spleen, right and left kidney), and also as a 3D coloured regions model (spleen is green, liver is orange, right kidney is yellow and left kidney is purple).Figure 7 shows the output of segmentation using DeepLabV3 with the default cross entropy loss function. Once again, we show the 3D model plus the 3D coloured regions. We can see from both models, and especially from the 3D coloured model, that there are several inaccuracies in the output, including regions classified incorrectly as another organ.Figure 8 shows the 3D coloured model of the results after post-processing. We can see that incorrect region classifications were corrected, especially in the spleen and left kidney regions, and in the right kidney region as well. Some noise was also removed.Deep learning-based segmentation is an established procedure for segmentation of MRI and CT sequences. In spite of the amazing quality of the results, there are still imperfections, and researchers search for approaches to improve the quality of the results. Many works have explored advances in segmentation network architectures and the use of ensembles and voting. We propose and evaluate a complementary approach of image post-processing for enforcement of semantic invariants (fencing, enveloping, noise removal, reassignment and smoothing) to improve the results. The approach is defined in detail, and in the experimental section we tested the degree of improvement achieved from the post-processing steps. We used a publicly available dataset to show that the approach improved segmentation scores by a sum of 12 to 25 percentage points over the four organs tested.Our focus in this work has been on improving the quality of segmentation output by means of post-processing, which can be applied regardless of the architecture of the segmentation CNN. Our future work on this issue will explore two major improvements. On one hand, we intend to find an approach to integrate post-processing into the network architecture itself as additional layers. The advantage will be to integrate this post-processing in the back-propagation learning procedure. On the other hand, we also wish to investigate several relevant evolutions to current state-of-the-art segmentation using deep learning. Concerning the use of U-Net for segmentation of MRI, those relevant innovations include Attention Gates (AGs) [26], Squeeze-and-Excitation (SE) blocks [27] and Squeeze-and-excitation networks [28], which improve generalization performance in multicentric studies. Our future work also involves testing the approach with CT scans and together with advanced techniques that include ensembles of networks with voting.The following are available online: https://github.com/pedronunofurtado/postprocess, base code functions used in this work for testing postprocessing; https://github.com/pedronunofurtado/codingLOSS, code used for testing networks with loss functions.This research received no external funding.Not applicable.Not applicable.The dataset used is public and all relevant details are provided. Additional information is provided on demand.We used a publicly available MRI dataset for our experiments. The dataset can be found in [2,29,30]. We acknowledge the organizers for allowing researchers to use these data.The authors declare no conflict of interest.Example MRI sequence segmentation. The ground-truth in the left shows the organ extents of the liver, kidney and spleen that are exposed in this specific slide (blue means organ). On the right we show the corresponding pixel classifications by the segmentation network, where the colours are coded as: light blue = kidney, gold = spleen, red = liver. It is possible to see that the liver segment spills well off the organ, a region above the spleen is classified as spleen and, finally, the borders of all three organs are overcome by the segments.Illustration of semitransparent dilated union of organ volumes. The figure shows the union of locations of organs in 3D, taken from the training dataset in our experiments. The union is shown from three perspectives, which are front, bottom and top. In the figure, the liver in blue, the two kidneys are gold and purple, respectively, and the spleen is grey. Regions shared by more than one organ are red. The fence is obtained using a dilation (imdilate operator [12]) of each individual organ.Illustration of space filling. Given a main volume of an organ (region r1) and another smaller region (region r2) classified as the same organ inside its fence, the algorithm interpolates between each pixel of the extremity of r2 and r1. (a) Two regions, same organ. (b) Illustration of space filling operation. (c) Resulting organ.Illustrating before and after transformations. (a) Original. (b) Segmented. (c) Post-processed.Architectures of segmentation networks used. (a) U-Net segmentation network. (b) FCN segmentation network. (c) DeepLabV3 segmentation network.Original test sequence with organs as 3D and coloured 3D models.Segmented test sequence with organs as 3D and coloured 3D models.Post-processed segmented organs.Summary of dataset configuration.IoU of segmentation networks with base crossE loss.Improvement (metric = IoU) when IoU used as loss function.Improvement (metric = Dice) when IoU used as loss function.Improvement (metric = IoU) using cross entropy loss function.Improvement (metric = Dice) using cross entropy loss function.Comparing to IoU of related approaches (CHAOS dataset).IoU as reported in some related approaches (MRI and CT).Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00008.txt

@@ -0,0 +1,7 @@
+Sleep deficiency impacts the quality of life and may have serious health consequences in the long run. Questionnaire-based subjective assessment of sleep deficiency has many limitations. On the other hand, objective assessment of sleep deficiency is challenging. In this study, we propose a polysomnography-based mathematical model for computing baseline sleep deficiency severity score and then investigated the estimation of sleep deficiency severity using features available only from wearable sensor data including heart rate variability and single-channel electroencephalography for a dataset of 500 subjects. We used Monte-Carlo feature selection (MCFS) and inter-dependency discovery for selecting the best features and removing multi-collinearity. For developing the Regression model we investigated both the frequentist and the Bayesian approaches. An artificial neural network achieved the best performance of RMSE = 5.47 and an R-squared value of 0.67 for sleep deficiency severity estimation. The developed method is comparable to conventional methods of Functional Outcome of Sleep Questionnaire and Epworth Sleepiness Scale for assessing the impact of sleep apnea on sleep deficiency. Moreover, the results pave the way for reliable and interpretable sleep deficiency severity estimation using single-channel EEG.Sleep is an important biological process and plays a key role in restoring energy, solidifying and consolidating memories, and repairing body cells. It is controlled by the circadian biological clock and sleep/wake homeostasis and also helps regulate metabolism and cardiovascular function [1]. With the rise of obesity, excessive usage of personal gadgets, rapid urbanization, and other socio-economic changes, sleep/wake homeostasis is increasingly impacted, disrupting the normal circadian rhythm and healthy sleep. Good quality sleep is essential for optimal health and improved quality of life. Neither body nor the brain can function properly without sufficient sleep. Research suggests that complete sleep deprivation significantly impairs attention and working memory [2]. Moreover, it also affects other important functions, such as long-term memory and decision-making. Even partial sleep deprivation can negatively impact attention, and vigilance in the long run [3]. Moreover, sleep deficiency can lead to physical and mental health problems, injuries, loss of productivity, and even greater risk of life-threatening diseases [4,5].The common practice to evaluate sleep health is to use standard questionnaires along with a sleep test where a polysomnogram is captured. The questionnaire-based approach has many limitations including high bias, long evaluation period, etc. Polysomnogram is expensive, has limited availability, and less user-friendly. In recent years, there has been a significant expansion in the development and use of multi-modal sensors and technologies to monitor sleep which includes sleep patterns monitoring, wellness applications, sleep coaching of individuals with chronic conditions, etc. [6]. Kuo et al. developed an actigraphy-based wearable device for sleep quality assessment [7]. Mendonca et al. proposed a method for sleep quality estimation using electrocardiogram by cardiopulmonary coupling analysis [8]. Azimi et al. reported an objective IoT-based longitudinal study for sleep quality assessment [9]. They estimated the sleep quality average to classify sleep into good or poor quality. Bsoul et al. developed a Sleep Efficiency Index based on ECG features using support vector machines [10]. Some research group explored deep learning based approaches for sleep assessment/scoring [11,12,13,14]. During sleep scoring information from the electrophysiological signal are extracted to detect sleep stages, arousals, respiratory events, movements, etc. Additionally, some commercial devices e.g., the Fitbit Charge smart band (Fitbit Inc., San Francisco, CA, USA), the Apple Watch (Apple Inc., Cupertino, CA, USA), the Oura sleep ring (Oura Health Ltd., Oulu, Finland) have attempted the estimation of sleep scores from non-polysomnographic measures. However, most of the previous approaches were focused on sleep quality assessment or sleep score estimation. Sleep deficiency provides a more specific evaluation of sleep disorder than sleep score or the sleep quality, as sleep deficiency includes lack of enough sleep (sleep deprivation), not getting all types of sleep that the human body needs, and poor quality of sleep [15]. Early detection of sleep deficiency is beneficial to avoid many linked chronic health problems, including heart disease, kidney disease, high blood pressure, diabetes, stroke, obesity, and depression. A method for objective assessment of sleep deficiency from wearable sensor data only is not well-established yet requiring further investigation.The main objective of this work was to develop a physiological sensor data-based objective sleep deficiency assessment method that can be integrated with user-friendly wearables e.g., smartwatch, smart band, etc. We proposed a mathematical model to facilitate a quantitative evaluation of sleep deficiency based on polysomnogram features. Then we addressed the same problem of estimating sleep deficiency severity when polysomnogram data is not available, with a machine learning-driven model using ECG/EEG based features that can be captured by wearables. The contribution of this work is summarised below:(i)Reported a mathematical model for quantifying Sleep Deficiency Severity (SDS).(ii)Reported a machine-learning driven model for estimation of SDS from wearable EEG based spectral features.(iii)Identified robust biomarkers with the help of Monte Carlo feature selection and Inter-dependency discovery for SDS monitoring in presence of confounding factors.(iv)Assessed the impact of obstructive sleep apnea on SDS.Reported a mathematical model for quantifying Sleep Deficiency Severity (SDS).Reported a machine-learning driven model for estimation of SDS from wearable EEG based spectral features.Identified robust biomarkers with the help of Monte Carlo feature selection and Inter-dependency discovery for SDS monitoring in presence of confounding factors.Assessed the impact of obstructive sleep apnea on SDS.The results pave the way for automated sleep deficiency severity assessment using single-channel EEG.We define “Smart Health (sHealth)” as a system that uses embedded artificial intelligence such as edge computing, machine learning, etc. and aims to deliver improved healthcare using users’ smart devices, wearables, and the Internet of Things (IoT) centric solutions. It not only benefits and monitors individual user’s health but also collects spatiotemporal community-wide data for collective and social well-being and informed policy makings. It is an emerging paradigm for efficient processing, sharing, and visualization of healthcare data, which is coming from different IoT devices, and wearable sensors. sHealth can be perceived as an upgraded and extended version of Mobile Health (mHealth). We previously reported a framework for sHealth and conducted a pilot study to evaluate the technical feasibility of the framework [16]. The system architecture for the sHealth framework is shown in Figure 1. The main components in the framework are various sensors, such as battery-less body-worn passive sensors with a scanner (i.e., reader for the passive sensors), commercial wearables, a custom smartphone app (SCC-Health app), and a custom web server (SCC-Health server). Details of the design and functionality of the sensors and scanner can be found elsewhere [17]. For physiological data collection, we utilized novel inkjet-printed (IJP) sensors in addition to commercial wearables such as a smart wristband (Mi Band 2, Xiaomi, Beijing, China) and a fingertip pulse oximeter (CMS50E, Crucial Medical Systems, Atlanta, GA, USA) [18]. The IJP sensors were zero-power, analog, wireless, and fully passive. Data collected in the IJP sensors were pre-processed and digitized by a custom-made scanner. Data from the scanner is transmitted to the smartphone via Bluetooth. Data reliability check, feature extraction, and classification or regression analysis using pre-trained machine learning models are performed in the smartphone for disease detection and severity assessment. The computed severity of the disease is then visualized in the smartphone as well as shared with the webserver using Wi-Fi (or cellular network) for observing the temporal and spatial distribution of the diseases. The webserver is accessible at http://sccmobilehealth.com (accessed on 30 June 2021), please see the supplementay materials for details. The app was developed using Android studio 3.1 with build tool version 25.0.2 and the minimum SDK level of 19. The pre-trained machine learning models integrated with the app for Events of Interest (EoI) computation were developed and evaluated in WEKA. By EoI we mean a change in disease-related biomarkers i.e., heart rate, oxygen saturation level, spectral power, etc. that help in detecting a disease or indicate a progression/exacerbation in the disease. Additionally, electronic forms were used to collect user-reported symptoms and user activity logs [19].The process of spatiotemporal visualization is fully automated and near real-time [20]. For personalized monitoring of diseases, temporal trends of disease severity for a participant can be visualized using a time plot graph [20]. Flow graphs have been used for community health trend monitoring over time. In addition to that, a spatial plot was used to visualize the severity of the disease in different areas at a period interval (averaged) [20]. Color coding has been used to indicate severity where red indicates the highest severity and green indicates the lowest severity.As shown in Figure 2 conventional methods of sleep health assessment fall under two broad categories-subjective assessment and objective assessment of sleep. Subjective assessment of sleep deficiency using standard questionnaires is well investigated and is widely used in clinical practice. Some of the well-accepted and popular methods for subjective sleep quality assessment are- the Pittsburgh Sleep Quality Index (PSQI), the Epworth Sleepiness Scale (ESS), and the Functional Outcome of Sleep Questionnaire (FOSQ).PSQI uses a seven component questionnaire and the subject assigns a score of 0–3 for each component [21]. The components are-subjective sleep quality, sleep latency, sleep duration, habitual sleep efficiency, sleep disturbances, use of sleep medication, and daytime dysfunction. A global score > 5 indicates poor sleep quality. FOSQ has 21 questions related to activity levels, vigilance, intimacy and relationships, general productivity, and social outcomes [22]. The potential range of scores for each subscale is 1–4 with higher scores indicating greater insomnia severity. Similarly, in ESS the subject assigns a score of 0–3 for 8 questions aimed at assessing daytime sleepiness. A total score of 16–24 indicates excessive daytime sleepiness suggesting the need for medical attention [23]. The Karolinska Sleep Diary (KSD) is another questionnaire that was developed to assess subjective sleep quality [24].Subjective reports of sleep quality are important in the clinical setting and can help determine whether further screening and/or treatment for a sleep complaint might be warranted [25]. However, subjective methods suffer from high bias, require active user participation, and a longer period (2 weeks–1 month) for sleep deficiency evaluation. Objective sleep quality consists not only of the total duration of sleep, but also of the architecture of sleep (amount of the different sleep stages across the sleep episode), the amount of wake time during the sleep episode, and the frequency and duration of awakenings across the night [26]. Prominent quantitative metrics that are used for objective sleep assessment are- the Sleep Quality/Efficiency Index and the Sleep Score.Several quantitative metrics have been developed to measure the quality of sleep from physiological sensor data. However, there is still a lack of standard and well-established definitions for the term ‘Sleep Quality’ [27]. Usually, it is used to refer to a score computed from the collection of quantitative sleep measures i.e., sleep duration, sleep onset time, degree of fragmentation, etc. [28].The concept of sleep score has been introduced mainly by commercial entities i.e., Fitbit, Polar, Oura, Apple, etc. Sleep score is usually tracked by smartphone apps based on data collected using a smart band or a smartwatch during sleep. There is no well-accepted and standard definition for Sleep score. Fitbit computes overall sleep score as a sum of individual scores in sleep duration, sleep quality, and restoration, for a total score of up to 100. Most people have a sleep score between 72 and 83. Sleep score ranges are: Excellent: 90–100, Good: 80–89, Fair: 60–79, Poor: Less than 60 [29]. A previous validation study showed that Fitbit smart bands performance is promising in detecting sleep-wake states and sleep stage composition relative to gold standard polysomnogram [30]. Oura sleep ring measures sleep using sensors that capture the body signals including resting heart rate (RHR), heart rate variability (HRV), body temperature, respiratory rate, and movement, to determine sleep patterns and compute the sleep score [31]. Sleep score-based monitoring has been criticized due to a lack of consistency in measurement and the impact of a sleep-related disease on sleep score is not well-investigated [32]. Scientific study suggests that current consumer sleep tracking technologies are immature for diagnosing sleep disorders, but they may be reasonably satisfactory for general purpose and non-clinical use [33].This is a new metric proposed by the authors aimed at pre-clinical early evaluation of sleep deficiency and is based on a fusion of features from ECG, EEG, SpO2, and other wearable sensors. Sleep deficiency is more than sleep quality or sleep efficiency as it includes sleep deprivation, sleep fragmentation, sleep duration etc. in addition to sleep quality which mainly measures the time spent in deep sleep stages. According to US National Institute for Health [34], sleep deprivation is a condition which happens when a person does not get enough sleep whereas sleep deficiency is a broader concept that occurs for the following reasons-A person does not get enough sleep (sleep deprivation)A person’s sleep is out of sync of his body’s natural clockA person does not get all the different types of sleep that his body needsThe person has a sleep disorder that prevents him from getting enough sleepThe person gets a poor quality sleepThe currently used sleep assessment metrics are not capturing all these components/causes of sleep deficiency e.g., sleep quality equation doesnot take SpO2 level into consideration.Sleep Health Heart Study (SHHS) is a dataset available from the National Sleep Research Resource [35]. SHHS was implemented as a multi-center cohort study in two phases by the US National Heart Lung & Blood Institute. Unattended home polysomnograms were obtained for both the phases of SHHS by certified and trained technicians. The polysomnogram data was saved in European Data Format (EDF). Data processing and initial scoring were done using Compumedics software (Compumedics Ltd., Abbotsford, Australia) as part of SHHS. Two manual scorings were included to annotate the database with sleep duration, sleep efficiency, arousal index, sleep stages, oxygen saturation level, etc. A dataset of 500 subjects containing good quality data for both ECG and EEG is available from the dataset provider and is recommended for use in a research study. In our study, for developing the regression models we used this dataset of 500 subjects. The gender distribution of records in the dataset is as follows: male: 231, female: 269. The age of the subjects ranges from 44 to 89 years old with a mean of 65 years old and a standard deviation of 10.41 years. The body mass index (BMI) of the subjects ranges from 18–46 with a mean of 27.51 kg per square meter and a standard deviation of 4.11 kg per square meter.Guidelines for computing a composite sleep health score from polysomnographic measures have been developed and reported in previous research studies [36,37]. In this study, we developd a generalized mathematical model for computing the baseline SDS score. The model is described in Equation (1) below, where Zneg is the sleep attribute (normalized) that increases sleep deficiency (i.e., higher is responsible for more sleep deficiency). The MAX function in the equation forces the lower bound of Sleep deficiency Severity to 0:(1)Sleep Deficiency Severity=MAX0,S × 100
+        where:S=1m ∑i=1mZnegi−1n∑j=1nZposj 
+        where:Z=X−minXmaxX−min X  Zpos is the sleep attribute (normalized) that reduces sleep deficiency (i.e., higher is better), m is the total number of negative attributes, and n is the total number of positive attributes. The positive attributes available from the SHHS dataset are as follows:Sleep time—Duration of entire sleep.Sleep efficiency—Percentage of time in bed that was spent sleeping, or the ratio of total sleep time to total time in bed, expressed as a percentage.Time deep sleep (%)—Percent time in sleep stages 3 and 4.Time REM sleep (%)—Percent Time in rapid eye movement sleep (REM).SpO2 (%)—Average oxygen saturation (SpO2) level in sleep.The negative attributes available from SHHS is as follows:Sleep Fragmentation Index (SFI)-Total number of arousals per hour of sleep i.e., ratio of the count of arousals to total sleep time in hours.In computing the sleep deficiency severity, all the attributes have been normalized on a scale of 0–1. To achieve a consistent “higher is better” rule the value of each negative attribute is subtracted from 1. Then, the attribute values have been summed up to develop a composite score. The composite score has been multiplied by 100 and divided by the total no. of positive and negative attributes to obtain the SDSin the range of 0–100. Statistical analysis has been conducted to investigate the relationship of baseline SDS with age, gender, and BMI. Additionally, the partial correlation of SDS (controlling for age and BMI) with HRV and EEG features has been investigated.The recording montage for polysomnograms consisted of data from 14 channels which include- ECG, EEG, electrooculogram (EOG), electromyogram (EMG), nasal airflow, thoracic and abdominal movement signal, SpO2, sleep hypnogram, etc. Hardware filters have been used for preliminary noise reduction. The cutoff frequency for hardware filters had been as follows: ECG 0.15 Hz, EOG 0.15 Hz, EMG 0.15 Hz, EEG 0.15 Hz, Thoracic respiration signal of 0.05 Hz, Abdominal respiration signal 0.05 Hz. The sampling rate is 125 Hz for EEG, ECG, and EMG signals. For EOG, the sampling rate is 50 Hz. In investigating a minimalistic approach, we considered the use of features from ECG, EEG, SpO2 signals considering the sensors are more user-friendly and widely used. For RR interval correction we used malik’s rule followed by a cubic interpolation for the determination of normal-to-normal (NN) intervals [38]. From the NN interval series time domain and frequency domain features have been extracted following the HRV guidelines using the HRV Toolkit available from Physionet [39]. For the power spectrum estimation, we used Lomb’s periodogram method. The entire ECG record has been divided into 5-min epochs to estimate short-term components of HRV. In total 20 HRV features have been extracted.From EEG, we computed spectral features as described in Table 1. The feature extraction was carried out using a MATLAB App called SpectralTrainFig [40]. SpectralTrainFig was specially designed for SHHS to conduct spectral analysis of EEG signals using European Data Format (EDF) files. The EEG signal was collected using two channels from the central region of the brain. One of the channels was C4-A1 and the other one was C3-A2. The power spectral densities for these two channels are very similar. In our study, we have only used the signal from the C4 channel as it has been designated as the primary EEG channel in the movement (NREM) power, and Total power at each frequency band. Also, 102 EEG spectral features i.e., REM, N-REM power at single frequencies have been computed for 51 frequencies from 0 to 25 Hz with a 0.5 Hz gap i.e., 0 Hz, 0.5 Hz, 1 Hz, 1.5 Hz, …, 24.5 Hz, 25 Hz.Feature selection has been done primarily to compare the relative importance of ECG and EEG-based features for SDS estimation. Monte-Carlo Feature Selection (MCFS) and inter-dependency discovery has been used for ranking the feature importance. In MCFS the relative importance of features is estimated by building hundreds of trees for a randomly selected subset of features [41]. In a mathematic notion, i subsets of m randomly selected features are constructed where m << n, n being the total number of features and for each subset, k trees are constructed and their performance is assessed for classification/regression. Finally, i × m trees are constructed and evaluated. The procedure has been illustrated in Figure 3. Weighted accuracy of a tree as defined by Equation (2) is used as a metric to assess the classification or regression ability of the tree.
+          (2)Wac=1c ∑i=1cnijni1+ni2+…+nic
+          where c = number of classes, i, j = 1,2, …, c; nij is the number of samples from class i classified as class j and ∑nij = n is the number of all samples.The Relative Importance (RI) of feature gd denoted by RIgd is defined by Equation (3):(3)RIgd=∑x=1m.kWac x  u ∑rgdxIGrgdxno.  in rgdxno.  in xv
+          where Wac stands for the weighted accuracy for xth tree, IGrgdx stands for the information gain for node rgdx, (no. in rgdx) denotes the number of samples in the node rgdx, (no. in x) denotes the number of samples in the root of the xth tree, and u and v are fixed positive reals. Information gain (IG) is measured by Gini index or gain ratio [42].Both ECG and EEG have correlated features that introduce the problem of multi-collinearity. To deal with this, the inter-dependency discovery was used to remove features with strong pairwise interactions. rmcfs package from R has been used for feature ranking using Monte-Carlo feature selection and interdependency discovery (MCFS-ID) method [42]. The steps of preprocessing, feature extraction, feature selection, and regression has been shown in Figure 3.For developing the machine learning-based regression model we only used the EEG spectral features and anthropometric measures and explored machine learning and deep learning methods as described below:For developing the regression model, the dataset was randomly shuffled and was divided into the training (70%), validation(15%), and the test (15%) set. The training set has been used for gradient computation and updating the network weights and biases. The validation set was used to monitor error during the training process. The validation set is also useful in monitoring overfitting, as overfitting causes a high validation error while training error goes down. The test set was used for comparing different models. For finding the best model, we investigated Bayesian regression method, and artificial neural network (ANN). Bayesian inference facilitates overcoming insufficient data or poorly distributed data as it allows to put a prior on the coefficients and the noise so that in the absence of data, the priors can take over. In a Bayesian framework, the regression model is stated in a probabilistic manner where the Bayesian sampling algorithm returns a probability distribution (known as the posterior of the effect) that is compatible with the observed data instead of a point estimate. The posterior distribution is obtained by the product of the prior distribution and the likelihood function. The model for Bayesian linear regression can be represented by Equation (4):(4)y ~ NWTX, σ2I 
+          where response data points y is sampled from a multivariate Gaussian distribution that has a mean equal to the product of W coefficients and the predictors X and variance of σ2. I–is the N × N Identity matrix [43]. In this work, we used Markov Chain Monte Carlo Method (MCMC) sampling and weakly informative prior for Bayesian regression. To verify convergence potential scale reduction statistic R-hat was used [44].An artificial neural network (ANN) is capable of approximating any linear or non-linear relationship including multi-dimensional regression mapping problems quite well. However, the ANN must have enough neurons in the hidden layers and the data distribution should be consistent. During the training process, an ANN fits a function on a set of inputs to produce a set of associated outputs. Once training is finished the network forms a generalization of the input-output relationship and can be utilized to generate outputs for unseen inputs. The structure of ANN has multiple layers with interconnected artificial neurons as the building blocks for each layer. Each neuron has weights that are adjusted during the training process. Training stops when any of these conditions occur: the maximum number of epochs (repetitions) is reached, or the maximum amount of time is exceeded or performance is minimized to the goal or the performance gradient falls below minimum gradient. The ANN used in this study is of feed-forward type and has 3 layers–input, output, and hidden layer. The number of neurons in each layer is input-117, hidden–10, output-1. The used activation functions are- relu for the hidden layer and softmax for the output layer. Levenberg-Marquardt optimization with backpropagation was used as the training algorithm [45]. The hyperparameters used for the ANN are as follows: max epochs = 1000, min gradient = 1 × 10−7, momentum (Mu) = 0.001, Mu decrease ratio = 0.1, Mu increase ratio = 0.1. To facilitate proper training and evaluation the input data was randomly divided into training (80%), and test set (20%). Root mean squared error (RMSE) and R-squared (R2) values were used for performance evaluation of both the Bayesian model and ANN. Additionally, Pareto smoothed importance sampling (PSIS) diagnostic plot was used for the Bayesian model. Good Pareto k estimates (k < 0.5) in the PSIS diagnostic plot show that the model fits the data. The version of PSIS used in this work corresponds to the algorithm presented in Vehtari et al. [46].It is well perceived that obstructive sleep apnea (OSA) has a negative consequence on sleep and is a reason for sleep deficiency. OSA induces behavioral sleep problems, bedtime resistance, and a significantly shortened sleep duration [47]. Apnea-hypopnea Index (AHI) is used to quantify the degree of OSA. We investigated the correlation of ESS, FOSQ, and SDS with AHI to examine which one better captures the impact of OSA on sleep deficiency.The probability density plot of SDS computed using the Equation (1) has been shown in Figure 4. The histogram of SDS follows a Gaussian distribution with a mean of 60 (N = 500) and a standard deviation of 22. A boxplot comparison between the sleep deficiency severities of males and females has been shown in Figure 5. No significant (p-value > 0.05) difference was observed between the average SDS of males with that of females. SDS shows a moderate (r = −0.35, p = 0.0) correlation with age, i.e., higher the age, the higher is the sleep deficiency. The scatterplot of age and SDS with a trend line has been visualized in Figure 6. Additionally, SDS shows a weak (r = −0.21, p = 0.0) positive correlation with body mass index (BMI). Boxplots of SDS for normal and overweight categories are shown in Figure 7. The overweight category has a higher SDS.The partial correlation (controlled for age and BMI) of HRV and EEG features with baseline SDS computed using Equation (1) indicated a significant correlation for several features. A co-variate analysis have been performed to investigate the relationship of SDS with these features when controlled for age and BMI. The best five features from each sensor showing a significant correlation with SDS have been listed in Table 2.Although both HRV and EEG features show a significant partial correlation with SDS, the correlation for EEG features is much stronger than HRV features indicating that EEG features have relatively higher importance than HRV features in estimating SDS. Hence, in developing the regression method, only EEG and anthropometric measures have been used.MCFS results indicate that out of 150 features 117 are important based on the cut-off value of feature relative importance (RI) as shown in Figure 8. The line with red/gray dots gives RI values, the vertical bar plot gives the difference δ between consecutive RI values. Informative features are separated from non-informative ones by the cutoff value and are presented in the plot as red and gray dots, respectively. The convergence of MCFS-ID algorithm has been shown in Figure 9. The distance function (red line) shows the difference between two consecutive rankings–zero means no changes between two rankings (see the left y-axis). The common part (in blue color) gives the fraction of features that overlap for two different rankings (see the right y-axis). The ranking stabilizes after some iterations: the distance tends to zero and the common part tends to 1. Beta1 shows the slope of the tangent of a smoothed distance function. If beta1 tends to 0 (the right y-axis) then the distance is given by a flat line. The top-ranked 20 features based on normalized relative importance by MCFS-ID have been shown in Figure 10.The distribution of posterior R2 for estimated SDS using Bayesian regression indicates an approximately normally distributed pattern. In MCMC diagnostics R-hat values for all parameters were less than 1.1. A posterior predictive check on MCMC sampler has been shown in Figure 11a. The dark blue line shows the observed data while the light blue lines are simulations from the posterior predictive distribution. The patterns for both distributions are in agreement with some deviations for the peak. Figure 11b shows the probability density plot for the estimated SDS using Bayesian regression where the solid line indicates the point estimate from the ordinary least squares method. The plot–of–fit for the Bayesian regression model is shown in Figure 12 where R-squared value = 0.60 and RMSE = 5.63. The PSIS diagnostics plot for the Bayesian model has been shown in Figure 13 which reveals that only a few points are outside the acceptable threshold. The estimated shape parameter k for each observation is used as a measure of the observation’s influence on the posterior distribution of the model.For SDS estimation, ANN achieved a performance of RMSE of 4.65 and R-squared value of 0.86 in the training set, and an RMSE of 5.47 and R-squared value of 0.67 in the test set. The fit of the regression plot for the ANN model has been shown in Figure 14. The dashed line indicates the ideal trend line and the solid line indicates the fitted trend line for the actual versus predicted values. The histogram of prediction error showed symmetrically skewed and almost normally distributed patterns with a higher frequency in the error bin ± 2. The residual plot for the regression analysis shows a random scattering around the zero lines.A comparison of ANN performance with other models has been shown in Table 3. ANN outperforms other models with a lower error measured by RMSE. Figure 15 shows a direct comparison of the developed SDS scale with MW8 sleep quality score reported earlier by Landry et al. [37]. An inverse relationship between SDS and MW8 sleep quality score is observed with a negative correlation (r = −0.34), meaning that a lower sleep quality is associated with a higher SDS. Figure 16a shows the impact of OSA as captured by the Epworth Sleepiness Scale (ESS). It can be seen that ESS does not reveal an informative trend and failed (r < 0.1) to capture the impact of severe OSA on the sleep deficiency of OSA patients. Similarly, Figure 16b shows the correlation of the Functional Outcome of Sleep Questionnaire (FOSQ) with the apnea-hypopnea index. The trend in this case also fails (r < 0.19) to capture the impact of OSA severity on sleep deficiency. Figure 16c shows that SDS, as computed using the proposed method, shows a good positive correlation (r = 0.31) with AHI. As OSA severity increases, SDS also proportionately increases.While quantification and longitudinal monitoring of sleep deficiency are beneficial for early diagnosis and continuous monitoring of sleep disorder and may facilitate corrective habitual actions and practices that adversely affect good sleep, it is noteworthy that sleep deficiency is not only linked with the physiological disorder but also emotional stress and other factors. To reduce the variability in everyday measurement a moving average over a week or longer period as well as sleep pattern visualization may provide better insights when added to the SDS score. Signal quality and data reliability also impact the measurements and hence a data reliability metric may be used to enhance the usability of the method. It is noteworthy that we could not directly compare the utility of SDS with that of sleep score as sleep score formulae used by commercial entities are not publicly available to the best of our knowledge. Moreover, the mathematical model which served as a baseline for machine learning-driven models is not a gold standard but based on recommendation and findings of previous research studies. There is a gap between the polysomnogram based model and the machine learning based model as machine-learning model is only using limited information/features available from the EEG sensor. It actually points out that there is a trade-off between comfort and accuracy. While minimizing no of ensors do increase the user comfort level and make the process more user-friendly it does compromise some performance.In this study, we analyzed SDS and its relationship with HRV, and EEG-based features. We performed a feature ranking using MCFS-ID for identifying the most informative features for SDS estimation. Finally, we developed a regression method using ANN for SDS score estimation from spectral features of single-channel EEG. The findings of this study increased the interpretability of SDS and paves way for the usage of SDS as a potential indicator for automated sleep disorder checks using wearables. In future studies, we are aiming a large scale deployment of the model for longitudinal monitoring of SDS with wearables.A Smart Health(sHealth) framework was developed previously for implementing the sleep deficiency severity algorithm. The framework is available online at www.sccmobilehealth.com. Associated codes and resources are available at our GitHub repo at https://github.com/esarplab.M.J.R. and B.I.M. conceived of the presented idea. M.J.R. developed the algorithm and performed the computations including the data analytics. B.I.M. and C.P. supervised the findings of this work and verified the analytical methods. All authors have read and agreed to the published version of the manuscript.This material is based upon work supported by the US National Science Foundation under Grant No. 1637250.Approval has been obtained from an institutional review board at the University of Memphis (IRB# PRO-FY2017-474).The data used in this study was obtained from the National Sleep Research Resource. Anyone interested may apply here https://sleepdata.org/data/requests/shhs/start for getting access to the dataset.We acknowledge the contributions of B.E. Harmon, Md Sabbir bin Zaman, Sharmin Afroz, and Mamun Rahman in developing the overall Smart Health (sHealth) framework.The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.Description of HRV features used.Definition of spectral features used.Workflow for Events of interest capture and spatiotemporal visualization in sHealth.Subjective and objective methods for sleep health assessments.Method for feature extraction, feature selection, and regression for sleep deficiency severity.Probability density plot for SDS distribution.Comparison of SDS between males and females.Correlation of SDS with Age.SDS for BMI Categories.Relative importance of features.Convergence of Monte-Carlo feature selection.Top 20 features by MCFS-ID algorithm.(a) posterior predictive check on MCMC sampler (b) density plot of Bayesian model estimated SDS including the point estimate.Plot–of–fit for the Bayesian regression model.PSIS diagnostic plot and regression plot for Bayesian method.Performance of Regression model in the train, validation, and test set.Scatterplot showing correlation of SDS with MW8 sleep quality score.Correlation of (a) ESS with AHI (b) FOSQ with AHI (c) SDS with AHI.Spectral Features Extracted from EEG.Partial Correlation of SDS with HRV and EEG features.N.B. Variables in the table have been described in the Appendix A. A Comparion of ANN Performance with other models.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00009.txt

@@ -0,0 +1,4 @@
+A whole-body physiology model of inflammatory burn injury was used to train an algorithm to correctly detect patients’ states. The physiology model of a thermal injury takes the surface area of patient skin burned as an input to the model and responds to common treatments. This model is leveraged to build a database of patient physiology as a function of total body surface area burn, without treatment, over a 48-h window. Using this database, we train a model to determine patient injury status as a function of the available physiology data. The algorithm can group virtual patients into three distinct categories, corresponding to long term patient health. The results show that, given an initial virtual patient and injury, the algorithm can correctly determine the placement of that patient into the corresponding category, effectively classifying long term patient outcomes.Burn and related injuries in the United States account for approximately 40,000 hospitalizations each year [1,2]. The patient inflammatory reaction due to large total body surface area burns (TBSA) create complex physiological patient responses. Generally, the larger the surface area covering of the burn, the more complex the treatment protocols become [3]. Current clinical best practice guidelines note that burns exceeding 20% require changes in the patient care provided. These treatment changes account for fluid shifts into the interstitial spaces of the tissue due to inflammatory mediators that the body releases in response to the injury [4]. Inflammation is the body’s physiological response to thermal injury and is directly impacted by the size of the burn [5].The increasing complexity of patient treatment requires advances in detecting long term patient health. These advances may increase triage services and provide information to care providers not trained in burn-specific injury treatments [5]. Additionally, not all patients have ready access to quality burn-centric treatment facilities [6]. Military combatants, for instance, who suffer severe thermal injury usually receive initial treatment on-site prior to evacuation to a burn center. The importance of proper care, as a function of initial patient injury TBSA, cannot be overstated, as early fluid resuscitation has crucial implications for long-term recovery [7]. Machine learning has been leveraged in a variety of ways in hopes of improving patient care. Some of the most successful examples include using convolutional neural networks (CNNs) for classification, localization, segmentation, and registration in image analysis [8]. These algorithms are employed to detect abnormalities in the medical image and may one day be used in a clinical pipeline to improve patient outcomes. In addition, ML models have been studied to detect patient physiological decompensation when in the ICU. Most applications are trained on sepsis data in order to detect the onset of sepsis and septic shock earlier than tradition clinical analysis [9,10]. Indeed, these methods have shown some promise, reducing patient hospital stay and mortality when deployed in a clinical setting [11]. Burn injury and patient physiology detection has not been as robustly studied in a research setting. Much of the current research employs ML algorithms to determine kidney injury [12] and to classify the TBSA sustained by the patient [13,14]. The lack of large temporal clinical data sets of patient physiology is a barrier for training robust ML models that can determine patient outcomes. Detecting patient health as a function of initial patient injury requires a large amount of physiology data with TBSA as an input parameter. Though burn physiology data do exist in electronic health records, there are issues with the amount and types of data recorded. To avoid these issues, we leverage a whole-body physiology model of the burn pathophysiology to generate sufficient data to train our statistical learning model. The BioGears model of patient burn injury allows us to select specific, physiologically relevant data, as an output and lets us generate patient states, which encompass all available data provided by the engine. For this research the model is used to generate 79 data sets corresponding to a 24-h simulation of a patient with an initial burn injury. The initial model conditions increment the burn TBSA by one percent up to 40 percent TBSA. This data set is provided to the machine learning community as an open access database (see the data availability statement). For each patient, output data, in comma separated format, as well as each BioGears patient state, recorded each hour over the 24-h period is provided. A BioGears patient state records all the relevant data needed to begin a simulation and collects around 36,000 data entries. While we can assign a BioGears patient a TBSA, we do not know how a patient’s health evolves over time and consequently we do not know when and what interventions are appropriate for a given patient. The model is trained to classify longer term patient health by training in patient physiology data generated through a BioGears computational model. Hourly BioGears burn patient states are used as an input to a clustering algorithm. Three clusters are used when configuring the K-means clustering algorithm. The number of clusters is a function of subject matter expert feedback on general patient trends and trajectories present in the constructed BioGears model. Second, we predict a patient’s current category given their physiological state from BioGears. Using the same patient physiology data, a nearest neighbor classifier is trained that accurately predicts a burn patient’s category given their physiological state. These two contributions demonstrate preliminary results that an integrated BioGears-machine learning model could be an effective method to preemptively evaluate patient health trajectories. This model may also be able to predict treatment protocols that a caretaker may chose, based upon an initial patient injury and will be an area of future investigation.The physiological model is developed by simulating the patient’s acute inflammatory response (AIR) as a function of the burn TBSA suffered by the patient. Clinical data shows that the TBSA of the patient plays a large role in the inflammatory response of the patient [15,16] This model is based upon the Diverse Shock States model of Chow [17] as it accounts for the major inflammatory mediators implicated in patient response to a thermal injury. This includes tumor necrosis factor (TNF), interleukin-6 (IL-6), interleukin-10 (IL-10), and nitric oxide (NO). The model simulates downstream tissue damage incurred by the AIR that if the inflammation is left unchecked. This contributes to volume shifts into the interstitium, late-stage hypoperfusion and associated death. The model connects the inflammation to tissue damage, effectively simulating the whole-body physiological changes induced by large thermal injury [18].The model of the patient TBSA response in the BioGears physiology engine allows for dynamic runtime simulation of the patient state which allows us to design treatment scenarios in response to the patient injury [19]. In addition, we may leverage existing models of the cardiopulmonary and nervous systems to generate other downstream interactions with the thermal AIR model. The state equation for tissue integrity (T) is modeled after the approach of Reynolds [20]. The equations are adapted to provide a functional relationship for T to be constrained between 1.0 (healthy) and 0.0 (irreversible damage), which greatly simplifies downstream integration with BioGears. 
+      (1)dTdt=kD(1−T)−T(kD6IL64xD64+IL66+kDTRTr)(1xDNO2+NO2)Equation (1) describes the evolution of T as a function of cytokine concentrations. IL-6 and NO, and the insult stemming from thermal trauma (Tr). The constants kD6 and kDTR describe the rate of tissue health depletion due to IL6 and trauma, respectively, while kD captures the rate of tissue healing. Likewise, xD6 and xDNO represent half-max values for IL-6 and NO effects. Though mediators such as TNF and IL-10 are not explicitly included in (1), they affect the IL-6 population, indirectly impacting the calculation for T.Vascular compartments in the BioGears cardiovascular system maintain a pathway to an interstitial compartment, representing the tissue, Figure 1. Tissue paths consist of three elements: a resistor that may describe vascular permeability (kP) and two pressure sources capturing vascular and interstitial colloid osmotic pressure (COP). We calculate COP from the total plasma protein concentration (CPP) via the Landis-Pappenheimer approximation [21]:(2)COP=2.1CPP+0.18CPP2+0.009CPP3The majority of the blood protein represented in the BioGears vasculature is Albumin, thus we assume a linear relationship between albumin concentration (CA) and CPP such that CPP = 1.6 ∗ CA. We update CA each at iteration by calculating local albumin flux (JA) from the plasma (p) to the interstitium (i) via the Patlak equation [22,23].
+      (3)JA=JV(1−σ)(CA,P−CA,jexp(−Jv1−σPS)1−exp(−Jv1−σPS))According to (3), we non-linearly couple the albumin flux across the vascular endothelium to the volumetric flux (JV) which is precomputed by BioGears lumped circuit model. We buffer this transport by the reflection coefficient (σ), representing the degree to which a membrane maintains an osmotic gradient, and by the endothelial permeability to albumin (PS). Under normal physiological conditions, σ is approximately 1.0 and thus transport of albumin from the vasculature is approximately zero. The lymphatic transport re-circulates the small amount that does cross this member under healthy conditions, Figure 1. We tune the resistance across each vascular-tissue pathway (kp) to match the rates of filtered and transported albumin in circulation during steady-state.In response to hyper inflammation, the nervous system drives much of the physiological markers generated by the BioGears model, Figure 2 [24]. As the inflammation model reduces vascular volume, blood pressure and oxygen carrying capacity both decrease. A model of the central nervous system includes aortic, carotid, chemoreceptors, cardiopulmonary receptors, and pulmonary stretch receptors properly models the autoregulatory feedback due to hypovolemia. Each of these receptor contributions are gathered in the form of an efferent, sympathetic efferent, and an efferent vagal (parasympathetic) signal that update the existing cardiopulmonary BioGears model. We define the three signals as the sum of their respective receptor contributions:(4)fES,H=fes,∞+(fes,0−fes,∞)exp(kes(∑(wifi)i=14−θSH))fES,P=fes,∞+(fes,0−fes,∞)exp(kes(∑(w˜if˜i)i=14−θSH))fEV=(fev,0+fev,∞exp(fAB−fAB,0kev)1+exp(fAB−fAB,0kev))+∑(w˜if˜i)i=14−θSH
+      where fES,H is the sympathetic heart fiber signal, fES,p is the sympathetic peripheral fiber signal, and fEV is the vagal heart signal. Other parameters define the weights, wi associated with each receptor response, fi, the steady and minimum firing rate fes,0,∞, the slope parameter of the computed response, kes, and the signal threshold for a given response, ƟSH. We omit the equations determining the response to hypoxia for brevity but note that they are calculated to produce a functional relationship between the partial pressure of oxygen and the sympathetic firing rate. Once we have computed each of these signals, the cardiopulmonary model is updated using effector equations to vary the resistance and compliance of the lumped circuit elements that control the BioGears fluid circulation model. These updates effectively connect pressure and oxygen partial pressures to the heart rate, respiration rate, and other major physiological markers. This ensures that the inflammatory response that results as a function of TBSA influences the overall physiology of the patient.In the computational design, the burn action in BioGears is configured according to the severity of the wound as measured by the TBSA of the patient. The model takes, as an input, the TBSA and maps it to a value, Tr, which initiates the inflammatory cascade model. As time progresses, tissue integrity declines which proportionately decreases kP and σ (3). These changes increase both JV and, JA. The concentration gradient generated by albumin crossing the endothelium decreases vascular COP, thus increasing interstitial COP (2). The transport of albumin increases fluid shifts from the vasculature and further promotes albumin leak, causing a negative feedback loop that results in the increasing plasma volume depletion, Figure 3.Volume depletion initiates a nervous system response in BioGears, simulating the patient physiology in response to large TBSA, Figure 4. In addition, we note that burn wounds are generally coupled by a severe pain response in the patient [25,26]. The pain response is modeled by simulating the pharmacological effects of epinephrine released in the blood as a response to the pain suffered by the patient. This model drives further sympathetic outflow via the BioGears drug system—and increases respiration rate (RR), Figure 5.Detecting patient trajectories is formulated as a supervised learning problem: “Given an hourly state file, what trajectory does a patient belong to?”. Our solution to this problem consists of four steps. First, we extract features from BioGears hourly patient state files. BioGears is a high-fidelity simulator making it difficult to isolate a small number of key values. As a result, we extract 20,397 numerical values from each state file that represent the feature space of a patient trajectory. Our second step creates a patient embedding by reducing the dimensionality of the feature space using UMAP [27,28]. Reducing the dimensionality supports both learning from a small number of samples (there are only 79 total state files) and density-based clustering methods that use Euclidean distances (i.e., K-means). In our third step, we seek data-driven evidence that burn categories can be separated using physiological data. We note that observing the physiological response as a function of TBSA leads us to believe that there exist 3 distinct patient groupings. We use an unsupervised learning method (K-means clustering [29,30] to cluster patient states into three desired trajectories (K = 3). Our final step is a 4-fold cross validation that splits our 79 patient states into train and test groups. The test is kept separate and is used to validate the trained model. A patient’s current state is predicted using a nearest neighbor classifier trained on the UMAP patient state embeddings with ground truth labels from K-means and report on typical performance metrics.The model is demonstrated by simulating burns of 10%, 20%, 30%, and 40% TBSA, Figure 3, Figure 4 and Figure 5. These simulations show the progression of a patient without any intervention following the burn incidence. The results show the physiology appropriately responding to increases in TBSA. Note that for each graph in Figure 6, the lines for 30% TBSA and 40% TBSA stop early due to the patient not surviving the full 24-h simulation without treatment. The data demonstrate how a greater severity burn decreases tissue integrity which can affect circulatory flow through the affected compartments and increase the heart rate and respiration rate. Additionally, blood volume decreases at a greater rate with more severe burns, contributing to the death of the simulated patients with 30% and 40% TBSA.For the 79 patient physiological data sets, 20,397 values are extracted to create an initial feature space. Figure 7a represents the pairwise Euclidean distance between the standardized, scaled feature space. The small changes in shading indicate that there is little difference within and between different TBSA patients. Figure 7b contains the Euclidean distance between embedded patient states where the dimensionality is reduced from 20,397 to 15 features. The overall scale in the color bar indicates more confidence in the Euclidean distance metric. Additionally, changes in shading within each TBSA patient occur over the course of time, indicating overall physiological changes. Between each TBSA patient we can see that 10 and 15% TBSA are similar by darker red shading. The 20% TBSA patient comparisons to 10/15% TBSA are white and light red and light blue when compared to 30/40% indicating that 20% is its own burn category. The 30/40% hours are shaded blue when compared to other patients and light red to one another indicating that they are their own category. Overall, the UMAP dimension reduction highlights differences between TBSA patients that are difficult to identify when using the full feature space. This is likely due to removing unnecessary features from our coarse-grained extraction of physiological data from BioGears patient state files.After the UMAP dimensionality reduction, 79 patient states were clustered using K-means clustering (with K = 3). We use three clusters as it is consistent with the US Army’s use of rounding to the nearest 10% TBSA and considering a TBSA over 20% as severe (categories are 10%, 20%, >20%) [31]. Figure 6 shows two UMAP dimensions over time along with the resulting clusters. There are three distinctive clusters, 10/15 TBSA, 20% TBSA and 30/40% TBSA that correspond to trajectories of mild, moderate, and severe burns, respectively. Given the large initial feature space and small number of samples, this in an encouraging result and is supported by the US Army’s burn care heuristics. The only point of ambiguity is the initial first hour of all patients are together in the top left corner. It may be that during the first hour post burn it is difficult to identify the severity category on physiology alone. Should this observation hold with a greater number of samples, it would be acceptable since a trainee will also have the ability to visually inspect a patient.Lastly, we separate the 79 patient states into training and test data sets and use the initial K-means cluster as the ground truth labels. We performed a 4-fold cross validation (cross-fold validation is the typical method used to assess how well a data-driven model will generalize beyond the dataset) and despite the few number of training samples in each fold, a trained nearest neighbor classifier (using three nearest neighbors) is able to correctly predict all but two samples correctly, Figure 8. The two incorrect predictions were the first hour of patients with TBSA 15 and 20. Despite the small number of samples, this is an encouraging result that supports our goals of predicting a burn victim’s physiological trajectories given an initial burn TBSA. This is not an operational ML model but is an encouraging incremental result towards constructing a clinical pipeline that may, in the future, increase patient health related to burn outcomes.BioGears simulates the whole-body physiological response of burn injuries by coupling a model of inflammation with a cardiopulmonary circulatory and nervous system models. The model shows systemic physiological responses that follow appropriate trajectories, given an initial TBSA. The patient will die because of TBSA burns over 30%. A large data set of patient responses for a given burn TBSA is generated by simulating patients using this model. This database can then be leveraged to train a model to predict patient physiological trajectories. There appear to be three distinct grouping of patients within this data set that correspond to what we label as mild, moderate, and severe injuries. The trained nearest neighbor model can accurately sort patients into these three bins, given an initial patient state. The algorithm struggles to properly label a patient for the first few hours of the simulation for moderately burned patients, TBSA’s of 15 and 20. Future work that leverages real world data to test the efficacy of the trained ML model will be essential in determining clinical relevance. This will also be critical in determining if a computational model can properly train an ML algorithm as it relates to real patient data and will be an area of future work.To continue this work the ML model can be trained on a more complex data set that includes common patient treatments such as fluid resuscitation. Given proper treatment we would like to determine if this model can properly predict patient trajectory changes, given a certain level of care provided. To validate this model, we would like test it against available electronic health records, to determine how well it performs with real patient data. Given these tests, we would hope to provide a tool that may eventually be used by care providers to predict patient trajectories, aid in triage, and suggest patient care in order to transition patients into safer labels.All code, data, and results presented in this paper are available to the community via permissible Apache 2.0 open source license. Links to these resources can be found here. In addition, all patient state files and data associated with said simulations are also available under the creative commons 4.0 attribution and are available here.Conceptualization, A.A.-B., A.B. and M.H.; methodology, A.B. and A.A.-B.; software, N.T., S.W., A.B. and A.A.-B.; validation, A.B., M.S.-M. and M.H.; formal analysis, A.A.-B. and N.T.; investigation, A.B. and A.A.-B.; data curation, S.W., N.T. and A.A.-B.; writing—original draft preparation, A.B., N.T. and A.A.-B.; writing—review and editing, A.B., N.T., M.S.-M., M.H. and A.A.-B.; visualization, A.A.-B., and N.T.; supervision, M.H.; project administration, M.H. and M.S.-M.; All authors have read and agreed to the published version of the manuscript. This work and all data gathered herein was funded by Department of Defense contract number W911NF-18-C-0037. The initial design and idea for the research that led to this application was led and managed by Dr. Matthew Hackett out of Army Research Lab.No review boards were required for this study.No human subjects were used for this study.Python code to generate the results found in this document can be found on github. All the data used to train the models can be found on zeonodo, under the creative commons license. Data can be re-generated using the BioGears physiology engine, available under the open source Apache 2.0 license. We’d like to acknowledge and thank the large team at USISR who helped with discussions on the model construction: Michael Rowland, Alicia Williams, Jose Salinas and many others. The authors declare no conflict of interest.A sample vascular (red) and tissue (yellow) compartment pair for a given BioGears system circuit. The vascular and interstitial colloid osmotic pressures (COP_v and COP_i) are determined by (2) and (3), while kp is a property of the tissue. R_pre and R_post are pre- and post-capillary vessel resistance, while C_ves and C_tis are tissue compliance for each compartment.The BioGears physiology burn model overview [24]. The state of inflammation in the patient cascades into other downstream models to simulate the correct patient physiological response. This model allows us to generate simulated data sets of patient physiology as a function of burn total body surface area and is a safe way to collect data to train a machine learning algorithm.The BioGears response to burns of 10% (blue), 20% (orange), 30% (green) and 40% (purple) TBSA. Hypotension due to blood volume decreases is a function of the destruction of the tissue integrity and associated albumin flux into the tissue. Burns over 30%, left untreated, result in the patient dying (green and purple simulations).The BioGears response to burns of 10% (blue), 20% (orange), 30% (green) and 40% (purple) TBSA. Heart rate spikes due to blood volume decreases and the associated nervous system response. Burns over 30%, left untreated, result in the patient dying (green and purple simulations).The BioGears response to burns of 10% (blue), 20% (orange), 30% (green) and 40% (purple) TBSA. Respiration rate spikes due the patient’s pain response and decreases in circulating oxygen in the blood. Burns over 30%, left untreated, result in the patient dying (green and purple simulations).TBSA patient states (color indicates TBSA) are clustered (marker indicates K-means cluster) then displayed over time. While each patient begins in the top left in the first hour, they follow their own trajectories after this, corresponding to three burn categories.(a) is the Euclidean distance of the standardized, scaled full 20,397 feature space. (b) is a UMAP embedding of the feature space reduced to 15 dimensions. Reducing dimensionality is a helpful method when measuring Euclidean distance and when working with smaller datasets.A confusion matrix that summarizes a 4-fold cross validation of 79 patient states when training a nearest neighbor classifier. The only incorrectly classified patient states were the first hours of 15 and 20% TBSA.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00010.txt

@@ -0,0 +1,3 @@
+Metagenomics promises to provide new valuable insights into the role of microbiomes in eukaryotic hosts such as humans. Due to the decreasing costs for sequencing, public and private repositories for human metagenomic datasets are growing fast. Metagenomic datasets can contain terabytes of raw data, which is a challenge for data processing but also an opportunity for advanced machine learning methods like deep learning that require large datasets. However, in contrast to classical machine learning algorithms, the use of deep learning in metagenomics is still an exception. Regardless of the algorithms used, they are usually not applied to raw data but require several preprocessing steps. Performing this preprocessing and the actual analysis in an automated, reproducible, and scalable way is another challenge. This and other challenges can be addressed by adjusting known big data methods and architectures to the needs of microbiome analysis and DNA sequence processing. A conceptual architecture for the use of machine learning and big data on metagenomic data sets was recently presented and initially validated to analyze the rumen microbiome. The same architecture can be used for clinical purposes as is discussed in this paper.Current studies are showing the importance and contribution of communities of microorganisms, known as the microbiota, for human development [1], diet–microbiota interactions [2], interactions with the immune system [3,4], and diseases [5,6]. Although the relationships between individual microorganisms and host status are straightforward, we still lack information regarding the exact role of the vast majority of individual microorganisms in their respective environment and how they work together. Metagenomics studies, or the study of the whole genomic content of a given microbial community, or microbiome, are attempting to answer these questions [7,8].The traditional way to attempt to answer these and other research questions would be to take samples of the microorganisms from their environment and to culture these in a lab. Afterward, they could be studied and compared to other samples to detect similarities or differences in the composition of microorganisms between samples. This process is, however, fundamentally flawed as less than 1% of microorganisms in microbiomes can typically be cultured in this way [9]. A more modern approach is the sequencing of such microbiomes using high-throughput sequencing (HTS) platforms. Until a few years ago, sequencing could be quite expensive. At the beginning of the 21st century, when the human genome project terminated, the cost for sequencing one million bases was still several thousand dollars. Nowadays, sequencing the same number would cost less than one cent [10]. This price reduction opens up new opportunities for research and practical applications [7]. The increase in the number of metagenomic applications together with the decrease in costs and the desire for a deeper understanding of microbiota functions leads to a rapidly growing quantity of genetic data. The total quantity of data produced by sequencing in 2025 is estimated to be on par or above that of astronomy, YouTube, or Twitter [11]. Metagenomics will contribute a significant subset of this data. A single microbiome study can contain hundreds of gigabytes or more of raw sequencing data. During processing, this can get multiplied many times as intermediate results in different formats need to be produced. Therefore, it is crucial to have algorithms and system architectures capable of handling this quantity of data [12,13].Another topical research trend in the last decade has been the development of new and improved Machine Learning (ML) algorithms and techniques often summarized under the term, “deep learning” [14,15]. The word “deep” refers to the fact that ML models often use many processing layers. The level of abstraction and the ability to learn complex relationships increases with every layer [14]. For example, a deep learning algorithm, such as a Convolutional Neural Network (CNN) that is trained on images might detect simple edges in an image in the first layer. In the second layer, it might combine several of these edges to detect simple shapes such as rectangles. Finally, in the last layer, it could combine these shapes to detect complex objects. Adding yet another layer could enable the network to recognize the composition of objects to describe or classify a scene.These algorithms are only possible by taking advantage of the increase in processing power and especially General-Purpose computation on the Graphics Processing Unit (GPGPU) as they can be computationally expensive and often require a large set of data for processing [15]. Deep learning achieved promising results (often record-breaking) in multiple classification benchmarks as well as real-life applications, with a broad range of input data such as image, video, audio, or text [14]. It has also been successfully applied to the field of genetics [16] including metagenomics [17].CNNs are one of the most popular deep learning models. CNNs are able to automatically detect significant features from biological data and eliminate the need for manual feature extraction. However, challenges do exist in applying deep learning models to metagenomics classification problems [18]. Deep models have been used in prediction tasks, but how users interpret such a model remains an open challenge. A concern also arises in the application of deep learning models to metagenomics classification of phenotypes (linking metagenomic data to observable characteristics of the microorganisms or hosts), where there are more features than samples, which is often the case in predictive modeling of metagenomes. Therefore, accurate classification of diseases or disease subtypes is a key challenge in biomedicine driven by metagenomics [19]. These, and other challenges, are detailed in Section 5.The overall idea in metagenomic studies is to sequence (read) and analyze the metagenomic content of one or more samples [7]. Analyzing these samples is a process involving multiple steps. Although there is not a single template that can be applied to every study, there are some recurring steps that are similar or identical in multiple research studies.Most approaches can be distinguished into two broad categories: (1) those using amplicon sequencing, and (2) those using shotgun sequencing [20]. Amplicons, in this case, are short deoxyribonucleic acid (DNA) or ribonucleic acid (RNA) sequences that are specifically selected and “amplified” to serve as a unique marker and molecular clock to identify species and judge evolutionary distances in the phylogenetic tree [21] after sequencing. A popular choice for amplicons is the ribosomal RNA (rRNA) and more specifically the 16S subunit for bacteria and archaea and the 18S subunit for eukaryotes.In contrast to amplicon sequencing, shotgun sequencing (or “whole genome sequencing”) attempts to sample sequences from the whole genome. As most sequencing platforms can only sequence up to a sequencer-specific length, the DNA needs to be fragmented first [22]. After this step, the fragments from all contained microorganisms in the sample will be mixed. This also explains the name of “shotgun sequencing”, as the exact fragments sampled from the complete metagenome are random like the pattern produced by firing a shotgun.Having access to the complete metagenome allows for a broader range of studies than using amplicons alone, e.g., by analyzing genes instead of taxa. Regardless, amplicon sequencing is a wise choice for many applications. Compared to shotgun sequencing, the cost can be reduced significantly, as only a tiny part of the metagenome is sequenced and less data needs to be analyzed and processed [23]. At the same time, the taxonomic composition of a sample can often be determined with higher precision because all reads are focused on regions that are well known and vast reference databases exist for them. Hence, amplicon sequencing is a popular choice in particular for ecosystems where most organisms are expected to be closely related to existing entries in a reference database.Figure 1 shows the typical process for these two approaches as a loose sequence of steps. The diagram was adapted from Krause et al. [24] and focuses on bioinformatic processes. Sample preparation steps necessary before sequencing are not included. On the left side of the diagram processing steps typically associated with amplicon sequencing are shown, while on the right side, typical steps for shotgun sequencing are shown. Most studies will only use a subset of these steps depending on the goals of the study. Arrows in the diagram indicate a relative order, but do not necessarily imply that one step is a prerequisite of another step. To simplify discussions and to provide a generic structure in spite of this heterogeneity, the steps are grouped into five phases. These phases are displayed as separate sections in Figure 1 and labeled on the left side.The “Preparation” phase encompasses preprocessing of the raw sequence data to make it available to subsequent processes. In metagenomic processing solutions (see Section 6), this can include all necessary steps for data ingestion and format conversion. It also includes steps to ensure the necessary quality of data by trimming or removing low-quality sequences [20].In the “Aggregation” phase, the sequences are sorted, grouped, and assembled as necessary. For amplicon sequencing, this usually includes a clustering step to identify clusters of closely related microorganisms or Operational Taxonomic Units (OTUs) [25]. In shotgun sequencing, it is often desired to reassemble whole genomes or larger sequences from the short reads produced by the sequencing platform. “Read Binning” can support this process by identifying reads that have a high probability of originating from the same or at least a closely related microorganism so that the “Read Assembly” can be performed more rapidly and precisely [26]. For the assembly, overlaps between sequencing reads are identified and used as an indicator to determine the relative order of fragments. A set of overlapping sequencing reads is called a “contig” [20].The “Feature Engineering” phase was previously split into two distinct phases called “Annotation” and “Summarization” in the original paper [24]. As “Feature Engineering” is a more established and broader term, and better describes the intent of creating features for the analysis phase, we decided to merge these two phases. During this phase, databases can be used to map sequences to known reference data [26]. In the case of amplicon sequencing, they contain common amplicons and their taxonomic interpretation (e.g., name of the species). Shotgun sequencing, on the other hand, is more often used for “Functional Annotation”, where the focus is no longer on organisms as a whole, but rather on individual genes, which might or might not be shared between organisms. Common databases for gene sequences can include much metadata that can reveal specific functions of a gene within an organism or within an environment as a whole, which can then be used for further analysis. As a general optimization, or in case it is desired to identify unknown genes, it can be useful to perform “Gene Prediction” on sequences beforehand. This process attempts to identify patterns that indicate coding regions within the sequence.In general, the feature engineering phase is a transitional phase where the focus shifts from gathering data to the analysis and interpretation of said data [24]. This includes all transformational tasks such as the computation of relative abundances, where the previously identified genes or taxa are counted and put into relation to other genes or taxa in the same sample, or other more complex transformations known as “Feature Selection” [27] and “Feature Extraction”, that simplify and reduce the amount of input data for the analysis phase while preserving important information.As the name implies, the “Analysis” phase contains analytical processes that try to derive new information from the input features. Often it will be the most important step in metagenomic studies, directly supporting the study goals. Therefore, the processing steps are quite heterogeneous, depending heavily on the goal of the study. That being said, frequently used steps include the classification of phenotypes or computing correlations between genes or taxa. Finally, during “Interaction & Perception”, results from analysis or previous steps are presented to the user in a suitable fashion such as interactive visualizations or dialog systems. As with the analysis phase, visualizations are also heavily dependent on the study goals. Note that “Interaction & Perception” was originally named “Visualization” in  [24] but was changed to match the AI2VIS4BigData reference model established by Reis et al. [28] and to better describe the possible interactive nature of result presentation.Wassan et al. [29] considered various data sets containing identified OTUs from 16S rRNA amplicon data on various human body sites. They then demonstrated how metagenomic ML models can benefit from using hierarchical, phylogenetic information (e.g., considering evolutionary distances between OTUs). This section provides an overview of the study and shows how it can be aligned with the five phases mentioned earlier. For the purpose of brevity, only the Human Microbiome Project dataset [30] is used in this description (Figure 2).The tasks matching the “Preparation”, “Aggregation”, and part of the “Feature Engineering” are described in the supplementary documents of The Human Microbiome Project Consortium [30]. For “Preparation”, QIIME [31] and custom scripts are used to discard low-quality sequences and to de-multiplex sequence data from different samples. To cluster sequences and to pick representative OTUs, UCLUST [32] is used (“Aggregation” phase). Then, the RDP classifier [33] is used in the “Feature Engineering” phase to assign a taxonomy and an OTU table is created while applying additional quality controls.Wassan et al. [29] continued the “Feature Engineering” phase by creating a phylogenetic tree from the OTU table and then using this tree to conduct feature selection using various algorithms, such as random forest (Section 3.4). Figure 2 outlines this idea. For the “Analysis” phase, Support Vector Machines (SVMs) (Section 3.2), naïve Bayes (Section 3.5), and logistic regression (Section 3.6) were compared for classifying phenotypes given in the original dataset.The analysis phase forms a promising direction of metagenomic research by uncovering knowledge from the information of compositional profiles obtained from the steps of aggregation and feature selection.An additional key focus of studies by Wassan et al. is integrating biological domain knowledge of phylogeny [34] with abundance profiles of OTUs, aiding in inferring metagenomic functions from the microbial profiles. Research by Wassan et al. [29] suggests that the use of computation techniques using ML over integrative profiles could improve our understanding of microbial profiles and their functions.The random forest classification method has typically been employed in the prediction of high-dimensional metagenomic data [35,36] assuming independence between features. By integrating biological knowledge of phylogeny with the quantitative profiles, the prediction of metagenomic functions built upon considering these relationships could better inform metagenomic studies. Wassan et al. [29] employed the incorporation of both quantitative (abundances) and qualitative (biological domain knowledge of phylogenetic relationships) aspects into the construction of new feature space for the prediction of metagenomic phenotypes. A novel framework was proposed to determine metagenomic functions involving (i) engineering of new feature space via integration of biological and quantitative profiles, (ii) application of feature selection strategies over the engineered space, and (iii) application of the classical predictive models (such as random forest, SVM, naïve Bayes, and logistic regression) over the selected features. By employing the integrative approach, an improvement in predictive performance was observed by Wassan et al. [29], in comparison to previously employed approaches [35,36]. Furthermore, features space modeling has been characterized by different levels of varying phylogenetic depth (phylum to genus) in the study by Wassan et al. [29] which seems to be advantageous over traditional work where only features at genus level were considered [35,36]. The studies by Wassan et al. [29] improved modeling over the high-dimensional and integrative microbial feature space for determining functional predictions by involving phylogenetic analysis. Proposed approaches demonstrated high prediction accuracy.The phylogenetic similarity between microbial genes may also play an important role in determining metagenomic functions. Applying the ordination of Principle Coordinate Analysis (PCoA) over phylogeny and/or abundance wise similar matrix could provide some more support for differentiating microbial samples based on the functional phenotype [37]. As this paper focuses on demonstrating classification improvements for ML algorithms, the visualizations included herein consist of tables showing Accuracy and Kappa values for the different configurations. These visualizations can be generated by using validation metadata obtained during the analysis phase. The same data could also be used in graphical representations.The two primary uses of ML within metagenomics are classification and clustering. Classification algorithms learn to classify data points into a predetermined set of categories or “labels” from a set of training samples (supervised learning). Clustering algorithms on the other hand try to find and group similar data points without using predetermined classes (unsupervised learning).Another unsupervised type of ML often used in metagenomics that can be distinguished from classification and clustering is dimensionality reduction [13]. The goal of dimensionality reduction is to decrease the sparsity and complexity of data while preserving key information. These algorithms can be used as a preparational step for other algorithms that do not work well with highly-dimensional data or in the creation of visualizations that are by nature limited to few dimensions.Most ML techniques require their input to be vectors with n dimensions [38], where each dimension corresponds to a feature in the input data that can be used to distinguish the data. A dimension could be something as detailed as the nucleotide present at a specific position within a sample or something more general as the fraction of guanine and cytosine present in the sample (GC content). As the vector components must be numeric, suitable transformations must be applied if the original input data does not fit this criterion [38].Encoding methods can be distinguished into sparse encoding methods, where most vector components are zero, and dense encoding where the same information is encoded using fewer dimensions and most components are non-zero [39]. For example, the nucleotide at a specific position could be encoded in four dimensions using 0 and 1 to determine the presence or absence of a base or in a single dimension if each possible base is assigned a unique number (e.g., A=0.25,C=0.5,G=0.75,T=1).Some (simple) ML techniques perform better on sparse vectors, while more powerful techniques tend to perform better on dense vectors.Support Vector Machines (SVMs) [40] are used to separate input data into exactly two distinct classes. This works by computing a hyperplane in the vector space with the largest possible margin between data points of the two classes. A hyperplane is defined as a subspace with one dimension less than the vector space it is contained in. Therefore, in a 2D space, the hyperplane would be a simple line, while in a 3D space it would be a plane and so on.If such a hyperplane exists the data is called linearly separable. Training an SVM on representative, linearly separable data would result in a perfect classifier with 100% accuracy. Unfortunately, most real-world data is not completely linearly separable due to outliers or complex dependencies between the dimensions that cannot be expressed in a linear way. Thus, most SVM implementations, allow for such errors and try to minimize the error introduced by the outliers while at the same time maximizing the margin between the data points that are linearly separable.Sparse vector representations improve linear separability as there are more dimensions than the hyperplane can “use” to separate the data. Figure 3 visualizes this using a two-dimensional vector space. The features x1 and x2 are arbitrary examples of features that could occur during a metagenomic analysis. The dots represent samples that were used to train the SVM. The correct classification is indicated by the color of the dots—white dots belong to one class and black dots to another. The hyperplane (in this two-dimensional case a simple line), separates the two classes almost perfectly with only one outlier (the white dot below the line). The dots on the edge of the margin (i.e., the vectors that define the margin) on both sides of the hyperplane are called “support vectors” (circled in Figure 3).Another way to improve linear separability is to project the input vectors into a higher-dimensional space using a kernel function [40]. This is known as the “kernel trick”.SVMs are a popular choice as they are efficient to train and provide favorable results on many data sets [41]. Their performance drops if the data has complex dependencies and cannot be made linearly separable. Another disadvantage of using SVMs is that they require careful selection of features used as input vectors, as having many dimensions that do not contribute to the classification decision can reduce performance [41].A decision tree is a tree structure where each node splits a data set into two subsets based on a predicate (e.g., x1 dimension has a value of 0.5 or less). This process can be repeated at subnodes which divide the dataset further into smaller and smaller subsets. A decision tree can be used as a classifier by assigning class labels to the leaf nodes of the tree. The nodes should be constructed in a way that maximizes the predictive value at each level of the tree, i.e., the dimensions and values should be chosen in a way that best separates the data points into the target classes. Figure 4 demonstrates how a decision tree splits the dataset used in the previous section into distinct subsets that only contain one class of points (black dots or white dots). The lines represent a node in the decision tree and the predicate is written in the label. Each side of the line is a sub-branch of this node and can contain further decision nodes.Using a decision tree alone as a classifier is one of the fastest ways to build a classifier as the tree can be constructed very efficiently [38]. By using a simple binary decision at each level of the tree, the classification decisions are also easily comprehensible by humans which improves the overall explainability of the model (see also Section 5.4).Unfortunately, decision trees are too simple for many real-world problems and are prone to overfitting, meaning the tree can classify the training data perfectly, but fails to classify new data accurately [42].The random forest ML method [43] tries to tackle the shortcomings of the decision tree classifier. It uses not one but several (often hundreds or thousands) decision trees and then bases the final classification decision on a majority vote. As constructing multiple decision trees from the same dataset using the same dimensions would result in the same or very similar trees without improving the classification performance, a special construction logic must be used for the trees. This construction logic consists in using only a subset of the total data for each tree and limiting the dimensions from which predicates can be chosen to a random subset of the total dimensions in each level of the constructed trees. It can be proven that constructing the trees in this way avoids the risk of overfitting even if a large number of trees is used [43].Naïve Bayes classifiers are simple models that assume that all dimensions of the input data are completely independent of each other [38], meaning they contribute on their own to the probability of a data point belonging to a specific class or not without taking into account the values of other dimensions. Even if this assumption is often not true, naïve Bayes classifiers can still perform quite well in many cases. Depending on the distribution and type of values (e.g., categorical or continuous), various subtypes exist. For example, the Gaussian naïve Bayes classifier assumes a normal distribution for continuous values, while a multinomial naïve Bayes classifier can deal with categorical values.Figure 5 shows an example where the normal distribution for black and white dots was computed and shown as crosses originating at the mean and extending to one standard deviation (SD) in each dimension. The probability that a data point belongs to a class depends on this probability distribution. The curve shows the decision boundary where the probability for both classes is the same. Even though there are two outliers, the overall separation works well and categorizes almost all points correctly. Naïve Bayes is popular because of its simplicity and its capability to perform well with very few training samples.Logistic regression [44] is another simple model that can be used for classification. The model computes a weighted sum of its input vector and then applies the sigmoid function, which is also known as the “logistic function”, to this sum [44]. The result of the sigmoid function can be interpreted as the probability that a data point belongs to a given class. The training of a logistic regression classifier consists in adjusting the weights for the input vector so that the output function best matches the training data set.As the only input to the logistic function in the classifier is a linear combination of the input values (“weighted sum”), the classifier requires the input data to be mostly linearly separable to perform well (see Section 3.2). Figure 6 shows the result of applying logistic regression to our example dataset. The line represents the decision boundary where the sigmoid function has the value 0.5, indicating a 50% probability for both classes.The computations performed in a logistic regression classifier are the same as those in a single neuron of a neural network (see Section 3.8) that uses the sigmoid function as the activation function.As the clustering algorithms used in metagenomics are generally simpler and less diverse than the classification algorithms, they will be described only briefly in this section. In general clustering algorithms try to create clusters by finding similarity between input data and grouping it in a way that maximizes the similarity within a cluster while simultaneously minimizing the similarity between clusters. Similarity, in this case, has to be defined in some way using a distance metric. For metagenomic sequences this could be character-based, i.e., checking what percentage of characters between two sequences is the same, or based on more general properties (e.g., GC-content).Finding optimum solutions to these problems is generally not possible—even with small datasets—as it requires trying prohibitively many combinations. Therefore, heuristic algorithms are used that aim for a satisfactory balance between runtime and quality. One of the popular algorithms is Lloyd’s algorithm [38], which is often referred to as “k-Means clustering”. In Lloyd’s algorithm, a list of data points is chosen randomly to act as starting cluster centers (centroids). All data points are then assigned to the cluster with the nearest centroid. Afterward, new cluster centroids are chosen by calculating the mean of all data points within each cluster. This last step could change the nearest cluster centroid for some of the data points, thus assignment to clusters and calculating new centroids are repeated until there are no more re-assignments.While popular due to its simplicity, Lloyd’s algorithm has several problems, which are also relevant in metagenomics. First, the number of clusters must be chosen beforehand, which can be difficult as the best number of clusters is often not known. However, this is solvable by using variants where the number of clusters is chosen dynamically. Another problem is that calculations of mean and distance are based on Euclidean distance and hence require input data to be presented in a vector space (see Section 3.1). While this is also true for most classification algorithms described in this section, clustering algorithms in metagenomics are often used early within metagenomic workflows (see Section 4.3) where other forms of distance metrics (e.g., based on character data) are easier to compute. Finally, Lloyd’s algorithm has a worst-case superpolynomial complexity [45]. Processing huge data sets, as sometimes required in metagenomics, is therefore not always feasible.While Lloyd’s algorithm has several problems that limit its use in metagenomics, its simplicity and general principles are a sound basis for understanding clustering algorithms in general. Some specific uses will be discussed in Section 4.3 and Section 4.4.Neural networks are inspired by biological neurons. Common to both biological and artificial neurons is the idea that they are interconnected and that each neuron only performs simple computations while the combined network can perform complex tasks such as classification based on many input variables and complex dependencies [46].The neurons in most models take several real numbered inputs, compute a weighted sum, and then output a value based on a threshold. If the weighted sum is below this threshold the output strives towards one value (e.g., 0 or −1 depending on the model) and if the sum is above the threshold the output strives towards another value (e.g., 1 or the sum itself depending on the model).Artificial neural networks are usually organized in layers where the output of the neurons in one layer is the input of the next layer [47,48]. The input of the first layer is the features obtained from the samples and the last layer outputs the desired result, e.g., one output for each possible category in a classification task. Intermediate layers serve to enhance the computational strength of the network by allowing the network to “recognize” more complex dependencies within the input data that are required to determine the result [46].Compared to, e.g., SVMs, neural networks depend less on the data being linearly separable as they are inherently more powerful. An SVM can be mathematically expressed as a single neuron (single layer network), but in contrast, the computations performed in multi-layer networks are generally not expressible in an SVM [49]. This increased power also allows the use of dense encoding techniques as discussed in Section 3.1 and empirical data shows that deep neural networks (see Section 3.9) benefit from this approach [39,50].On the downside, the complexity of neural networks also demands more computational power to find the optimum values for all weights in the network. Furthermore, the number of variables in the model increases the risk of overfitting [51]. Overfitting occurs when a classifier tries to optimize its parameters to correctly classify all training samples as perfectly as possible at the expense of generalization.Figure 7 shows the same data points as were used in the previous examples. Here, the dots were used to train a hypothetical neural network. The gray background shows which area is classified as the “black dot” category, whereas the white background shows which are classified as the “white dot” category. The neural network was able to classify all points in the training set correctly. However, it seems unlikely that these irregular areas represent an underlying truth in the data. If the network is overfitted to the data used during training it is likely to misclassify new data points given to it. Intuitively the SVM did a better job at generalization in this example.Both this risk of overfitting and the required computational power limited the application of larger neural networks for some time. A collection of new techniques together with increased computational power using GPGPUs led to new popularity for neural networks under the label, “deep learning”.Deep learning summarizes a collection of techniques and advancements in recent years [14]. While some of these techniques can be applied to other ML methods, they are most often used in the context of neural networks. The term “deep” in deep learning refers to the higher number of layers compared to classical neural networks, which can improve the predictive power of these networks significantly [14]. In classical neural networks several factors limited the use of many layers:
+		Increasing the number of neurons leads to a higher number of variables which increases the risk of overfitting [51].The processing power for training large networks was not available [15].The popular “activation functions” used in neurons to compute the final output value of the neuron based on its inputs were proven to perform poorly in the context of many layers (known as the vanishing and exploding gradient problems in the literature) [52].Increasing the number of neurons leads to a higher number of variables which increases the risk of overfitting [51].The processing power for training large networks was not available [15].The popular “activation functions” used in neurons to compute the final output value of the neuron based on its inputs were proven to perform poorly in the context of many layers (known as the vanishing and exploding gradient problems in the literature) [52].With deep learning these challenges are addressed as follows:Deep learning networks are often trained with a significantly larger quantity of data. Many breakthroughs come from big tech companies like Google that have access to huge datasets (e.g., billions of images) which decreases the risk that the network only memorizes its training input and fails to generalize. Some computational techniques were also developed which keep the network from overfitting (e.g., “Dropout”, see Srivastava et al. [51]).The processing power was increased by offloading the training to the Graphics Processing Unit (GPU) and utilizing large computing clusters. The advent of cloud computing enabled this possibility not only for big companies but also for smaller research teams.The problem of vanishing or exploding gradients was mitigated by using vastly simpler activation functions, which do not exhibit this problem.Deep learning networks are often trained with a significantly larger quantity of data. Many breakthroughs come from big tech companies like Google that have access to huge datasets (e.g., billions of images) which decreases the risk that the network only memorizes its training input and fails to generalize. Some computational techniques were also developed which keep the network from overfitting (e.g., “Dropout”, see Srivastava et al. [51]).The processing power was increased by offloading the training to the Graphics Processing Unit (GPU) and utilizing large computing clusters. The advent of cloud computing enabled this possibility not only for big companies but also for smaller research teams.The problem of vanishing or exploding gradients was mitigated by using vastly simpler activation functions, which do not exhibit this problem.Due to these improvements over classical neural networks, deep learning has been able to solve problems that have previously been impossible. It broke records in many disciplines such as speech or image recognition [14] and many products such as digital assistants or search engines now use deep learning.In metagenomic studies, there are several steps that use ML algorithms or could potentially use them in the future. This section will cover typical processing steps and identify ML applications. Unless otherwise mentioned, the process will focus on sequencing results from high-throughput platforms like Illumina.The section is structured along the phases and steps shown in Figure 1. Readers are invited to use the figure as a reference while reading this section.The sequencing of DNA fragments works by reconstructing one of the two DNA strands base by base using specially modified fluorescent nucleotides with different dyes [22]. The original bases can then be detected by taking a photo at each step of the reaction and finding the dominating color. Determining the correct base from the images is a process called base-calling. Base-calling is complicated by multiple sources of errors during this reconstruction.Machine learning has been applied to this problem. One example is the use of SVMs, which showed promising results in a comparison of various base-calling algorithms [53]. However, the literature on base calling using more modern deep learning techniques is sparse. One explanation for this may be that the results of simple algorithms already have low error rates and that there is therefore little room for improvement. In nanopore sequencing (an alternative, emerging sequencing method), where the error rate for base calling is significantly higher, deep learning has been applied successfully, which further supports this assumption [54,55].After base calling the (digital) result of the sequencing operation will be a collection of “reads” representing the fragments that were sequenced with corresponding bases and quality metrics [20].Common preprocessing steps to be applied on raw sequencing data include demultiplexing reads from various samples by identifying attached barcode sequences, discarding or trimming low-quality reads, and removing sequencing artifacts.A review of relevant literature revealed no significant use of ML in these areas. State-of-the-art toolsets like QIIME 2 [31] use relatively simple rules for these steps.Some authors [19], however, do see potential for deep learning models that cover the entire processing pipeline, including preprocessing steps, in the future.For amplicon sequencing a common next step is to cluster reads belonging to the same species or higher taxonomic rank into OTUs for later processing by searching for reads with similar sequences (typically, 97%) [25]).The goal of clustering into OTUs is to reduce the quantity of data that needs to be processed in later steps. Clustering was discussed in Section 3 as one of the two important branches of ML types. Accordingly, there are several algorithms that can be used in this task [13].One popular algorithm is UPARSE-OTU [56]. UPARSE-OTU is a greedy algorithm (Section 3.7) and thus assigns clusters in a single iteration. The algorithm starts by ordering input sequences by abundance as a high abundance indicates a non-erroneous read and therefore might be good OTU candidates. Starting from the first sequence, the algorithm will either assign the sequence to an existing cluster, create a new cluster using the sequence as the OTU reference sequence, or discard the sequence if it is believed to be chimeric. A sequence is assigned to an existing OTU cluster if its similarity is above 97%. Similarity is defined here by the UPARSE-REF algorithm [56] which uses a maximum parsimony model, i.e., defining similarity by the number of “events” that need to occur to go from one sequence to another sequence (e.g., a sequencing error). As the Maximum Parsimony model is a satisfactory approximation of similarity, there is little practical usage of more advanced ML models.Another use of clustering algorithms is for the purpose of “Read Binning” in shotgun sequencing—often used as an optimization step before sequence reassembly. Binning tries to separate reads from different organisms so that the reassembly process is easier and less prone to false overlaps between only far-related organisms [26].In contrast to OTU clustering, where sequence reads are expected to map to the same position/amplicon, the actual overlap between two sequences belonging to the same “bin” can be less or non-existent as the individual reads map to a longer sequence. Instead of relying on sequence similarity, binning algorithms hence often rely on other similarity measures such as k-mer distribution or probabilistic models [13]. The former relies on the fact that the distribution of k-mers (i.e., short subsequences) is similar within a genome while the latter attempts to model probabilities for reads to belong to the same bin and then uses expectation maximization to find the most likely result. Early attempts also used neural networks for this task [57].HTS platforms are limited to DNA fragments with a few hundred bases at best for technical reasons [22]. This can be sufficient for many metagenomic applications. For example, phylogenetic studies based on amplicon sequencing can use rRNA regions that are short enough to be read completely but still unique enough to differentiate species [23].When complete genomes or larger sequences need to be sequenced, they need to be split up into fragments smaller than the maximum read length of the sequencer first [22]. However, this creates two problems: First, the order in which the fragments originally occurred within the genome is lost, and second, the organism from which the fragment originated is also unclear as the fragment gets mixed with fragments of all other organisms in the sample [9]. This is not necessarily a problem if only the presence and abundance of certain genes in the sample need to be known. In this case, it can be sufficient to match the fragments directly to adequate databases for the research topic [26]. However, in other cases, the reassembly of the original genomes present in the sample is desired.Reassembly happens by finding overlaps between the reads and aligning them on each other. Several strategies and tools exist for this [20] and some of them are specialized for metagenomics [26]. There are some promising results for genome reassembly using both classical ML algorithms as well as deep learning techniques, but most of them are still relatively new and their performance on real-world samples is yet to be determined [58].Taxonomic assignment is the task of choosing a label for each sequence or cluster based on a chosen taxonomy, e.g., finding the species that a sequence is associated with. To facilitate this, reference databases with sample sequences for the various labels can be used. The methods for taxonomic annotation can broadly be separated into two groups. The first group of methods is used in amplicon sequencing and utilizes rRNA databases. Here, one of the most popular tools is RDP classifier [33], an ML approach using a naïve Bayes classifier. The second group constitutes methods for assignment of reads or clusters of reads taken from shotgun sequencing in whole-genome studies. In this group, non-machine learning-based approaches like Kraken [59] which rely on exact substring matches seem to be more popular.One challenge faced by ML algorithms is the large number of possible species and similarities between substrings in even distant species [60]. However, there are some indications that less reliance on exact matches could also be a strength for ML algorithms when dealing with species not part of the training set [13,60]. This is particularly important when comprehensive reference databases for the chosen environment do not exist. For many microbiome environments, whole-genome samples only exist for common or otherwise interesting species [26].While taxonomic annotation tries to answer the question of “what is a sample composed of?”, functional annotation tries to answer the question of “what do the components do?”. Initially, the task can be very similar to that of taxonomic annotation. However, the reference databases used in functional annotation consist of individual genes and associated functional categories. In this case, the challenges for ML algorithms are very similar to taxonomic annotation, and ML is hence not often used for this task [13].However, machine learning is often used in gene prediction [13]. Gene prediction tries to find genes within samples without matching them to existing databases and is thus often one of the first steps when exploring new genomes. A challenge in metagenomics is that complete genomes can often not be assembled so that the analysis can only be based on individual reads or partial assemblies. Most gene prediction algorithms are based on hidden Markov models [13]. Other popular tools for metagenomic gene prediction use neural networks [61] and there have also been successful attempts to apply deep learning to the task [18,62].“Feature Selection and Extraction” is usually applied for removing irrelevant and redundant features from the high-dimensional metagenomic datasets. Three standard techniques are [27] (i) filter-based techniques that use heuristic functions over general statistical characteristics of data to determine important features such as correlation between features, (ii) wrapper-based approaches that iterate an ML algorithm over the input features to determine important features, and (iii) embedded methods intended to determine features by measuring performance while the model is being constructed.Wassan et al. [63] conducted a comprehensive analysis of metagenomic data from human microbiomes with application of different feature selection strategies. They recommended the application of embedded feature selection methods, namely, extreme gradient boosting (XgBoost) [64] and penalized logistic regression to determine important microbial genes in metagenomic studies. Automatic selection and weighting of features is one of the promises of deep learning, which could make this separate step obsolete in the future [14,15].In phenotype classification, one trains a classifier to predict certain characteristics based on metagenomic sample data. Starting from training data where both the metagenomic sequences and the associated phenotype are known, the classifier learns to apply the categories to previously unseen data. This helps us to answer questions like “How can sick and healthy individuals be distinguished by their blood microbiome” [65], “How does the maternal microbiome affect the development of newborns?” [1], “What role do microbiomes play in the relationship between diet and metabolism?” [2], “How do microbiomes interact with the immune system?” [3,4], and “How do they relate to diseases?” [5,6]. These classifications can be based on relative abundances of taxa or genes, and additional features obtained from the samples [13,63].All of the supervised learning algorithms discussed in Section 3 can be used for this task. The example introduced in Section 2.2 used, for example, SVMs (Section 3.2), naïve Bayes (Section 3.5), and logistic regression (Section 3.6). While still rare, some studies also use deep learning for phenotype classifications [17], although successful application is often a challenge due to an insufficient number of samples [19] (see also Section 5.7).Nevertheless, recent advances open up a new avenue for applying deep learning models in data analysis [66]. For example, Zhu et al. [67] successfully applied a new deep learning model called Deep Forest [68] to investigate microbiome associations. One key advantage exhibited by Deep Forest is that it can work well even with small-scale training data.Other frequent analysis tasks include, for example, diversity estimation, phylogenetic analysis, or correlation analysis. Diversity estimations are statistical measures to determine the variety of species present within a sample (alpha diversity) or the different composition between samples (beta diversity) [37]. Correlation analysis measures the relationships between different genes or species across samples. In a positive correlation, the increase of one gene or one species increases the abundance of another gene or species. In a negative correlation, the increase in one leads to a decrease in the other. Both analyses are based on statistical measures and can be computed directly from the compositional data of samples. To our knowledge, there are no attempts to provide alternative measures using ML.Phylogenetic analysis tries to determine the evolutionary relationships between microorganisms and to order them within a phylogenetic tree. In general, algorithms in this category treat genetic mutations as a distance measure. The closeness on the phylogenetic tree depends on the number of mutations required to explain sequence differences between two species. The algorithms to compute these distances and to construct phylogenetic trees from the set of all sequences range from simple distance methods to more complex probability-based models [69]. Machine learning algorithms are not frequently used.Visualizations are an important tool to understand the data generated at various steps. One example is the visualization of sample composition using taxonomic or functional annotation data. Figure 8 shows a sunburst diagram created with the free software Krona [70] and using the dataset from [71]. The rings of the sunburst diagram represent different levels in the phylogenetic tree, with the outer rings being more and more specific. The size of the ring sections is determined by the relative abundance within the sample.Figure 9 shows another example of taxonomic composition using a phylogram. It was created with the software Megan [72] using a dataset from [73] (sample file AS53_18). In contrast to the sunburst diagram, the hierarchical structure of the phylogenetic tree is emphasized more while the relative abundances are only shown as a heat map and, thus, more difficult to compare. Both visualizations are examples of combining a “Composition Diagram” with a “Phylogenetic Tree”, emphasizing one or the other aspect.For simple graphics or diagrams, ML algorithms are not needed. However, there are often cases where highly-dimensional data needs to be presented to the user in a way that highlights important features while discarding non-essential data. This is a typical use case for dimensionality reduction algorithms. Examples include the creation of correlation networks or visualizations of clustering and binning (see Section 4.3 and Section 4.4). A practical application for this is described in Laczny et al. [74].Machine learning can be used successfully in many key areas of metagenomics. However, there are several challenges when applying ML to metagenomics, some of which might also impede broader usage of more advanced techniques like deep learning.As seen in Section 3, many ML approaches can be used in the field of metagenomics. Within complex approaches such as neural networks, the models themselves have infinite possible configurations (e.g., number of layers and neurons) to choose from. Finding a satisfactory model is often a difficult and time-consuming task even for experts in the field. It can involve a lot of trial and error without a guarantee that the result will outperform older or simpler models [75]. The necessary time and knowledge to select the best model for a specific task from a myriad of options are often lacking.Deep Learning requires new approaches to really gain an advantage compared to simpler ML models. While simple models require careful feature selection and will fail to detect meaningful patterns if the input space is too large and complex, deep learning promises to find suitable representations and meaningful connections in a large feature space without requiring the same level of domain expertise and feature engineering beforehand [14,15]. Just switching out algorithms and using the same features for deep learning neural networks that were used for simple ML models will limit their potential or might even make them perform worse than the simpler models.On the contrary, the feature space should be extended by adding data from additional data sources where possible to take full advantage of these advanced ML models. Examples in the clinical field include models that not only take into account data from the metagenomic data itself, but also the patient’s medical history or other laboratory data. As with the model selection itself, using deep learning and selecting relevant features is a difficult task requiring suitable architectures and ML expertise, which are not always available. Including additional data sources increases its complexity even further.Deeply related to the challenges of selecting a model and features is the accessibility for these processes in analysis systems. Assuming that experts in metagenomics are not usually experts in ML and that hiring ML experts for a metagenomics project is not always an option, the process of selecting models and features should also be accessible for non-data scientists. This is a challenge specifically relevant for the user interface used in these systems, which needs to allow flexible configuration of analysis tasks and model configuration for all kinds of use cases while still being easy to use.Explainability strives to make the reasoning behind decisions taken by automated algorithms, such as ML methods, more transparent. More explainability is in high demand. For example, the European Commission has committed itself to a more trustworthy and secure use of artificial intelligence (AI), which includes explainable AI [76]. This is particularly important for the field of medicine as decisions can have far-reaching consequences and there is a high demand for understanding what specific reasoning led to them [77].However, explainability is not straightforward with complex ML algorithms as these are often black boxes. Specifically with deep learning, the large number of layers and neurons can make it impossible to understand the reasoning behind a decision taken by the algorithm. Designing models in a way that improves explainability is possible but could lead to decreased accuracy, although there is some debate about whether this is necessarily true [77]. It is worth questioning the practical importance of explainability [78]. Assuming that there is in fact a trade-off between accuracy and explainability, there are certainly patients and physicians that would prefer a hypothetical black-box algorithm that is 99% accurate in predicting a medical condition instead of another hypothetical algorithm that is only 90% accurate but provides full transparency in its decision process. There are also numerous examples in classical medicine of effective treatments that have been used without knowing why they work [78].Both determining the need for explainability and managing the possible trade-offs between explainability and accuracy are challenges when implementing new clinical solutions utilizing ML.Closely related to explainability is the concept of reproducibility. In a metagenomic analysis, multiple steps are needed between the initial raw data and the final visualization or result data. It is crucial to be able to reproduce these steps whenever necessary. While this can be done by vigorously documenting each processing step, it is not an easy task as results can depend on the exact parameters of the involved tooling or even a specific environment or version number [79]. It is also easy to make mistakes if these steps have to be performed by hand. The use of ML and trainable classifiers, such as neural networks, can increase this problem as the classification results can change with every retraining of the model, even if the input data remains the same.The microbiome in human hosts and other environments can contain millions of species and genes [13]. Many of those are still unknown, which can make taxonomic or functional annotation of sequencing data difficult. Existing computational methods and databases also have a bias towards bacterial data, which makes analyzing the whole metagenome including archaea and fungi more difficult [13,80].Even phylogenetically distant species often share a lot of similarities, which can be a problem with using ML algorithms looking for these patterns. Simple algorithms relying on exact matches are thus often preferred to less strict ML algorithms in taxonomic annotation tasks based on shotgun sequencing (see Section 4.6) [60]. On the other hand, less reliance on exact matches could also be a strength for ML algorithms when dealing with species that are not part of the training set [13,60]. This can be important when comprehensive reference databases for the chosen environment do not exist yet [26].The nature of metagenomics often leads to relatively few samples (e.g., blood samples from a dozen patients) with very high dimensionality, i.e., thousands or millions of species in the sample [19]. This combination makes it difficult for ML algorithms to find clear patterns in the data that can be used in their decision processes. Modern ML algorithms like deep learning neural networks work by training many parameters and do not need the same level of feature selection as classical algorithms. The ability to infer suitable data representations on its own is one of the greatest strengths and a defining characteristic of deep learning [14,15].However, this ability usually requires access to large data sets, to allow the algorithm to recognize complex patterns [15,81]. Having access to such datasets with labeled data is, therefore, crucial for metagenomics. Large datasets of, e.g., patient data, can be difficult to obtain, however, as the process of collecting and processing a large number of samples is still expensive, and legal requirements also need to be taken into account.Big Data is often defined by the three V’s: volume, velocity, and variety [82]. Each of these can be relevant in metagenomics and represent a unique challenge. The volume of data results from the usually large number of sequencing reads required to identify and differentiate the microorganisms found in metagenomic samples. These can lead to many gigabytes or even terabytes for a single study. During processing (e.g., for metagenomic assembly), the required system memory can also reach several terabytes. This is out of reach for desktop computers and requires special infrastructure [13,26]. The secure archival and storage of intermediate and final results presents another challenge as archiving of all intermediate results can increase the amount of storage needed manyfold [79]. Archival must also comply with all legal requirements including minimum and maximum retention times for various types of data in an automated way.The second “V”, “Velocity”, refers to the speed at which new data is being generated and specifically for metagenomics the rapid growth of data. The total quantity of sequence data has been growing exponentially, doubling approximately every seven months, and is expected to reach the quantity of data produced by other big data applications such as astronomy, YouTube, or Twitter [11]. Metagenomic applications need to be able to scale with this increasing quantity of data, which is a challenge both for algorithms that need to be efficient and for infrastructure, which needs to be scalable [79].For deep learning the volume and growth of metagenomic data could become a problem for training the algorithms “on-site”. Looking at fields like image or text processing some models are trained using such large quantities of data in order to improve classification performance that the cost of training becomes prohibitive for all but a few very large corporations [83,84]. There is a growing trend of separating training and usage of ML models or to use transfer learning, where existing models are adapted and retrained to specific use cases [85]. Metagenomic systems that integrate ML models should be prepared to allow these hybrid approaches.Last, the “Variety” of data results from different data formats used and the combination of sequencing data with other data sources for analysis, which can include unstructured or semi-structured data such as scientific publications or medical histories. Integrating all these types of data so that they can be processed together is a challenge [86].A common approach in big data applications is to split the processing of data into individual steps. These steps can then be optimized separately, e.g., by using parallelization. This separation also facilitates configuration of study-specific workflows in the form of metagenomic pipelines, which is the topic of the Section 6.Many projects aim to facilitate metagenomic or biomedical analyses in general by providing step-by-step processing pipelines. In a pipeline, individual components work together to create a result. The output of one step in the pipeline is the input for one or more following steps. This allows flexible reconfiguration of the pipeline or individual steps depending on the requirements. At the same time, configured pipelines can be saved and reused later for similar studies. Individual steps can easily be replaced with alternative approaches as long as the input and output formats stay the same. This section will provide some examples of these projects.The Galaxy project [87] is a web-based platform for biomedical analyses including tools for metagenomic research [88,89]. It integrates several thousand tools in its “ToolShed” ready to be used in custom workflows that can be defined in a visual interface. There are tools to cover the complete metagenomic workflow from the processing of raw sequence data up to visualizations. Galaxy also provides access to a wide range of ML tools and algorithms including popular libraries like scikit-learn [90] and Keras [91]. The project is enabled to support many concurrent users using an infrastructure scalable across multiple computing nodes. It can be used on free public servers, pay-as-you-go cloud services, or installed locally.Galaxy processing is based on files. Individual processing steps can consume files from previous processing steps and output their results as new files. This facilitates adding existing tools to Galaxy by creating a wrapper that describes its expected inputs and outputs. However, this reliance on files can have disadvantages. First of all, data has to be constantly written into suitable output formats only to be parsed again by the next step in the pipeline, which is inefficient. Furthermore, the execution of new steps cannot start until the files from previous steps have been written completely. Finally, parallelization of individual processing steps is more difficult as the data contained in files cannot be easily split without reading and parsing the whole file.MG-RAST [92] is an open submission platform for metagenomic data, that integrates automatic processing like taxonomic and functional annotation of sequence data. Data submission is possible using the web-based interface, a scripting interface, or a REST-based API. As the focus of MG-RAST is on ease of use, fast processing, and using standard procedures across all submissions, the pipeline is fixed and customization is limited to setting several parameters before starting an analysis. As an archive, MG-RAST has an extensive repository of metagenomic data with over 473,000 metagenomes containing more than 2000 billion sequences (as of November 2021, www.mg-rast.org, accessed on 4 November 2021). As MG-Rast does not allow custom pipelines to be used, it is not possible to extend the processing with new machine learning-based steps. There is also little use of ML in the fixed pipeline itself.A similar solution is offered by the European Bioinformatics Institute (EBI) as MGnify (formerly EBI Metagenomics) [93]. Like MG-Rast the platform allows free submission and analysis of metagenomic data using fixed pipelines. Depending on the type of study performed (e.g., shotgun or amplicon-based analysis) different pipelines are available providing all usual processing steps like functional and taxonomic annotation and a range of visualization and comparison options.In contrast to the other tools in this section, QIIME 2 [31] does not aim to be a full-fledged platform with ready-made workflows accessible through an easy-to-use interface. Rather than that, it is a collection of python scripts that can be used together to do metagenomic analyses locally. Using the command line as the primary interface provides a lot of possibilities for customization and facilitates integrating other command line-based tools into the workflow. Documentation of the steps in a metagenomic study is also straightforward through providing the executed command lines within the paper or in a shell script in an accompanying source repository. The commands are designed to be run locally by default, although it is possible to run some of the jobs in parallel or on a cluster. For developers, the script-based nature of QIIME 2 facilitates integrating ML algorithms as new processing steps. The source code repository includes an example of training a random forest model for phenotype classification [94].Although QIIME 2 provides a GUI as an API as alternatives mean to access its functionality, the focus is clearly on the script-based interface, which makes it difficult to use the tools as a user not familiar with command-line interfaces and scripting. Setting up clusters to run commands in parallel is also not straightforward and to our knowledge, there is no hosted, web-based interface that allows complete access to the functionality of QIIME 2. QIIME 2 is not a complete workflow system on its own as the user is entirely responsible for the execution of tasks and the management of intermediate and final results. However, QIIME 2 can be integrated into other biomedical workflow systems such as Galaxy.The idea of MetaPlat is not only to provide comprehensive analysis tools for metagenomic data, but to support the complete life cycle of metagenomic studies, including archiving, taxonomy management, and visualization of results. To achieve this, it is integrated with the Knowledge Management Ecosystem Portal (KM-EP) [12]. The architecture was designed based on best practices of big data systems, to ensure scalability. Another goal was to use new and innovative ML models and visualizations to help researchers in understanding the collected data [95,96].The system supports reproducibility and tracks the phenotypic information associated with each sequence, including its origin, quality, taxonomical position, and associated biological genome, and produces automated reports and visualizations. Additional objectives of MetaPlat include the following:
+		Sample gut collection, from cattle, for sequencingCollection of publicly available databases to create a new classification of previously unclassified sequences, using ML algorithmsDevelopment of accurate classification algorithmsReal-time or time-efficient comparison analysesProduction of statistical and visual representations, conveying more useful informationPlatform integrationInsights into probiotic supplement usage, methane production and feed conversion efficiency in cattleSample gut collection, from cattle, for sequencingCollection of publicly available databases to create a new classification of previously unclassified sequences, using ML algorithmsDevelopment of accurate classification algorithmsReal-time or time-efficient comparison analysesProduction of statistical and visual representations, conveying more useful informationPlatform integrationInsights into probiotic supplement usage, methane production and feed conversion efficiency in cattleThe bioinformatic workflow engine used by default in MetaPlat is called Simplicity [97], developed by the company NSilico. MetaPlat is an EU-funded Horizon 2020 project developed by several universities and other organizations, including some of the organizations represented by the authors of this paper. MetaPlat was validated and used for the use case of rumen microbiome analysis.Partly inspired by MetaPlat, but with an even stronger focus on AI and the goal of expanding its use beyond rumen microbiome analysis, another conceptual architecture was recently presented and initially validated. The “AI2VIS4BigData Conceptual Architecture for Metagenomics supporting Human Medicine” is described in detail in the Section 7.To support the applications discussed in this paper, including addressing some of the challenges described in earlier sections, a conceptual architecture for AI and big data supporting metagenomics research (Figure 10) was introduced in Reis et al. [98]. It is based on the general AI2VIS4BigData reference model [28] for AI and big data applications.The AI2VIS4BigData architecture is split vertically into three pillars to separate data ingestion from analysis and visualization, applying the design principle of Separation of Concerns (SoC) [99]. Each of these pillars is split into three layers, following the Model View Controller (MVC) [100] pattern, to separate the persistence (Model) from the user interface (View) and the application logic (Controllers). The persistence layer is shared between all three pillars.The first pillar is responsible for data ingestion. Data such as metagenomic sequences or other laboratory data as well as subject-independent reference data is imported into the system and processed using a mediator/wrapper approach [101]. The wrapper is responsible for transforming the data into a common schema and the mediator is responsible to facilitate communication between the individual data sources and the rest of the system. The intent behind this is that the data formats that can be imported should be easily extensible. Another important aspect of the data ingestion pillar is the storage of raw data in a data lake. This allows better transparency and reproducibility as all steps can be repeated starting from raw data.In the second pillar, data is taken from the persistence layer and then processed for analysis using a workflow engine to orchestrate the required metagenomic tasks. The configuration for these tasks is provided by domain and data science experts using a configuration user interface. The results are persisted into structured storage again.Finally, the third pillar is responsible for presenting the persisted analysis results to the end-user and possibly entering into a dialog with the user to further adjust the presentation.In Section 2.2, an example was introduced in which human metagenomic data is used for the classification of phenotypes. So far, the conceptual architecture presented in this section was only validated for the use case of rumen microbiome analysis. While it seems trivial at first to apply the same architecture for human microbiome analysis in a clinical setting, there are some details to consider that merit another validation by comparing both use cases.Looking at “Knowledge & Data Input”, there is no fundamental technical difference between the raw sequencing data used in our example compared to the examples in Krause et al. [24] using rumen samples. There might be different supplementary data like medical history or specific databases targeted for human diseases that differentiate these use cases. As the conceptual architecture does not impose limits on the type of data to be used this is not a problem. Likewise, the roles of “Domain Expert” and “Diagnostic Expert” were chosen broadly so that they can be filled by experts in biology, medicine, or others as needed. The “AI Integration & Fusion” layer allows the integration of various data types, bridging their technical differences. This will allow different types of data by implementing suitable wrappers, thus supporting both use cases. The “Persistence” layer does not impose schema restrictions that would either impede clinical use cases.For the “Model & Configuration Input” layer there may be different legal requirements, like stronger data protection requirements, when comparing both use cases. As the architecture uses a policy model and the role of a “Governance Officer”, these legal requirements can be presented in implementations using the architecture, for example, by using access policies and auditing the use of data. The analysis steps and algorithms, like the use of QIIME 2, UCLUST, and RDP classifier in the human example are similar or identical to the tools used in rumen microbiome analysis. The specific tools used vary from study to study in both use cases, which is addressed in the model by using a configurable workflow engine and a registry allowing for different analysis and processing services to be used.The “End User Interface” and “AI Input/Output” layers are also very generic to support a wide range of use cases. For rumen microbiome analysis the use of dialog systems or multilingual reports might not be necessary if the role of “End User” is filled by a researcher, who is also a “Domain Expert”. For clinical settings, the “End User” may actually be the patient, who wants an easily understandable report of his diagnosis. As it is, the conceptual architecture does support both use cases.This initial validation shows that all layers are generic enough to support both use cases. It can thus be assumed that the conceptual architecture as a whole also supports both use cases.The increased use of sequencing technology in the biomedical field and the voluminous quantity of associated data provide several challenges for data processing and opportunities for leveraging advanced ML techniques. This paper provides an extensive overview of both metagenomics and ML in the hope of further bridging the gap between these two disciplines.With the same goal, several existing use cases for ML in biomedical workflows have been documented. The use case of using phylogenetic information to improve phenotype classification was discussed in detail. Integrating phylogeny directly at the level of the ML model, instead of at data preparation and modeling level, has the potential to provide new microbial features for classifiers such as deep learning neural networks and to improve overall phenotype classification performance.Possible reasons for the perceived lack of advanced ML techniques like deep learning have been identified, like the lack of sizable training data, accessibility, and general challenges related to big data. Automatization pipelines can help to address some of the challenges associated with big data processing, metagenomics, and ML by improving reproducibility, accessibility, and dividing the work into smaller processable units. With this in mind, they also provide a foundation for the use of advanced, computationally-expensive ML algorithms like deep learning neural networks.The AI2VIS4BigData Conceptual Architecture for Metagenomics was given as a solution to integrate automated biomedical pipelines while taking into account additional research aspects such as data ingestion, management, archiving, and visualization. It was built with advanced, machine learning-based analysis methods in mind and supports large quantities of data using a distributed and modular architecture. Taken together, these features improve the accessibility of machine learning-based analysis methods and the handling of big data. It was demonstrated that the architecture is suitable for studies in human metagenomics such as the one discussed in this paper, therefore providing an initial validation for the architecture that was previously only validated for use in rumen microbiome analysis.Future research could address the remaining challenges of advanced ML methods like the quality and size of training corpora, problems arising from the biological diversity of samples, the need for explainability, and further advancements to increase the accessibility of these methods for scientists without an ML background. The AI2VIS4BigData conceptual architecture could be further validated by evaluating additional use cases, adding a technical architecture, and finally providing an implementation.Conceptualization, T.K. and M.H.; investigation, T.K.; writing—original draft preparation, T.K. and J.T.W.; writing—review and editing, P.M.K., H.W., H.Z. and M.H.; visualization, T.K. and J.T.W.; supervision, H.W., H.Z. and M.H.; project administration, T.K. All authors have read and agreed to the published version of the manuscript.This research received no external funding.The authors declare no conflict of interest.Typical processing steps for amplicon and shotgun sequencing. Most studies will use only a subset of these steps. Steps are grouped into five phases.Study performed by Wassan et al. [29]. Abundance information is combined with phylogenetic information to improve classification performance.Linear separation of samples in a two-dimensional vector space using SVM. Dots are colored according to their correct classification. The line represents the hyperplane that determines the classifier result.Consecutive splitting of dimensions in a decision tree. Colored points represent samples belonging to two classes, lines represent the binary decisions taken at the nodes of the tree.Naïve Bayes classifier with decision boundary. Mean and SD for both classes is shown as two crosses. The curve represents the decision boundary resulting from the probabilities.Classification using logistic regression function. The line shows the decision boundary where the sigmoid function is equal to 0.5.Neural network classification indicating the decision boundaries using distinct background colors. The network is possibly overfitted as the complex boundaries are unlikely to represent an underlying truth in the data.Sunburst diagram created with Krona. The rings represent different levels in the phylogenetic tree.Phylogram created with Megan. The phylogenetic hierarchy is visualized as a tree. The individual nodes are colored according to their relative abundances.A conceptual architecture for AI and big data supporting metagenomics research. Architecture is split into three pillars with each pillar having three layers. Diagram from Reis et al. [98].Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00011.txt

@@ -0,0 +1,5 @@
+Electronic health records (EHRs) can be very difficult to analyze since they usually contain many missing values. To build an efficient predictive model, a complete dataset is necessary. An EHR usually contains high-dimensional longitudinal time series data. Most commonly used imputation methods do not consider the importance of temporal information embedded in EHR data. Besides, most time-dependent neural networks such as recurrent neural networks (RNNs) inherently consider the time steps to be equal, which in many cases, is not appropriate. This study presents a method using the gated recurrent unit (GRU), neural ordinary differential equations (ODEs), and Bayesian estimation to incorporate the temporal information and impute sporadically observed time series measurements in high-dimensional EHR data.One of the biggest challenges to work with electronic health record (EHR) data is that there are many missing values. This issue incorporates uncertainty in the predictive model if the missing instances are imputed. Common imputation methods usually do not consider the temporal information, which is crucial for time series analysis. Moreover, most time series analysis methods ignore the time gap between measurements or assume that the time gaps are equal. In this study, we investigated time series imputation with irregular time gaps and propose a method based on neural ordinary differential equations (ODEs), recurrent neural networks (RNNs), and Bayesian estimation. This method offers a robust imputation of sporadically sampled multivariate time series measurements obtained from different patients.Many data imputation techniques have been developed over the years. In real-life problems, it is very common to have multiple missing attributes for a particular dataset. In the literature, most datasets have 30% to 50% missing values, and they have been imputed using various techniques [1]. There are mostly two techniques widely used for data imputation. These are statistical techniques and machine-learning-based techniques. Among the statistical techniques, expectation minimization (EM), the Gaussian mixture model (GMM), Markov chain Monte Carlo (MCMC), naive Bayes (NB), principal component analysis (PCA), etc., have been used frequently [1]. Among the machine-learning-based techniques, the Gaussian process for machine learning (GPML) (see [2]), support vector machines (SVMs) [3,4], k-nearest neighbors (k-NNs) [5], decision trees (DTs) [6], and artificial neural networks (ANNs) [7] have been heavily used in the literature.In recent years, longitudinal data imputation has been necessary specially in EHRs. However, many imputation methods only consider the data without the very important element–temporal information. However, there are many time series imputation methods that only consider equal time steps. Our research focuses on a time series imputation method that can deal with sporadically observed time series measurements obtained from EHRs. The major key components to implement this imputation method are neural ODEs [8], which parameterize the derivative of a neural network’s hidden state. As compared to the popular residual neural networks, neural ODEs have superior memory and parameter efficiency. Neural ODEs can easily deal with continuous time series, unlike recurrent neural networks, which require discretization. In a follow-up study, latent ODEs were proposed for irregularly sampled (e.g., sporadic) time series [9]. This method (called ODE-RNN) is presented as an alternative to autoregressive models. However, neural ODEs [8] use RNNs as the recognition network to estimate the posterior probabilities. However, that approach is more appropriate for continuous time series modeling with regularly sampled data. Therefore, the ODE-RNN [9] has been introduced as the recognition network to deal with irregularly sampled continuous time series analysis. Combined with neural ODEs and the GRU, a Bayesian update [10] is proposed that uses a predictive (with observation masking) method (called the GRU-ODE-Bayes method) to include only the available observations to update the predicted values along the multivariate time series. However, this method does not impute the missing values; rather, it performs zero- or mean-value padding. The GRU-ODE-Bayes method assumes that the observations are sampled from a multidimensional stochastic process whose dynamics can be explained by a Weiner process. The examples include the stochastic Brusselator process [11], the double-Ornstein–Uhlenbeck (OU) stochastic differential equations [12], etc. The authors showed that their method achieved better results than the GRU-D [13,14], minimal gated unit or minimal GRU [15], and other popular methods. In another recent study [16], a bidirectional recurrent imputation for time series (BRITS) was proposed. This algorithm uses both a forward and backward feeding of inputs to the RNN and simultaneously imputes the missing values. However, the BRITS does not allow the stochastic imputation of time series data. The BRITS is composed of a recurrent component and a regression component for imputation. The authors also presented a unidirectional approach called the RITS and claimed that the process was slower compared to the BRITS. Among other RNN-based models, the multidirectional RNN (M-RNN) provides good imputation results [17]. This study aims to develop a robust multivariate stochastic imputation technique for irregular time series that will fill this research gap.To develop a good predictive model, a reliable database is very crucial. The data need to be authentic and mostly accurate. Any big-data-driven research highly depends on the quality of the data being used. In this study, a very popular dataset [13,18,19,20] named the Medical Information Mart for Intensive Care (MIMIC) clinical database was explored and analyzed. This dataset contains intensive care unit (ICU) admission records of patients admitted to Beth Israel Deaconess Medical Center in Boston, Massachusetts, from 2001 to 2012. This database has several versions. In this study, MIMIC III Version 1.4 was used, which is the latest. It contains de-identified electronic medical records, demographic information, and billing information for ICU-admitted patients. Many of these records contain timestamps of clinical events, nurse-verified physiological measurements, routine vital signs, check-up information, etc. However, this database is only available after completing a required course and acknowledging a data use agreement.After accessing the electronic records from the MIMIC III clinical database, a clean dataset needs to be formed that can be useful for analysis. As most other databases, MIMIC III provides its users with different types of structured and unstructured data records that must be cleaned prior to performing any data-driven analysis. Furthermore, the electronic records sometimes have different artifacts associated with them. Data cleaning tends to be more difficult in the case of large databases such as MIMIC III. The search algorithms for the desired attributes need to designed in a way that they can extract the necessary information efficiently. There are hardware and software limitations due to which it becomes very difficult to carry out large matrix operations in traditional computers. Since the analysis highly depends on the data quality, a proper cleaning process should be selected and performed carefully. There might be anomalies in the chosen dataset that should be properly dealt with before any analysis.The MIMIC-III database is an information storehouse for critical care patients. Therefore, it needs to deal with proper care and privacy. To access the data, one needs to request access formally through the Physionet website (https://www.physionet.org) (last accessed on 15 November 2021). Two important steps need to be followed to access the data. The first one is to take a recognized course and comply with the Health Insurance Portability and Accountability Act (HIPAA) requirements. The second step is to sign a data use agreement that includes the appropriate data usage policy, security standards, and preventing identification efforts. Once the request is submitted, the approval comes within a week. Then, the data can be accessed from the server or can be stored in local storage. More information can be obtained by visiting the official MIMIC website (https://physionet.org/content/mimiciii/1.4/) (last accessed on 15 November 2021).There are 26 tables in total in the MIMIC-III database (see Appendix A, Table A1). In this subsection, we mainly discuss the tables that were directly used for the analysis. Since our goal was to extract as many CHF-related variables as possible, we needed to explore different tables and match records with mostly unique patient IDs and sometimes with admission IDs. Different tables are linked together to compile the dataset that we needed. The tables are described in the following paragraphs.The “Admissions” table provides unique hospital admission information for each patient. It reports the admission ID, ICUstay ID, date of admission, admission type, discharge location, diagnosis, insurance, language, religion, marital status, age, ethnicity, etc. This can be linked with other tables via the admission ID and patient ID.The “Patients” table provides demographic information for 46,520 unique patients. It contains the patient ID, admission ID, date of birth, date of death, gender, hospital expire flag, etc.The “Chartevents” table is the largest table in the entire MIMIC III database. This has about 330,712,483 records. This table provides the patient ID, admission ID, item ID, and the corresponding routine physiological measurements of each patient from time to time.The “D_ICD_diagnosis” table contains unique patient IDs, unique hospital admission IDs, and International Coding Definitions Version 9 (ICD-9) for each of the 14,567 diagnosis categories. The code for CHF diagnosis is 4280.The “D_items” table has 12,487 records of items used to treat different patients. Routine vital signs such as blood pressure, heart rate, white blood cell count (WBC), respiratory rate, and other numerical variables are listed here with distinct item IDs.Before any type of analysis, the subset of the entire database, that is of interest, needs to be extracted. Data cleaning can be a very tedious process, especially in the case of such huge clinical databases as MIMIC III. Figure 1 shows the data-cleaning steps. The data-cleaning steps are quite challenging at times. At first, all patient records (56,320) were assessed. These records mostly contain the demographic information such as age, gender, religion, etc. There is some hospital-related information available as well, such as admission type, discharge location, length of stay, etc. Since this study focuses on CHF (and some related diagnoses)-diagnosed patients only, those were extracted using the ICD-9 diagnosis codes. This totaled 13,295 patients. Among these patients, many of them had been readmitted multiple times with the maximum number of readmissions as high as 13. If a patient is never readmitted or followed up, we kept their admission record as is. In the case of multiple readmitted patients, we only considered their first time readmission record and discarded the subsequent ones. Then, all the numerical and categorical features were joined to the patient list (total 10,027 patients) according to their unique ICU stay IDs.As mentioned earlier, mostly CHF patients were considered for this study. Along with CHF, the following diagnoses were also considered, as shown in Table 1.EHRs contain time series measurements of different patients. It is sometimes very difficult to understand the contributing features of a certain outcome. Table 2 shows the chosen predictor numerical variables [21] for this analysis. The variables were selected based on multiple studies [22,23,24] and their importance to CHF readmission prediction.This section presents the filtering criteria and the personalized time series extraction process from the MIMIC-III v1.4 database. As mentioned in earlier sections, there were about 10,000 unique patients having CHF and other related diagnoses. For each of these patients, a unique time series was extracted that contained different types of measurements obtained for different items (e.g., heart rate, glucose, BUN, etc.). Figure 2 shows a sample patient with different item measurements taken at irregular intervals along the horizontal axis. Although it shows measurements beyond 48 h from discharge, this study only considered measurements up to 48 h from discharge. It is important to note that all the item measurements might not be available for all patients. Therefore, the time series imputation became more challenging due to the lack of data.First, the patients were sorted using their unique ICU stay IDs. This ID distinguishes every single ICU stay of a patient. There were some patients who had been readmitted to the ICU more than once during the same hospital admission. In these cases, they usually had a the same admission ID, but different ICU stay IDs. That is why we chose to identify patients by their ICU stay IDs. However, their subject IDs and admission IDs are also stored in the patient database. A search algorithm was deployed to find each patient’s measurement during his/her unique ICU stay.This section describes and implements the multivariate irregularly sampled time series imputation method that was based on neural ODEs, GRU, LSTM, and Bayesian estimation. It is important to discuss the useful technical details of this method so that it can be explained easily. The flowchart and mathematical notations are used for the explanation as necessary. are applicable here, as well. In the following sections, the imputation problem is introduced, and the methods for overcoming this challenge are discussed.Any EHR database contains numerous vital measurement information for ICU patients. Due to many physiological factors, these measurements are not taken at the same time intervals. For predictive modeling and other data scientific procedures, regular intervals are usually expected. Therefore, EHR data can be challenging due to these irregular measurements.The problem is to develop an imputation method that can perform two challenging tasks: capture the temporal information from irregularly sampled time series measurements and impute time series with high missing ratios. However, the standard imputation methods hardly consider temporal information and are mostly suitable for regular time series. This section describes the technical and mathematical details for the multivariate imputation method. The important components of the method—neural the ODE, GRU and Bayesian estimation—are discussed sequentially. To summarize the steps, the algorithm is presented in a compact version that is easier to understand. From this point, the proposed method is called the GRU LSTM ODE BAYES Imputation (GOBI) method.Recurrent neural networks are a special type of artificial neural network that are able to exhibit temporal dynamic behavior in sequence data. They can process temporal information from the current state to the next state using hidden layers. However, the conventional RNN suffers from the vanishing gradient problem. This means that the weights of the neural network are more difficult to train further down the sequence because the loss function tends to be very close to zero. To avoid this problem, mostly two types of gated RNNs are used—long short-term memory (LSTM) and the gated recurrent unit (GRU). These two special types of RNNs were used in this study. Defining X as the input and ht and ht+1 as the previous and current hidden layers, the simple structure of an RNN layer is shown in Figure 3.Neural ODEs [8] are useful for continuous-depth neural networks. This method involves performing a reverse-mode differentiation technique (e.g., backpropagation), which is quite difficult to train. While solving the ODEs, the adjoint sensitivity method is used. This approach usually takes less time and calculation effort. A scalar-valued loss function L [8] as described in Equation (1) is minimized, whose input comes from the ODE solver.
+        (1)Lht1=Lht0+∫t0t1fht,t,θdtHere,h(t1) = current hidden state;h(t0) = initial hidden state;t0 = initial time;t1 = current time;θ = weight of neurons at synapses;f = hidden unit dynamics function.The loss function L incorporates all the time points (e.g., states) in the time series. The adjoint states help to calculate the gradients with respect to θ, as needed along the time sequence. Defining the adjoint a(t)=δLδt and taking the derivative yield [8],
+        (2)datdt=−atTdfht,t,θdhNote that Equation (2) allows the computation of gradients along the time sequence. However, the gradient of L with respect to θ is calculated by propagating backwards. This yields (See [8])),
+        (3)dLdθ=∫t1t0atTdfht,t,θdθdtTherefore, the above two equations can be efficiently calculated by any regular auto-differentiation packages.If the observations yi are sampled from the realizations of a D-dimensional stochastic process Y(t), the internal dynamics can be expressed as an unknown stochastic differential equation (SDE) as the following:(4)dYt=μYtdt+σYtdWtHere,dW(t) = Weiner process;μ = mean of the probability density function of Y(t);σ = covariance of probability density function of Y(t).Then, the distribution of Y(t) evolves according to the Fokker–Planck equation [10].The GRU is a type of RNN that requires less time and calculation effort than its counterpart LSTM. However, there is no clear indication of the superiority of their performance on real-world datasets. As reported by many authors [13,15,17], the GRU and LSTM usually perform head to head with a slight edge over each other on different datasets and in different configurations. Figure 4 shows a standard GRU cell configuration. There are two main gates in a GRU: reset gate and update gate. This configuration makes the GRU very simple to train and take less time to compute all the parameters.As seen in Figure 4, a standard GRU cell contains the following elements [10]:(5)rt=σWrxt+Urht−1+br(6)zt=σWzxt+Uzht−1+bz(7)gt=tanhWhxt+Uhrt⊙ht−1+bhHere, rt, zt, and gt denote the reset, update, and forget gate, respectively. Furthermore, ⊙ corresponds to the elementwise product. Two matrices W∈IRH×D and b∈IRH×H denoting the weight and bias vectors are the cell parameters. H and D are the dimensions of the hidden process and given inputs. Therefore, the hidden state, h, of the GRU can be updated as follows [10]:(8)ht=zt⊙ht−1+(1−zt)⊙gtIn order to construct a first-order differential equation similar to Equation (1), the following can be obtained from Equation (8):(9)Δht=ht−ht−1=zt⊙ht−1+1−zt⊙gt−ht−1 = 1−zt⊙gt−ht−1Taking the differential with respect to t, the following can be obtained readily: (10)dhtdt = 1−zt⊙gt−htEquation (10) can be solved using any regular first-order ODE solvers such as Euler, midpoint, Dormand–Prince (Dopri), etc.Bayesian estimation allows the updating of prediction after the model is run on GRU cells. The most important feature of Bayesian estimation is the ability to incorporate new information as it becomes available. In this imputation method, only the available values can be used as a source of information since the missing values play no part. After the initial estimation from the GRU cells, a Bayesian estimation is necessary to include the information from observed values. This helps reduce the gaps between the observed and estimated values, which, in turn, improves the imputation performance.In the original version, the model in [10] uses an integrated GRU cell to incorporate the Bayesian update with observed values. However, in this study, we used an LSTM cell to perform the Bayesian update since it provides a more accurate estimation [25]. However, the missing values are not updated since they do not have any effect on the hidden layers. The proposed hidden layer can be described as follows:(11)ht+ = LSTMht−,fprepyk,ht−Here,h(t+) = hidden state after Bayesian update;h(t−) = hidden state before Bayesian update;fprep = perception layer;y = observations;k = observation mask.Figure 5 shows the Bayesian estimation effects on the prediction.Two loss functions are used to evaluate the updating performance as shown in Figure 5. The first function is the negative log-likelihood, described as follows:(12)Losspre=−∑j=1Dmjlogpyj|θ=fobsh−jHere,D = number of samples;m = mask;yj = current sample;fobs = observed hidden layers.The second function is the Kullback–Leibler (KL) divergence, which basically compares two probability distributions.
+        (13)Losspost=∑j=1DmjDKLpBayes,j||ppost,jHere,DKL = KL divergence;pBayes,j = probability distribution after update;ppost,j = posterior distribution of observations.The imputed mean and variances are calculated using the following equations:(14)μBayes=σ2σpre2+σobs2μpre+σpre2σpre2+σobs2μobs(15)σBayes2=σpre2·σobs2σpre2+σobs2Since, the algorithm has many parameters and steps, it might be difficult to understand sometimes. Therefore, the algorithm sequences are outlined [10] in Figure 6 (λ = trade-off parameter between post- and pre-losses) below.The experiments were designed according to the needs and objectives of this study. As mentioned before, the missing ratio is very high in most of the numeric features. This scenario is not suitable for any ML-based techniques since they require complete datasets with much information available prior to training.The training and testing ratio was 9:1, which means 90% of the samples were used for training and validation, while the rest were used for testing. To validate the results, 10-times 10-fold cross-validation was used. The following hyperparameters, shown in Table 3, were used to conduct all the experiments.Since this is an imputation task, the performance measures were the mean-squared error and mean absolute error. The corresponding formulae are described in the next section. All the experiments were performed on a workstation with the specifications as shown in Table 4.The following results were obtained from the five folds of the dataset with 1345 samples each. In the following Figure 7, Figure 8 and Figure 9 the solid lines represent the predicted values, the solid dots represent the observed values, and the dashed lines represent the confidence bounds.The proposed method was compared with some baseline methods and RNNs. The results for both training and testing are shown in Table 5 and Table 6, respectively. It is clearly seen that the GOBI method works better than most baseline and RNN-based methods. Two metrics are reported for comparison: root-mean-squared error (RMSE) and mean absolute error (MAE), as described by the following two equations:(16)RMSE=1N∑i=1Nyi−y^i2(17)MAE=1N∑i=1Nyi−y^iHere,yi = true value for instance i;yi^ = imputed value for instance i;N = number of instances.The proposed model was based on neural ODEs, RNN units, and Bayesian estimation and is suitable for imputation tasks involving temporal information. In most real-life scenarios, temporal information is very sensitive and determines many aspects of our day to day lives. Therefore, it is not always right to ignore the time-sensitive information found in EHRs or other similar databases. Furthermore, it makes more sense to have a probabilistic imputation rather than a deterministic one since there is always some level of uncertainty associated with the imputed data.The novel contribution of this model is that it is tailored specifically for multivariate irregularly sampled time series imputation. As mentioned earlier, most other imputation methods deal with regular time series and provide deterministic imputation. Both of these issues are addressed by the GOBI method in this study.The performance of the GOBI method was satisfactory, as shown in the comparative analysis section. Many state-of-the-art methods have been developed for classification tasks, but RNN-inspired stochastic imputation is still a growing area of data scientific research. In this study, only EHR data in the MIMIC databases were used for imputation. However, the GOBI method works well for many other datasets, as well, and provides fairly good estimation [10]. As seen in Table 6, the GOBI model has greater accuracy and lower variance.GOBI method has several advantages, as well. It takes less time to train since GRU cells are simpler to compute and have fewer parameters. Even with very large datasets, it works relatively faster than most RNN-based techniques [15,26]. Besides, the GOBI method is quite accurate as compared to many algorithms currently available. Although the performance might vary from dataset to dataset, still it should be fairly competitive overall.The GOBI method has high potential in data science sectors. It should be very useful for analyzing large datasets with a high amount of missing values. It might have broad impacts in healthcare, manufacturing industries, process improvement, etc., because most of these sectors typically deal with missing data or have physical constraints for time-sensitive data collection. The GOBI method can deal with these types of problems, providing a good solution with an acceptable error margin.The imputation algorithm presented here is of great importance due to the increasing number of missing values in EHRs. It is very common to have more than 50–60% missing values per channel in EHR time series data. Each patient has some set of demographics, which vary greatly from one patient to the other. Therefore, it is difficult to impute records of one patient based on another. There is hardly any way around this since some patients might not have any measurement for a particular channel or item. As mentioned earlier, the EHR time series data might have some underlying dynamics that might not be approximated well by the stochastic Weiner process. This opens up a great opportunity for further research. As for standard oscillating behaviors, there are many well-established equations to represent the internal dynamics. As shown in many recent articles, the simulated data for standard oscillations can be accurately estimated by the Weiner process. However, this might not be the case for most real-life EHR time series data. The proposed method not only imputes the time series data, but also provides an estimation of the level of uncertainty that the imputed values represent. In the proposed model, the Bayesian estimation is coupled with an LSTM cell. This allows for the update to happen only when there are available values. The missing values are then inferred from the resulting imputed time series. The target is to predict the mean imputed values as close as possible to the actual values. The log variance needs to be smaller as well, since higher levels of uncertainty are much more difficult to propagate and can easily jeopardize the entire predictive model. In most common imputation methods, the mean-squared error is minimized. The proposed method uses two losses (pre and post) before and after Bayesian estimation, which are useful to decide whether the new observational update reduces the error. The two loss functions (negative log likelihood and KL divergence) are well established and widely used in ML techniques. Considering these issues, further research can be concentrated on developing an imputation method that could efficiently learn the dynamics of the underlying process instead of assuming a Weiner process in every case. This would significantly boost performance in complex EHR data.Conceptualization, M.A.U.Z. and D.D.; methodology, M.A.U.Z.; software, M.A.U.Z.; validation, M.A.U.Z., D.D.; formal analysis, M.A.U.Z.; writing—original draft preparation, M.A.U.Z.; writing—review and editing, M.A.U.Z.; visualization, M.A.U.Z.; supervision, D.D. All authors have read and agreed to the published version of the manuscript.This research received no external funding.Not applicable.Not applicable.The data will be available only after completing the requirements from the MIMIC website (https://physionet.org/content/mimiciii/1.4/ (last accessed on 15 November 2021)).The authors declare no conflict of interest.All tables of MIMIC III and their total records.Data-cleaning steps.A sample patient time series showing heart rate and arterial blood pressure (diastolic and systolic) measurements.A simple RNN layer.Configuration of a GRU cell.Bayesian jump during the update of hidden layers.GOBI algorithm sequences.Imputation results (heart rate).Imputationresults (systolic blood pressure).Imputationresults (diastolic blood pressure).All the diagnoses considered for patient cohort selection.All numeric variables and their missing ratios.Hyperparameters used and their values for the GOBI method.Computer infrastructure.Performance comparison of the imputation methods (training).Performance comparison of the imputation methods (testing).Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00012.txt

@@ -0,0 +1,2 @@
+According to the WHO, approximately 50 million people worldwide have dementia and there are nearly 10 million new cases every year. Alzheimer’s disease is the most common form of dementia and may contribute to 60–70% of cases. It has been proved that early diagnosis is key to promoting early and optimal management. However, the early stage of dementia is often overlooked and patients are typically diagnosed when the disease progresses to a more advanced stage. The objective of this contribution is to predict Alzheimer’s early stages, not only dementia itself. To carry out this objective, different types of SVM and CNN machine learning classifiers will be used, as well as two different feature selection algorithms: PCA and mRMR. The different experiments and their performance are compared when classifying patients from MRI images. The newness of the experiments conducted in this research includes the wide range of stages that we aim to predict, the processing of all the available information simultaneously and the Segmentation routine implemented in SPM12 for preprocessing. We will make use of multiple slices and consider different parts of the brain to give a more accurate response. Overall, excellent results have been obtained, reaching a maximum F1 score of 0.9979 from the SVM and PCA classifier.Wavelet transform [1] and its applications are a powerful tool in image processing. Since the rise of wavelet analysis in the early 1980s, it has been shown to be useful in many fields of applied mathematics, specifically in signal and image processing.The recent advances in neural networks and transfer learning methods [2] have opened the door to discussions about the goodness and supremacy of neural networks over the traditional transforms in pattern recognition and classification.At the same time, Alzheimer’s disease has established itself as one of the great epidemics of the 21st century, being the most common case of dementia. It has been proved to present intermediate stages before the severity of the dementia becomes moderate or serious. These stages can be noticeable and, on occasion, patients are even aware of their own cognitive impairment. However, such symptoms can go unnoticed until the disease presents an advanced stage. Detecting and diagnosing early stages can be the key to provide patients with better prognosis and preventive treatments in order to improve their quality of life.The latest experiments were able to find a correlation between the information contained in brain MRI images and the presence of dementia. These experiments made use of a wide variety of methods and algorithms. The first investigations were based mainly on traditional transforms [3] and VBM plus DARTEL analysis [4] techniques for feature extraction and state of the art Support Vector Machine (SVM) classifiers. Generally, these experiments obtained good results (92.36% and 96.32% accuracy respectively) classifying binary (Alzheimerś disease (AD) and control (CN)) stages.In the last five years, more stages of the disease, including mild cognitive impairment (MCI), have been included in the classification and neural networks have gradually assumed greater importance. These classifiers normally outperformed previous accuracies, as stated in [5] (97.50%), [6] (97.51%), or [7] (99.1%). By combining them with other techniques such as latent transition analysis [8] it is possible to predict status changes in Alzheimer’s disease. For this investigation, the line of research proposed by [9,10] will be followed. In particular, we will make use of the Segmentation routine implemented in SPM12 [11,12] for preprocessing and the most relevant volumes of the brain related to Alzheimer’s disease discovered in [10].This research will thoroughly analyze the full content of the MRI images and contrast them again the afore-mentioned experiments, aiming to predict up to six different stages of dementia. The objective is to use the nine best slices proposed in [10] and simultaneously apply two different approaches as suggested in [13,14]: two-dimensional multiresolution analysis (2-D MRA) in L2(R2), aligned with SVM and convolutional neural network (CNN) [15,16].Following, also, the strategy used in [10,17], the mRMR algorithm will be used to extract the coefficients to feed the SVM classifier. Subsequently, its performance will be contrasted with the traditional PCA algorithm.Overall, this research aims to answer to several questions. First, whether it is possible to predict and prevent the appearance of the Alzheimer’s disease, considering up to six stages of dementia: cognitively normal, significant memory concern, early mild cognitive impairment, mild cognitive impairment, late mild cognitive impairment, and Alzheimer’s disease. Secondly, which of the aforementioned algorithms performs better after the application of SMP12 preprocessing? Thirdly, what is the result of processing multiple slices in conjunction instead of using single slices for classification? Lastly, how to address the problem of memory management when using larger volumes of information for model training.One of the key points is to decide which of the following algorithms performs better when detecting Alzheimer’s disease from MRI images: SVM or CNN. An additional experiment will be conducted using transfer learning based on VGG16, a convolutional neural network model proposed by K. Simonyan and A. Zisserman, from the University of Oxford, in the paper “Very Deep Convolutional Networks for Large-Scale Image Recognition”. Similar transfer learning approaches were suggested in [6]. Another key point is to learn the best strategy to use in terms of data handling. By example, we will address the question of whether it is better to process a single slice at a higher level of detail or to process, simultaneously, more than one slice with at lower level thereof. Additionally, we need to determine the classes that are of most importance. It is evident that the sooner the disease is detected, the better the prognosis will be for the patient. It will be also relevant for us to test the model’s performance in detecting Alzheimer’s in its early stages. Restrictions apply to the availability of this data. Data was obtained from http://adni.loni.usc.edu (accessed on 20 May 2020), and is available with the permission of The Alzheimer’s Disease Neuroimaging Initiative (ADNI). The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). The complete database of ADNI including ADNI1, ADNI-GO, ADNI2 and ADNI3, are used for this investigation. The set downloaded comprises brain images in NIfTI (Neuroimaging Informatics Technology Initiative) format. Corrupt and duplicated images were deleted, resulting in a total of 6028 T1-weighted MRI images. The methodology described in [9,10] is used for the pre-processing of the information from the sample data. This pre-processing, which requires segmentation, bias correction and spatially normalisation, has been performed using the Segmentation routine implemented in SPM12 [11,12]. The number of samples in the data constitutes a considerable improvement to [9,10], where 400 and 242 images were obtained, respectively. The full project and code are available on https://github.com/juliopradom/alzheimer-classifier (accessed on 6 December 2021).As mentioned before, the set of images provided are classified into six groups:AD (Alzheimer Disease): the images in this group correspond to patients diagnosed with Alzheimer.CN (Cognitively Normal): corresponds to healthy individuals (control).MCI (Mild Cognitive Impairment): causes a slight but noticeable and measurable decline in cognitive abilities.EMCI (Early Mild Cognitive Impairment): an early stage of MCI with milder episodic memory impairment.LMCI (Late Mild Cognitive Impairment): a more advanced stage of MCI previous to AD.SMC (Significant Memory Concern): patients with SMC are characterized by self-report significant memory concern, quantified by using the Cognitive Change Index and the Clinical Dementia Rating (CDR) of zero. SMC participants score within the normal range for cognition, and the informant does not equate the expressed concern with progressive memory impairment. SMCs have been shown to be correlated with a higher likelihood of progression [18].AD (Alzheimer Disease): the images in this group correspond to patients diagnosed with Alzheimer.CN (Cognitively Normal): corresponds to healthy individuals (control).MCI (Mild Cognitive Impairment): causes a slight but noticeable and measurable decline in cognitive abilities.EMCI (Early Mild Cognitive Impairment): an early stage of MCI with milder episodic memory impairment.LMCI (Late Mild Cognitive Impairment): a more advanced stage of MCI previous to AD.SMC (Significant Memory Concern): patients with SMC are characterized by self-report significant memory concern, quantified by using the Cognitive Change Index and the Clinical Dementia Rating (CDR) of zero. SMC participants score within the normal range for cognition, and the informant does not equate the expressed concern with progressive memory impairment. SMCs have been shown to be correlated with a higher likelihood of progression [18].With these classes, we can build a progression diagram for Alzheimer’s disease, as shown in Figure 1.Unlike other experiments and investigations, by achieving the goals set for this classifier we will not only discern between healthy patients and patients with dementia but also provide a more accurate diagnosis about the stage of their dementia. This may be useful to detect the disease earlier and prescribe a more effective and personalized treatment to those individuals situated in intermediate stages.By looking at Figure 2 we can observe the number of samples per category. LMCI, MCI, AD and CN have approximately the same number of samples (∼1200) whereas SMC and EMCI are clearly unbalanced. Particular attention is paid to these minority classes given that, as commented upon at the beginning of this chapter, our specific focus is on Alzheimer’s early stages.Each of the samples is composed of nine different images corresponding to nine different slices from an MRI image. Slices 55, 56, 61, 72, 82, 90, 104, 106 and 114 are used, according to the most promising slices obtained in [10], after the application of multi-objective genetic algorithms to find most relevant volumes of the brain related to Alzheimer’s disease and mild cognitive impairment.The images have three different colour channels (RGB) and are already resized to 656×875. As a result, every image is an array with shape (656,875,3). In Figure 3 shows the nine images for patient 1 in AD.In practice, the management of all this information is very resource-consuming. This will be a limiting factor throughout our investigation. Smart approaches must be taken to overcome these obstacles and will be covered in depth in the following sections.Having analysed and described our data, two lines of research are going to be opened. First, we are going to proceed to extract and select features with which to apply SVM. Secondly, we will use a custom CNN and CNN with transfer learning to train our model without further preprocessing.Global F1 (macro) is the measurement that we will use to compare the performance of the different algorithms. However, it is essential to remember that another relevant indicator that we will use in practice, when facing similar scores, is the specific performance (F1) of the algorithms identifying SMC and EMCI.Building a support vector machine classifier using wavelet coefficients extracted from our images is the first approach. By using wavelet coefficients, the identification of relevant anomalies and slight changes within our samples will be easily detected, which is, actually, what we are seeking to achieve in distinguishing between one class or another.The coefficients extracted for a given patient can be used to construct a vector in a high-dimensional space. The set of these vectors labelled with their corresponding classes will help determine a hyperplane able to separate each group from the other. In practice, one hyperplane will be created for every class. This research is carried out using the scikit-learn library.For our experiment, only the wavelet coefficients of the approximation image at level three and four are accessed for memory reason. It is mandatory to first convert the images from RGB space to grey so the wavelet module used can extract the coefficients using a single channel. In particular, we will perform ten extractions; one for every slice available and one combining all the coefficients consecutively. Extracting one slice means obtaining the wavelet coefficients at the third level for every patient at this specific slice, which gives a total of 6028 arrays of 10,120 coefficients each. Extracting all the coefficients means obtaining the wavelet coefficients at the fourth level for all the available slices and combine them, which, again, gives a total of 6028 arrays of 9 × 2867 = 25,803 coefficients each. In Table 1 are shown some of the coefficients we could extract per image.Nevertheless, as previously mentioned, going beyond the third level becomes computationally difficult for the machine used in our research. For such purpose, mini-batch training techniques must be used, which is beyond the objectives of this investigation. Sensitive information loss could be a possible consequence to this limitation, which could result in a worse F1 performance. After the coefficients are successfully extracted, it is important to note that we have arrays of 10,120 and 25,803 coefficients, respectively. In order to generate an input small and meaningful enough to train and work with our models it is necessary to apply feature selection. For this objective, we will perform two different experiments using two methods: mRMR (as used in [7,10,19]) and the traditional PCA.mRMR is an iterative algorithm that starts with an empty set S, and for each iteration searches for an input feature p that maximizes the mutual information (MI) [20,21] with respect to the output feature. After this, it minimizes MI with respect to the rest of input features, that is, maximizes the relevance and minimizes the redundancy. This feature is added to S. The algorithm finishes when all features are added. Moreover, the final ranking is the order in which they have been added to S. On the other hand, principal component analysis (PCA) [22] is another technique for reducing the number of variables in the sample data. However, unlike mRMR, this algorithm does not preserve the original features, but it applies dimensionality reduction. The objective of PCA is to transform the original random vector into variables called principal components. These components are all orthogonal and ordered so that the first few explain most of the variation of the random vector. PCA is a widely used algorithm for feature selection. Given the interest and good results [7,10,19] provided by mRMR, this research finds it appropriate to compare how both techniques affect the outcome of the classification.For the first experiment, mRMR is used to select the best 100 coefficients. This algorithm is extremely resource consuming and it is not possible to pass the whole set of arrays at once. In order to overcome this issue, a recursive version of this algorithm was designed and built. Both recursive and original versions have been tested on a small set of 100 features, obtaining a similar set of features, as expected (maximum relevance and minimum redundancy principle). The functioning of this recursive version is shown in Figure 4.It is worth noting that the application of mRMR will give a set of existent features as a result. Before starting the process, the data is standardized and rescaled in order for the mean of the observed values to be 0 and the standard deviation to be 1. This can be particularly practical in avoiding different scales between our images and to take a step further to ensure normalization.The same process is repeated, but using PCA this time. Contrary to mRMR, using PCA we obtain 100 new features as a result of dimensionality reduction; that is, we keep the information of all the original features in a smaller set. No particular considerations needs to be taken to perform PCA as the module used is powerful enough to be utilized with an input of our size.The next step is to find the best hyperparameters with which to tune our SVM model. After applying a grid search using slice 55 on the set of hypermarameters, as shown in Table 2, considering that C is the regularization parameter and gamma is the kernel coefficient, the most promising hyperparameters using the output of mRMR are kernel=rbf, C=1000 and gamma=10−2. Correspondingly, using the output of PCA we obtain the parameters kernel=rbf, C=10 and gamma=scale.Following [23], when creating the SVM, the strategy to use will be one-against-one since our kernel is based on a radial basis function (rbf). In all the experiments conducted the set of patients is split into training and test sets. These sets will remain unchanged to contrast the results more effectively. Specifically, the training set covers 75% (4521) of the available samples and test set covers the rest 25% (1507), where the number of representatives for each class follows the original distribution in both sets. In addition, the number 104,729 is used as a random seed for shuffling. The pre-processing and classification pipeline is summarized in Figure 5.The second part in our investigation is to train a convolutional neural network using, directly, the images available as input. Along with a custom CNN, we will also test VGG16, modifying the input and output layers. For both experiments, the number of epochs is set to 500 and the images are resized to 224×224 according to VGG16’s requirements. The objective is to make both experiments as similar as possible.By using CNN, the pre-processing steps covered in SVM can be skipped, since CNN is able to learn features from the image itself. Additionally, we can keep the RGB channels to make our input a tensor of order four. In order to be able to compare the performance of CNN with SVM, the same training and test set are considered. To construct our networks TensorFlow is used on a cluster containing two GPUs NVIDIA RTX 2080.To build our CNN, we perform two experiments with two common and effective structures consisting of the following sequence of layers:Experiment 1:1.One convolutional layer with 32 filters, a kernel size of (3,3) and using “relu” as its activation function. The number of filters determines the number of kernels to convolve with the input volume. Each of these operations produces a 2D activation map. Layers early in the network architecture (i.e., closer to the actual input image) usually learn fewer convolutional filters while layers deeper in the network (i.e., closer to the output predictions) will learn more filters. This layer will expect an input of size (224, 224, 3) in order to fit our images.2.One max pooling layer (2,2) to reduce the spatial dimensions of the output volume.3.Another similar convolutional layer, although using 64 filters this time.4.Another max pooling layer (2,2) to reduce the spatial dimensions of the output volume in the second convolutional layer.5.One flattened layer to connect the multidimensional data from convolution to dense layers.6.Two dense layers using the “relu” (64 units) and “softmax” (6 units) activation functions respectively. We use “softmax” to convert the scores to a normalized probability distribution.One convolutional layer with 32 filters, a kernel size of (3,3) and using “relu” as its activation function. The number of filters determines the number of kernels to convolve with the input volume. Each of these operations produces a 2D activation map. Layers early in the network architecture (i.e., closer to the actual input image) usually learn fewer convolutional filters while layers deeper in the network (i.e., closer to the output predictions) will learn more filters. This layer will expect an input of size (224, 224, 3) in order to fit our images.One max pooling layer (2,2) to reduce the spatial dimensions of the output volume.Another similar convolutional layer, although using 64 filters this time.Another max pooling layer (2,2) to reduce the spatial dimensions of the output volume in the second convolutional layer.One flattened layer to connect the multidimensional data from convolution to dense layers.Two dense layers using the “relu” (64 units) and “softmax” (6 units) activation functions respectively. We use “softmax” to convert the scores to a normalized probability distribution.Experiment 2:1.One convolutional layer with 16 filters, a kernel size of (3,3) and using “relu” as its activation function.2.One max pooling layer (2,2) to reduce the spatial dimensions of the output volume.3.Another similar convolutional layer, but using 32 filters this time.4.Another max pooling layer (2,2) to reduce the spatial dimensions of the output volume in the second convolutional layer5.A third convolutional layer but using 64 filters.6.A third max pooling layer (2,2) to reduce the spatial dimensions of the output volume in the third convolutional layer.7.One flatten layer to connect the multidimensional data from convolution to dense layers.8.Two dense layers using the “relu” (128 units) and “softmax” (6 units) activation functions respectively. We use “softmax” to convert the scores to a normalized probability distribution.One convolutional layer with 16 filters, a kernel size of (3,3) and using “relu” as its activation function.One max pooling layer (2,2) to reduce the spatial dimensions of the output volume.Another similar convolutional layer, but using 32 filters this time.Another max pooling layer (2,2) to reduce the spatial dimensions of the output volume in the second convolutional layerA third convolutional layer but using 64 filters.A third max pooling layer (2,2) to reduce the spatial dimensions of the output volume in the third convolutional layer.One flatten layer to connect the multidimensional data from convolution to dense layers.Two dense layers using the “relu” (128 units) and “softmax” (6 units) activation functions respectively. We use “softmax” to convert the scores to a normalized probability distribution.Both experiments will process 500 epochs. Replicating the experiments conducted in SVM, we build a model for every slice. Again, the training set covers 75% (4521) of the available samples and the test set covers the remaining 25% (1507), where the number of representatives for each class follows the original distribution in both sets. In addition, the number 104,729 is used as random seed for shuffling.Let us now begin analysing the results from the mRMR tests, in which our main objective is to maximize F1 score. Looking at Table 3, we can observe that the results obtained are decent. Classes in the middle are perfectly recognized and predicted, whereas control and sick patients present an observably worse score. The experiment with the best results is the one that used all the available coefficients at level four. This result is of interest since it shows that processing all the information as a whole is, in principle, more effective than using single slices for classification, even if the level of detail accessed is lower.In Figure 6 we can see more distinctly see the different scores between slices. After this experiment we have reached an F1 score of 0.9657. Regarding the models built using single slices, we can see that slice 82 slightly stands out from the rest. Below we will discuss if this trend is confirmed.The same experiment was reproduced using PCA, as shown in Table 4. We not that it is undeniable how well the classifier performed when using PCA for every slice. This time the slices with the best results were slices 104 and 106. As the method performed to extract the coefficients (PCA) transformed and resized all the available information, we could build the following hypothesis, considering, also, the conclusions from mRMR: slice 82 is the most informative slice when looking at specific regions of the image but slices 104 and 106 are more relevant when viewing the image as a whole. Again, the experiment with the best outcome continues to be the all-coefficients classifier, with an outstanding F1 score, in this case, of 0.9979.An extra experiment was conducted, also using PCA, and was noted as “Multi-Slice”. This test retrieves the output of every slice and selects the most common value, simulating a decision system between different experts, as shown in Figure 7. It is not a model in and of itself; rather, it applies the mode of the single-slice models.However, the all-slices classifier performs better, confirming the hypothesis opened after discussing the results from mRMR: processing all the information is more effective than focusing on specific areas. A visual representation of the scores for the different slices using PCA is shown in Figure 8.Thus far, our classifier is able to recognize SMC, EMCI and MCI perfectly. We will contrast the score obtained from the all-slices classifier against the outcome of CNN in the following section. By using PCA we have achieved an increment of more than 3% in F1 score, as shown in Figure 9. It is also notable that this approach provides a most homogeneous performance for every slice.The pattern observed in the previous section is repeated, using, now, CNN. Slices 104 and 106 are the most informative (although slice 72 was the most relevant in Experiment 2) and SMC, EMCI, MCI and LMCI are nearly perfectly matched, with a noticeable but not pronounced difficulty identifying CN and AD (See Table 5 and Table 6).The “multi-slice” aggregator (see previous section) is also included and performs better than considering the decision of every slice separately, reaching a maximum F1 score of 0.9902. Note that, for CNN, it is not possible to build the all-slices classifier because the input of our model would be too large for the machine on which this research was carried out.Despite the fact that the scores obtained using the custom CNN are more than adequate, it still performs worse than our SVM classifier based on PCA (recognising that we acheived an F1 of 0.9979 with this classifier)). In Figure 10 and Figure 11 we examine the final score for each slice and experiment.For the last experiment, we will use the VGG16 pre-trained model with a custom output (dense) layer using, again, “softmax” as the activation function. This model requires a longer time to train, since the number of layers and weights are significantly higher than in our custom CNN. Jumping directly into the results in Table 7, we can see that the previous pattern is not repeated: slice 56 is the best classifier. This is most likely due to easier recognition patterns present in slice 56 for VGG16. The highest score obtained belongs to “multi-slice” classifier as expected, with a total F1 score of 0.9620.It would be interesting to compare the results obtained from VGG16 with those outputted from other transfer learning models such as ResNet50, Inceptionv3 or EfficientNe. Although the preprocessing and training parts differ from the experiments conducted in [6], VGG16 can take second place between the pretrained models used in the previous classification. Nevertheless, this extension goes beyond the objective of this investigation. With respect to VGG16, a more pronounced difference in terms of performance is detectable between the slices (see Figure 12).Overall, the classifier performs very well. However, once we have covered all the possible approaches, the tested transfer learning solution proves not to be the best option between the available CNN alternatives, as shown in Figure 13.Finally, it is worth including the confusion matrix for the best classifier i.e., the confusion matrix for the SVM all-coefficients model. As shown in Figure 14, three out of the four wrong predictions were false LMCI negatives. This implies that the model can eventually make a wrong prediction when dealing with patients in advanced stages of the disease.It is worth noting that the accuracy score is 0.9973. This will be insightful in comparing the proposed classifier with previous investigations.The development of this work was motivated not only by medical purposes but also for the non-unified criteria regarding the usage of support vector machine and convolutional neural networks. Progressively, medicine and artificial intelligence tend to operate together and are proving to obtain outstanding results with respect to enhancing people’s lives. The limits of this symbiosis are still uncertain, although it seems undeniable that, from such combination, we are entering a new era in science, here, early in the 21st century.Considering our results, the following conclusions can be highlighted:It is possible to detect the early stages in Alzheimer’s disease and this prediction can be as precise as the prediction of dementia itself. Significant memory concern (SMC) and early mild cognitive impairment (EMCI) have been proven to have an effect on the brain that can be detected and measured. Patients with early symptoms of dementia can be localized and preventive treatments can be applied.If the MRI images reach a high level of normalization and enough samples are accessible it is possible to build an SVM classifier able to predict Alzheimer’s stages with an F1 score higher than 99.7%. As mention throughout the research, a key point of this investigation is the high quality of the dataset and the segmentation, bias correction and spatial normalisation applied by SPM12 [11,12] beforehand. The large number of samples (6,028) used it also relevant when compared with similar investigations [10]. The results show that our model has outperformed other modern experiments [5,6,7,14,24,25]. A more detailed comparison with some of the most promising investigations conducted to date is made in Table 8.In MRI images, some slices are more informative than others i.e., there are parts in the brain that contain more information and can be used more precisely to provide a diagnosis of the patient. Of the available slices in our dataset, slice 82 demonstrated the best results. This slice is located in the coronal plane, confirming the conclusions exposed by Luis Balderas in his thesis [9].In order to give a more accurate diagnosis, it is better to process all the information available in the brain rather than considering located regions only. Even if the disease has a more noticeable impact on specific regions, the information distributed throughout the brain’s mass makes a difference when seeking optimal results.Both SVM and CNN approached competent performances. Nevertheless, SVM stands out above CNN. A possible explanation for this is the normalization and regularity of the data. Since the available images are already resized, and the classifier is built using delimited and localized variation of the imaged zones, the edge identification power of CNN does not beat the capacity of SVM to allocate samples in Rn, n∈N and group them using the partitions generated by its trained hyperplane.The Mallat algorithm, revealed in [26], can be used to access the wavelet coefficients at deeper levels of the approximation image, LL. These coefficients are still very informative, exposing the power of the wavelet transform even in today’s image classification tasks. Using the wavelet coefficients from the approximation image at level four gave an outstanding F1 score of 0.9979. This classifier, which used all the available coefficients from the set of slices, performed better than slice-isolated classifiers accessing wavelet coefficients at level three.PCA performs better than regular feature selection algorithms when facing image classification problems where data has certain continuity properties. Features are highly correlated with each other and present small variations. Applying feature selection could lead to missing wider anomalies that would otherwise be detected using a dimensionality reduction system.It is possible to detect the early stages in Alzheimer’s disease and this prediction can be as precise as the prediction of dementia itself. Significant memory concern (SMC) and early mild cognitive impairment (EMCI) have been proven to have an effect on the brain that can be detected and measured. Patients with early symptoms of dementia can be localized and preventive treatments can be applied.If the MRI images reach a high level of normalization and enough samples are accessible it is possible to build an SVM classifier able to predict Alzheimer’s stages with an F1 score higher than 99.7%. As mention throughout the research, a key point of this investigation is the high quality of the dataset and the segmentation, bias correction and spatial normalisation applied by SPM12 [11,12] beforehand. The large number of samples (6,028) used it also relevant when compared with similar investigations [10]. The results show that our model has outperformed other modern experiments [5,6,7,14,24,25]. A more detailed comparison with some of the most promising investigations conducted to date is made in Table 8.In MRI images, some slices are more informative than others i.e., there are parts in the brain that contain more information and can be used more precisely to provide a diagnosis of the patient. Of the available slices in our dataset, slice 82 demonstrated the best results. This slice is located in the coronal plane, confirming the conclusions exposed by Luis Balderas in his thesis [9].In order to give a more accurate diagnosis, it is better to process all the information available in the brain rather than considering located regions only. Even if the disease has a more noticeable impact on specific regions, the information distributed throughout the brain’s mass makes a difference when seeking optimal results.Both SVM and CNN approached competent performances. Nevertheless, SVM stands out above CNN. A possible explanation for this is the normalization and regularity of the data. Since the available images are already resized, and the classifier is built using delimited and localized variation of the imaged zones, the edge identification power of CNN does not beat the capacity of SVM to allocate samples in Rn, n∈N and group them using the partitions generated by its trained hyperplane.The Mallat algorithm, revealed in [26], can be used to access the wavelet coefficients at deeper levels of the approximation image, LL. These coefficients are still very informative, exposing the power of the wavelet transform even in today’s image classification tasks. Using the wavelet coefficients from the approximation image at level four gave an outstanding F1 score of 0.9979. This classifier, which used all the available coefficients from the set of slices, performed better than slice-isolated classifiers accessing wavelet coefficients at level three.PCA performs better than regular feature selection algorithms when facing image classification problems where data has certain continuity properties. Features are highly correlated with each other and present small variations. Applying feature selection could lead to missing wider anomalies that would otherwise be detected using a dimensionality reduction system.Disease’s diagnosis using MRI is an open and extensive line of research. To continue and improve the steps followed in this investigation, I suggest these possible lines:It would be possible to access a higher level of detail either by using a machine with better specifications or performing mini-batch training techniques. This approach could lead to obtaining a more informative training dataset. Examples of this include using wavelet coefficients at lower levels or training a CNN classifier using all the available images as input.Investigate whether accessing different types of wavelet coefficients (diagonal, horizontal or vertical) can lead to a better outcome in F1 score or not.Research on the structure of CNN models to develop smarter and more suitable networks using different distribution and types of convolutional layers.Develop new paradigms of research to process 3D images and investigate the possible use and applications of the 3D wavelet transform.It would be possible to access a higher level of detail either by using a machine with better specifications or performing mini-batch training techniques. This approach could lead to obtaining a more informative training dataset. Examples of this include using wavelet coefficients at lower levels or training a CNN classifier using all the available images as input.Investigate whether accessing different types of wavelet coefficients (diagonal, horizontal or vertical) can lead to a better outcome in F1 score or not.Research on the structure of CNN models to develop smarter and more suitable networks using different distribution and types of convolutional layers.Develop new paradigms of research to process 3D images and investigate the possible use and applications of the 3D wavelet transform.Conceptualization, I.R. and J.J.P.; methodology, I.R. and J.J.P.; software, J.J.P.; validation, J.J.P.; formal analysis, J.J.P.; investigation, J.J.P.; resources, I.R. and J.J.P.; data curation, I.R. and J.J.P.; writing—original draft preparation, J.J.P.; writing—review and editing, J.J.P. and I.R.; visualization, J.J.P.; supervision, I.R.; project administration, I.R.; funding acquisition, I.R. All authors have read and agreed to the published version of the manuscript.This research was funded by Ministerio de Ciencia, Innovación y Universidades (program RTI2018-101674) and Junta de Andalucía (program Q1818002F).Not applicable.Not applicable.Data was obtained from http://adni.loni.usc.edu in May 2020 and is available with the permission of The Alzheimer’s Disease Neuroimaging Initiative (ADNI).The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.The following abbreviations are used in this manuscript:
+      CNCognitively NormalMCIMild Cognitive ImpairmentEMCIEarly Mild Cognitive ImpairmentLMCILate Mild Cognitive IMpairmentSMCSignificant Memory ConcernADAlzheimerś diseaseMRAMultiresolution AnalysismRMRMinimum Redundancy Maximum RelevanceSVMSupport Vector MachineCNNConvolutional Neural NetworkTLTransfer LearningStratification of classes according to their severity. From left to right, the different stages in Alzheimer’s disease are presented sequentially. The classes on the left correspond to early stages of dementia, whereas the classes on the right correspond to advanced stages.Frequency distribution of classes. The classes are sorted from left to right by number of samples.Example of images for a given individual. From the top left corner to the lower right corner, the images correspond to slices: 55, 56, 61, 72, 82, 90, 104, 106 and 114. The patient shown belongs to category AD.Recursive version of mRMR algorithm. The original set is divided into multiple subsets of 1000 features, which is the maximum number of features to which mRMR is applied. This means that, if we have N>1000 features, we will generate (N//1000)+1=C subsets, where N//1000 is the integer quotient resulting from the division N/1000. The ouputs of the C subsets are joined together to construct a new input set.SVM diagram. The input image is converted to BGR2 grey. The wavelet coefficients are later extracted and mRMR/PCA are used to perform feature selection. Finally, an SVM model is trained with the selected coefficients.F1 mRMR scores. Slice 116 presents the lowest F1 score as processed using SVM and mRMR, whereas the all-coefficients classifier performs the best, with an F1 score of 0.9657.Multi-Slice classifier functioning. The prediction of every single-slice model is retrieved for every patient and the mode is applied. A final output is obtained, which corresponds to the most common value between the single-slice model predictions.F1 PCA scores. Slice 61 presents the lowest F1 score when processed using SVM and PCA, whereas the all-coefficients classifier performs best, with an F1 score of 0.9979.F1 mRMR-PCA scores comparison.Custom CNN F1scores from Experiment 1. Slices 104 and 106 present the highest F1 score, although the multi-slice classifier performs the best, with an F1 score of 0.9902.F1 custom CNN experiment 2 scores. Slice 72 and the multi-slice classifier present the highest F1 score: 0.9857.F1 TL CNN scores. Slice 56 presents the highest F1 score although the multi-slice classifier performs the best with a F1 score of 0.9620.F1 tl-custom CNN scores comparison. The three CNN approaches’ results are shown together. Experiments 1 and 2 perform better than the transfer learning model based on VGG16. Experiment 1, in turn, stands out from Experiment 2, reaching the maximum F1 score: 0.9902.SVM All-coefficients confusion matrix. Classes SMC, EMCI, MCI and LMCI have perfect precision. Classes EMCI, MCI and SMC have perfect recall.Number of wavelet coefficients per image for some approximation levels. The lower the level we access, the better the image can be reconstructed. Levels higher than five are not relevant for this investigation and therefore are not included.Hyperparameter space considered for SVM. A total of 3×4×5=60 combinations are tested when applying a grid search. This number corresponds to all the possible combinations between kernels, C and gamma parameters.F1 results for SVM combined with mRMR. In the table we can see the performance of each slice when predicting specific stages. The average score is included on the right side. The all-coefficients classifier obtains the highest F1 score, 0.9657. The best single-slice classifier corresponds with slice 82.F1 results for SVM combined with PCA. In the table we can see the performance of every slice when predicting specific stages. The average score is included on the right side. The all-coefficients classifier obtains the highest F1 score, 0.9979. The best single-slice classifiers correspond to slices 104 and 106.F1 results for custom CNN, Experiment 1. In the table we can see the performance of every slice when predicting specific stages. The average score is included on the right side. The multi-slice classifier obtains the highest F1 score, 0.9902. The best single-slice classifier correspond to slices 104 and 106.F1 results for custom CNN, Experiment 2. In the table we can see how every slice performs when predicting specific stages. The average score is included on the right side. Multi-Slice classifier obtains the highest F1 score along with slice 72, 0.9902.F1 results for CNN using transfer learning. In the table we can see the performance of every slice when predicting specific stages. The average score is included on the right side. The multi-slice classifier obtains the maximum F1 score, 0.9620. It is closely followed by slice 56, which reaches an F1 score of 0.9618.More detailed comparison between recent relevant investigations. Some experiments have already passed the accuracy threshold of 0.99. The proposed method exceeds this threshold (0.997) and includes six different stages in its classification.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-01-03-00013.txt

@@ -0,0 +1,4 @@
+Modern network science has been used to reveal new and often fundamental aspects of brain network organization in physiological as well as pathological conditions. As a consequence, these discoveries, which relate to network hierarchy, hubs and network interactions, have begun to change the paradigms of neurodegenerative disorders. In this paper, we explore the use of thermodynamics for protein–protein network interactions in Alzheimer’s disease (AD), Parkinson’s disease (PD), multiple sclerosis (MS), traumatic brain injury and epilepsy. To assess the validity of using network interactions in neurological diseases, we investigated the relationship between network thermodynamics and molecular systems biology for these neurological disorders. In order to uncover whether there was a correlation between network organization and biological outcomes, we used publicly available RNA transcription data from individual patients with these neurological conditions, and correlated these molecular profiles with their respective individual disability scores. We found a linear correlation (Pearson correlation of −0.828) between disease disability (a clinically validated measurement of a person’s functional status) and Gibbs free energy (a thermodynamic measure of protein–protein interactions). In other words, we found an inverse relationship between disease disability and thermodynamic energy. Because a larger degree of disability correlated with a larger negative drop in Gibbs free energy in a linear disability-dependent fashion, it could be presumed that the progression of neuropathology such as is seen in Alzheimer’s disease could potentially be prevented by therapeutically correcting the changes in Gibbs free energy.The treatment and management of neurological dysfunction/neurodegeneration is an area of great medical need. The World Health Organization (WHO) estimates that neurological disorders contribute 10.9% and 8.7% of the global disease burden in high- and medium-income countries, respectively [1]. Moreover, as the average age of the populations in developed countries will increase in the coming decades, it is expected that the disease burden will continue to increase. However, the present treatment of neurological diseases constitutes, by and large, the management of disease symptoms, because the etiology remains unclear. In Alzheimer’s disease, for example, the earlier etiological hypothesis that amyloid deposits are caused by environmental stimuli is being supplanted by a growing body of evidence that it is the genomic dysregulation of cellular and molecular pathways that cause accumulation of beta amyloid and MAP tau proteins [2]. This may not only explain why targeting amyloid and/or tau proteins has been unsuccessful so far, but it also unlocks the possibility of using targeted therapies. However, to realize the full potential of targeting these newly identified genomic patterns in Alzheimer’s disease [3,4] and in order to pursue these novel, molecularly guided therapies, an improved understanding of the complexity of the system biology of neurological diseases is needed.Many pathologic conditions, such as, for example, neurodegenerative diseases, chronic inflammatory disorders or cancer, are associated not only with mutational activation of genes, but also with re-activation of developmentally silenced pathways. Because the processes of tissue invasion, proliferation, inflammation and angiogenesis are common not only to oncogenic induction, but also to embryonal development and normal host response to tissue injury, a simple DNA mutation analysis would not provide sufficient information. Therefore, the activation of pathways associated with inflammation and tissue injury is best studied using mRNA expression as a quantitative measure. However, until very recently, it has been very difficult to directly correlate the levels of gene expression with the cellular effect of specific proteins. In fact, high gene expression did not always imply increased protein function or activation of a biological process. The intensity of a biological effect is dependent on the interactions of the affected (overexpressed) gene with its neighbors, and the quorum effect of the innumerable feedback loops on the global protein–protein interaction network (PPIN).The realization that the perturbations of individual genes can be measured by their effect on the global response of a network has led many scientists to find alternative approaches for genomic interpretations. The one proposed and presented in this manuscript is based on using not only the level of expression of a gene, but also its topology as a measure of its connectivity within a network. This novel approach has been enabled by the vast amounts of well-curated information accumulated in publicly available PPINs over the last 4 decades, and has emerged quite recently [5,6,7,8].The underlying premise of our analysis is that biological systems are in essence complex chemical networks. A cell, for example, consists mainly of a large molecular network made of DNA, RNA, proteins, peptides, small molecules and lipids. Each of these molecules is associated with potential energy contributing to an even larger energetic network. The energetic state exists in or close to thermodynamic equilibrium, and any perturbation sets off a cascade of events striving to bring the overall network back to the same entropy level, which is a hallmark of homeostasis. The prime force pushing the reaction back toward equilibrium is the Gibbs free energy, an expression of the thermodynamic energy reflecting the chemical potential between interacting proteins. This Gibbs free energy is used as a measure of the changes occurring within a disease-related PPIN.In an earlier work, Rietman et al. [9] described an inverse correlation between the Gibbs free energy and 5-year survival probability for ten different types of cancers. The study showed that poor prognosis cancers such as glioblastoma multiforme have low thermodynamic entropy, less negative Gibbs free energy and very low 5-year survival. This was consistent with the clinical status of the disease, which has an average survival post diagnosis of 6 months, and a 5-year survival of 2%. Similarly, it was found that in breast carcinoma, which had a higher thermodynamic entropy, a more negative Gibbs free energy and a much higher 5-year survival were also congruent with the clinical observation, as the average 5-year survival for breast cancers of all stages is ~88%.In this manuscript, we describe a linear relationship between Gibbs free energy (a measure of thermodynamic energy for a specific disease), and disability weight (a clinically validated measurement of a person’s functional status) for several neurological diseases for which consistent data were available. The choice of thermodynamics for the analysis was not fortuitous. Thermodynamic energy represents an important driving factor of chemical and biological interactions for all living organisms, and its correlation with biological events is not unexpected. Gibbs free energy, which incorporates information on both the mRNA expression and the complexity of the protein–protein interactions, would be expected to have the ability to discriminate between “passenger genomic events” and “driving genomic events”. This remains the main quandary—finding ways to differentiate between causative, as opposed to ancillary molecular changes. The use of Gibbs free energy in this context represents a novel approach, and may be of particular usefulness in understanding neurological disorders/neurodegenerative diseases where complex and by and large unclear etiology, pathogenesis and clinical response prevent identification of effective therapeutic strategies.Gibbs free energy (G) is the energy associated with a chemical reaction that can be used to do work. The free energy of a system is the sum of its enthalpy (H) plus the product of the absolute temperature (in Kelvin) and the entropy (S) of the system.We propose that the use of this well-established thermodynamic measure is useful for analyzing the interplay between patient’s genomic information and the existing knowledge about PPINs [7]. Its use is based on a number of important observations. First, a system comprising proteins interacting with a large number of other proteins (even if not simultaneously) has a higher entropy. Because each protein–protein interaction has a different molecular configuration, given by Boltzmann’s classic equation (S = k ln(W) where k is the Boltzmann constant), the entropy, S, increases as the natural logarithm of the number of available configurations (microstates), W. As such, proteins with many interaction partners exhibit many possible configurations, and each protein–partner interaction leads to a different configuration. Highly interconnected proteins such as, for example, ubiquitin (UBC) or p53 (encoded by the TP53 gene) can undergo a simultaneous physical interaction with hundreds of their respective interaction partners, because at any given time one UBC molecule interacts with a protein and a different UBC molecule interacts with another protein within the same cell. An RNA transcriptome from a tissue biopsy thus represents an ideal mixture of UBC and its interacting partners for a given condition.Second, transcription (RNA expression levels) data are good surrogates for protein concentration values. Unlike the DNA-level gene alterations, which are transcribed with variable frequencies to RNA, the number of mRNA copies is translated into individual proteins with great fidelity. Several research groups have confirmed this fidelity: Greenbaum et al. [10] and Maier et al. [11] report a Pearson correlation in the range of 0.4–0.9 for a large set of experiments across five different species. Similarly, Kim et al. [12] and Wilhlem et al. [13] found an 83% correlation between human transcription data and mass spectrometry proteomic data for multiple tissue types, supporting the use of human transcriptome as a surrogate for protein concentration data.Third, the use of a real-world dataset contains an inherent level of noise, but as new mRNA data sets emerge from ongoing clinical trials, the accuracy of the information in protein–protein interaction databases as well as in the gene expression datasets will improve as well. For the time being, the preliminary findings represent a convenient way to discover new avenues for future analysis, but as data integrity and quality of our conclusions improve, we should be able to use these data for increasingly more reliable therapeutic decisions. In addition, the ability to combine different data sources (mRNA expression data from individual patients and existing PPINs) as introduced in this manuscript is likely to deliver new insight into biologically complex diseases than traditional approaches.Because the study of chemical thermodynamics embodies chemical potential, for two molecules A and B interacting to form a new molecule, or an A-B molecular complex, the amount of A-B formed would be dictated by the amount of A and B. In cases where A is present in a higher concentration than B, a chemical potential develops. As such, a protein, D, interacting with proteins, C, E and F, has a chemical potential represented by:(1)μD=lncDcC+cD+cE+cF
+      where the chemical potential is the natural logarithm of the concentration of protein D divided by the sum of the concentrations of protein D and the combined concentrations of all its neighbors. Because the argument of the natural logarithm is a ratio, we can use scaled “concentrations”, or in this case, scaled expression values. The log-2 normalized expression data typically fall in the range (–10,10). Rescaling sets the range to be [0,1], and we can compute the minimum, emin, and the maximum, emax, for a given expression data set. The normalized expression values for each gene are then computed as follows:(2)ci=ei−eminemax−eminThe rescaling is justified from a mathematical perspective by the fact that the argument of the natural logarithm must be positive. Furthermore, if a gene mutation leads to loss of its RNA transcription (RNA transcription is said to be downregulated), the concentration for the respective protein would essentially be zero. Likewise, when a gene alteration leads to constitutive activation of its transcription, multiple copies of its mRNA will be made (the RNA transcription is said to be highly upregulated), and very large quantities of protein will be produced. In this case, the protein concentration would be effectively set to the maximum value of 1.Thus, the computation of Gibbs free energy for a single protein in the PPI would be
+      (3)Gi=lnci∑j=icj
+      which suggests that we should first compute the chemical potential for protein i with neighbors j, and multiply it by the [0,1] scaled concentration to obtain the Gibbs free energy for protein i. For the overall Gibbs free energy for the network, the individual G’s for each of the protein within the network are summed up. In the final calculation, the normalized expression data are overlaid on the BioGRID PPI and Gi followed by use of the above equation. Note, the SI units in Equation (3) are J/mole; we use a modification of Equation (3) by multiplying moles, ci, on the right-hand side. Therefore, we have the mapping—a quasi-Gibbs energy:(4)qGiJmole←cilnci∑j=icjJThe mapping is simply a scaling factor, but the actual units on the graphs and in the table would, at best, be difficult to interpret because we do not have molar concentration of the proteins, only their relative expression levels.We used three main data sources for our analysis: (1) World Health Organization data on disability associated with neurological diseases, (2) protein–protein interaction data from Biological General Repository for Interaction Datasets (BioGRID, Human ver. 3.4.139, September 2016, https://thebiogrid.org/ accessed on 7 April 2020) and (3) RNA transcription data sets from Gene Expression Omnibus (GEO).Unlike terminal conditions such as cancer, there is no direct correlation between death rate and disability in neurodegenerative diseases, and death rate is not a meaningful measure of morbidity. For example, although epilepsy may cause more deaths, multiple sclerosis (MS) far outweighs its impact in the sense of morbidity and disability during a person’s lifetime. For this reason, the World Health Organization has selected Disability-Adjusted Life Years (DALYs) as a measure to use for evaluation of disability associated with neurodegenerative diseases. DALYs combine two components: years of life lost due to premature mortality, and years lived with disability. DALYs are an expression of the number of healthy years a person loses in life. It is a more accurate representation of the damage a disease exerts on a healthy human population, and its measure—Disability Weight—is a number between 0 (perfect health) and 1 (death). Although DALYs are population-dependent, and the same disease may lead to a larger apparent loss in a region where the disease is widespread, Disability Weight avoids the potential for biasing the numbers towards more prevalent diseases and away from uncommon diseases, because it rates the severity of a disease in an individual. The most recent public list of WHO for neurological diseases and their corresponding disability weights was published in 2006, and the “Neurological Disorders, public health challenges” [1] describes the demographics, geographic distributions, graphs and projections for many neurological diseases. It standardizes the comparison of neurological diseases, and was, therefore, used for our analysis.The source of the transcription data was Gene Expression Omnibus (GEO) repository of -omics and high-throughput data https://www.ncbi.nlm.nih.gov/geo (accessed on 7 April 2020). The data sets for each of the selected neurological diseases were collapsed from probe IDs to gene IDs using GenePattern software (Broad Institute, Cambridge, MA, USA; https://software.broadinstitute.org/cancer/software/genepattern, accessed on 25 November 2020), and the corresponding chip platform documented for the respective data set. The following datasets for specific diseases were examined: Alzheimer’s disease (GDS4136), Parkinson’s disease (GSE6613), Multiple Sclerosis (GSE19285), Epilepsy (GSE32534), Cerebrovascular disease (GSE36791) and Meningitis (GSE40586). Most data sets were already log-2 normalized, and we transformed those that were not. Table 1 lists the GEO dataset number and PubMed ID (PMID) for each of the diseases.Neurological disorders are common and represent a major public health problem. Even though neurological impairment and its sequelae constitute over 6% of the global burden of disease [1], the management of neurological disorders has not significantly changed in the past few decades, and the mainstay of therapy remains focused on symptomatic management. As such, the already very high disease burden is likely to continue to increase as life spans across the world’s population increase. According to the recently published Global Burden of Disease 2010 Study (GBD 2010) [14], stroke is the second leading cause of death globally and the third leading cause of premature death and disability as measured in disability-adjusted life years (DALYs).There are many reasons for the lack of effective therapies, but the principal challenge is the complexity of data and the inconsistency in assigning causality to genomic alterations. It is a great challenge to discriminate between incidental molecular findings, and those that may be driving disease pathogenesis. This not only hinders the search for effective therapies, but the absence of tissue targets also prevents effective clinical initiatives. Until recently, much effort was dedicated to statistical interpretation of RNA expression levels. However, the level of mRNA expression is not always reflective of the gene importance in a biological event or in a particular disease. An overexpressed gene that is peripheral to a major proliferative pathway will have a minimal effect on proliferation, whereas a mild elevation in the expression levels of a well-connected gene will have a crucial effect on the process. Minute changes in the levels of genes coding for very important growth factors such as VEGF, or inflammation regulators interleukins lead to biological events that are normally carefully managed through feedback loops.Any measurement of disease activity must, therefore, incorporate the intracellular protein–protein interactions. We introduce a method for interrogating individual tissue expression of mRNA against existing, well-curated PPINs. We focus on proving that thermodynamics, i.e., the molecular changes defined by Gibbs free energy, can be correlated with disease state and progression. Figure 1 shows the correlation of Disability Weight with average Gibbs free energy values for six neurological conditions. The relationship between disease-related disability and Gibbs free energy is linear, with progressively worsening disability correlating with increasingly more negative Gibbs free energy. The respective values used for the figure are shown in Table 1. We confirm that low entropy (less negative Gibbs free energy) for epilepsy correlated with the lowest disability, whereas the higher entropy (more negative Gibbs free energy) in Alzheimer’s disease correlated with the highest disability. These findings correlate well with clinical observations that the severity of neurological dysfunction in multiple sclerosis, bacterial meningitis and Alzheimer’s disease is certainly higher than in epilepsy.The mRNA expression values available from publicly available GEO data sets (Alzheimer’s disease GDS4136, Parkinson’s disease GSE6613, Multiple Sclerosis GSE19285, Epilepsy GSE32534, Cerebrovascular disease GSE36791 and Meningitis GSE40586) were used to calculate Gibbs free energy, a thermodynamic measure of protein–protein interactions. As would be expected based on the level of clinical disability, epilepsy has low Gibbs free energy (low entropy) and correlates with the lowest neurological disability. This is in stark contrast with the high Gibbs free energy (high entropy) and high neurologic disability in Alzheimer’s disease, as would be consistent with clinical observations in this disease. The respective values, and the size of the cohort are summarized in Table 1, and the error bars have been set to 5% of the average Gibbs Free energy value, given that the actual errors of the reported mRNA gene expression values were not reported.We further explored whether Gibbs free energy correlated not only with severity of the disease, but also with disease progression. Seven stages have been described in Alzheimer’s disease (AD) progression, from no impairment (Stage 1) to loss of ability to respond to their environment or communicate, needing assistance with all activities of daily living and loss of ability to swallow (Stage 7). The GEO data sets tend to simplify these stages into only four stages (Figure 2a), and we show a clear linear relationship (R = 0.7978) between thermodynamic energy (Gibbs free energy) and the four stages of the disease. The positivity of the slope is most likely reflective of neuronal loss, with less and less metabolically active tissue. This is in direct contrast with epilepsy, typified by an abnormality of conduction of an action potential across neuronal tissues rather than by a neuronal loss, where Gibbs free energy is less negative (i.e., more positive).Similarly strong correlation (R = 0.932) was observed between Gibbs free energy and the degree of amyloid deposition in Alzheimer’s disease (see Figure 2b). In this case, progressive histological changes in the form of extracellular deposits of amyloid β peptides, senile plaques and intracellular neurofibrillary tangles of hyperphosphorylated tau in the brain relate to neuronal death and correlate with severe disability. The amyloid metabolic cascade and the post-translational modification of tau protein, often considered causal in AD, are sufficient to explain the diversity of biochemical and pathological abnormalities in AD. There is a multitude of cellular and biochemical changes leading to the accumulation of extracellular senile plaques made of deposits of Aβ peptide, and all have the outcome of neural degeneration and loss. The correlation of Gibbs free energy with histological changes and disease progression implies a global value of our measurement and the need for future evaluation of the hidden metabolic information.The analysis of 2000+ Alzheimer’s Disease patients from GSE84422 data set revealed a correlation coefficient R = 0.7978 between Gibbs free energy and disease progression using disease stage (Figure 2A), and an even stronger correlation (R = 0.932) using tissue pathological analysis such as amyloid plaque density (Figure 2B).Yet, another scale is used by clinicians for evaluation of disability in multiple sclerosis (MS). The Expanded Disability Status Scale (EDSS) was developed by a neurologist (John Kurtzke) in 1983, and ranges between 0–10. It is based on neurological evaluation of pyramidal (limb movement), cerebellar (ataxia, coordination, tremor), brainstem (speech, swallowing and nystagmus), sensory, bowel and bladder, vision and cerebral (mental) functions. Its 0.5-unit increments represent progressively higher levels of disability. EDSS 1.0–4.5 refers to people with MS who are able to walk without any aid, whereas 5–6 refers to progressive motor and cognitive disability. Similarly to AD, we find a positive linear slope between Gibbs free energy and disability in MS (see Figure 3), as consistent with neuronal loss due to demyelination. The respective correlation coefficient between EDSS and Gibbs free energy in MS was found to be very strong at R = 0.913.As shown in Figure 3, thirty patients with multiple sclerosis from the GEO data set GSE19285 revealed a correlation coefficient R = 0.913 between Gibbs free energy and disease progression as reflected by The Expanded Disability Status Scale (EDSS).In this paper, we have provided early evidence for using not only expression data, but also connectivity/topology data obtained from the associated PPINs for analysis of genomic information in neurodegenerative diseases. One of the approaches is based on using a thermodynamic measure such as Gibbs free energy to identify the disease-related global changes in the associated PPIN resulting from changes in genomic profiles of individual patients. The ability to correlate these molecular profiles with disease severity and disease progression suggests that we can use Gibbs free energy in the future to evaluate causative versus ancillary molecular changes through mathematical simulations of the protein inhibition/stimulation. This is the first installment in developing mathematical algorithms that would facilitate identification of relevant, therapeutically targetable pathways in neurodegenerative diseases. The use of Gibbs free energy in genomic analysis would be beneficial not only from a therapeutic point of view, but also from a cost and sustainability perspective, because it would minimize futile clinical trials.Conceptualization, E.A.R. and J.A.T.; methodology, E.A.R.; software, M.K. and H.T.S.; formal analysis, E.A.R. and G.L.K.; investigation, G.L.K. and M.K.; resources, H.T.S.; data curation, M.K.; writing, M.K., E.A.R., H.T.S. and G.L.K. All authors have read and agreed to the published version of the manuscript.This research received no external funding.Not applicable.Not applicable.Not applicable.We acknowledge (NSF)/ECCS-1533693 NSC-FO: Col “Individual Variability in Human Brain Connectivity, Modeling Using Multi-scale Dynamics Under Energy Constraints”, and Office of Naval Research (ONR)/N00014-15-2126. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Science Foundation, or the US Navy. We thank Samuel McGuire for helpful discussions. EAR and GLK acknowledges partial funding support from CSTS Healthcare, Toronto, Canada. JAT acknowledges research support for this project provided by NSERC (Canada).The authors declare that they have no conflict of interest in regard to this manuscript.The correlation of Disability Weight and Gibbs free energy for six distinct neurological conditions.The correlation between disease progression and pathological findings with Gibbs free energy in Alzheimer Disease.The correlation between disease progression and Gibbs free energy in Multiple Sclerosis.Disability score and Gibbs free energy for the neurological disorders studied.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00001.txt

@@ -0,0 +1 @@
+The use of artificial intelligence (AI) systems in biomedical and clinical settings can disrupt the traditional doctor–patient relationship, which is based on trust and transparency in medical advice and therapeutic decisions. When the diagnosis or selection of a therapy is no longer made solely by the physician, but to a significant extent by a machine using algorithms, decisions become nontransparent. Skill learning is the most common application of machine learning algorithms in clinical decision making. These are a class of very general algorithms (artificial neural networks, classifiers, etc.), which are tuned based on examples to optimize the classification of new, unseen cases. It is pointless to ask for an explanation for a decision. A detailed understanding of the mathematical details of an AI algorithm may be possible for experts in statistics or computer science. However, when it comes to the fate of human beings, this “developer’s explanation” is not sufficient. The concept of explainable AI (XAI) as a solution to this problem is attracting increasing scientific and regulatory interest. This review focuses on the requirement that XAIs must be able to explain in detail the decisions made by the AI to the experts in the field.The terms artificial intelligence and machine learning are sometimes used interchangeably, although this is incorrect. In fact, artificial intelligence is a branch of computer science that deals with the automation of human activities that are normally considered intelligent human behavior [1]. These activities include understanding human language, representing and using knowledge, reasoning, planning, problem solving, and risk assessment, including guessing and learning from experience. Machine learning is currently by far the most popular method used in artificial intelligence and can be referred to in two different forms [2]: first, approaches in which a class of very general algorithms (artificial neural networks, classifiers, predictors, associative memories, etc.) are tuned based on examples to optimize the prediction or classification of new, unseen cases. This is skill learning. Second, methods that recognize structures, such as subgroups, in the data and describe these structures in such a way that such a description (knowledge) can be used to correctly classify new cases and can be understood by humans. This is the deduction of knowledge from data [3,4].The growing importance of artificial intelligence or machine learning algorithms in biomedical research is reflected in their increasing influence on clinical decision-making processes. This, in turn, has a direct impact on medical practice and communication between physicians and patients. The traditional doctor–patient relationship is based on personal trust and transparency of medical advice and therapeutic decisions. If the diagnosis or the selection of the most promising therapy is no longer made solely by the physician, but to a considerable extent by a machine with learning algorithms and artificial intelligence, the decisions become nontransparent. Physicians cannot be assumed to have the computer science skills necessary to understand the decision-making process of an algorithm and should be able to communicate this process in all relevant details to their patients in an understandable way.The European Union (EU) has recognized the problem that algorithm-based medical decision making poses to the information rights of affected patients and has published a landmark paper highlighting the need for explanations of computerized decisions so that they can be communicated to affected patients in an understandable manner [5]. The solution is found in the concept of explainable AI (XAI), which is attracting increasing scientific interest [6]. This is consistent with the U.S. military’s efforts to obtain explainable models that make decisions made by autonomous systems transparent (https://www.darpa.mil/program/explainable-artificial-intelligence [7], accessed on 15 December 2021). Without a deep understanding, machine learning relies on trial and error and has been compared to medieval alchemists [8]. Along the same line of reasoning, this review focuses on the requirement for XAI to be able to explain in detail the decisions made by an AI in a biomedical setting to the expert in the domain, e.g., the physician in the case of AI-based clinical decisions related to diagnosis, treatment, or prognosis of a disease.A common clinical situation is the communication of a diagnosis by the physician to the patient. The diagnosis is made on the basis of sound decision criteria, such as the presence of pathognomonic signs or the excess of a laboratory value over the generally accepted limit for healthy individuals. For example, the diagnosis of lymphoma can be made by observing specific cell types in a patient’s blood sample. With increasing automation, the assessment of the many cell subpopulations in a sample is increasingly performed by algorithms, including the analysis of microscopic images for cell type separation and counting [9].A typical but small data set, freely available with the R library “opdisDownsampling” (https://cran.r-project.org/package=opdisDownsampling [10], accessed on 15 December 2021), consists of d = 6 cytological markers measured by fluorescence-activated cell sorting (FACS) in a total of n = 111,686 cells obtained from 100 patients with chronic lymphocytic leukemia and 100 healthy controls, using the seed value of seed = 42, which is reported here for reproducibility of the results with the referenced R libraries and example data included there. After class-proportional downsampling to 3000 instances to speed up subsequent computations, the data space consisting of data space D=xi,yixi,yixi,d∈RX,yi∈Y{1,2},i=1⋯nxi,d∈RX,yi∈Y{1,2},i=1…n with input space X consisting of vectors xi = <xi,1,...xi,d> with d = 6 different cytological markers and the output classes yi consisting of the diagnoses of healthy versus diseased, two different algorithms were trained to map the cell marker pattern xi to the diagnosis classes yi, i.e., to automatically perform the diagnosis of leukemia. Since the present analyses were intended for the demonstration of an introductory example and not to discuss techniques of classifier training and tuning, the interactive R data mining tool rattle (https://cran.r-project.org/package=rattle [11,12], accessed on 15 December 2021) was used with the default parameter setting. The reasons for the choice of implementation details were explained in the cited publications and are not challenged here.Thus, the downsampled data set was split into training/test/validation subsets sized 70%/15%/15% of the total data set as advised in “rattle”. Using the seed of 42, a standard classification and regression tree (CART) [13] and a support vector machine (SVM) [14] were trained by calling the respective R libraries “rpart” (https://cran.r-project.org/package=rpart [15], accessed on 15 December 2021) and “kernlab” (https://cran.r-project.org/package=kernlab [16], accessed on 15 December 2021). The algorithms were trained and tuned on the training and test data sets, and the accuracy of class assignment was assessed in the validation subsample by calculating accuracy and the area under the receiver operator characteristic (AUC-ROC). The default settings of “rattle” do not include cross-validation or repeated calculations; this was considered sufficient for the present exemplary demonstration purpose and therefore was not changed or refined. For the same reason, no grid search for hyperparameter tuning and similar standard classifier tuning procedures were performed. The two algorithms provided a nearly similar accuracy of assigning a cell sample to “healthy” or “diseased” of 0.7711 for CART (“rpart”) and 0.7689 for the SVM (“ksvm”). The AUC-ROC values were also nearly identical (Figure 1A). However, the transparency of the class assignment decision was completely different for the two algorithms, as explained below.CART delivers a single simple rule as explanation for the decision, namely that the sample belongs to the disease if the expression of the CD19 marker has a value of 2.9 or more; otherwise, it is from a healthy subject (Figure 1B). This can be communicated to the physician, who understands the role of CD19. Such a biomedical expert will know that the expression level of CD19 on cell surfaces plays an important role for the functioning of B cells [19]. With this information, transparency is transferred from the field of informatics back to medicine, where the physician has to explain the meaning of CD19 to the patient, while the decision-making process of the algorithm is made transparent to both the physician and the patient. Thus, transparency of XAI does not necessarily mean transparency for the patient, but emphasizes the compressibility of a biomedical decision based on machine learning first for the (biomedical) field expert, who then takes over the establishment of comprehensibility for the layperson.SVM are machine learned classifiers, which use kernel functions to assign data to given classes [14]. Kernel functions represent distances from a hyperplane in a space, where the original data are mapped to. This decision space is typically much higher in dimension (up to infinity) than the feature space of the data. This may have the result that the classification is easier in that space. However, a representation of the decision surface in the data’s space is typically torn and may contain holes and bumps, i.e., senseless [20]. The SVM explanation of why a patient receives a diagnosis of leukemia based on his/her blood sample would be that because the patient’s blood sample (thick green dot in (Figure 1C) contains cell marker patterns that place it on the “sick” side of the decision surface (black line in Figure 1C), which separates healthy and sick cells in a unintelligible projection of marker expressions (Dim1/Dim2 in Figure 1C).Explainable AI is not a new field [21]. AI systems were extensively researched in 1980/1990. These systems were based on a precise and formal representation of human knowledge using predicate logics, graphs, e.g., directed acyclic graphs (DAG), and a type of approximate reasoning, such as Bayes [22], fuzzy reasoning, and Dempster–Shafer theory [23]. For example, one of the world’s highly successful systems of these knowledge-based systems is the GeneOntology knowledge base [24].However, a serious bottleneck of these systems is that the AI needs a knowledge representation created by hand before it can start working. Algorithms that can learn to appear to act intelligently seemed to be a solution to this problem. However, most of the machine learning models in use today are neither knowledge based nor knowledge producing. From the perspective of knowledge-based AI, these systems sacrifice understandability and explainability in favor of performance. A better name for most machine learning systems and many “AI” systems would, therefore, be artificial skills-based systems (AS), which is elaborated in the next chapter.A transparent decision based on AI could ideally be achieved when a sound scientific theory is available as a basis for how the underlying ML system works. Then, the trustworthy system can draw logical inferences based on this theory to reach its conclusions. This is like knowing Kepler’s laws for predicting the positions of planets in astronomy. In systems based on sound scientific theory, the scope and accuracy of a prediction can be estimated. In addition, a rationale (explanation) for the result can be given [25]. In astronomy, for example, Kepler’s three laws can be derived from Newton’s law of gravity. From these laws, it can be deduced why a particular planet is in a particular position.Machine learning systems are often used for tasks where a scientific theory is not given or even known. For example, machine learning systems are developed to diagnose patients based on various measurements of gene expression, even when the exact molecular processes involved in the disease are only partially known or understood. In such situations, the machine learning literature is content to measure the “quality” of a diagnostic system by measuring the accuracy of its predictions on a limited data set that was not used during the development (training, learning, adaptation, and tuning) of the system, i.e., the so-called “test data” [26,27]. That is, the algorithm is trained on a carefully selected training and test data set to develop the ability to perform a specific task, such as making a clinical diagnosis. In this way, the ability to generalize to new, unknown cases is evaluated. In this approach, confidence is determined by a measure of performance on unseen data. However, in most cases, these “unseen” data were already available when the model was developed.The limitations of skill-based ML systems are obvious: for data that are very similar in structure to the training data, the system will perform well. For data that have a different structure, the skill-based ML system will fail, and may not even recognize that such data do not fall within the algorithm’s skill domain. In astronomy, this is like the epicycle model of planetary motion. It can be thought of as an empirical Fourier analysis of planetary motion, with a series of larger and smaller circles superimposed [28]. For small periods of time and under “normal” circumstances, the epicycle model can predict the position of a planet to some accuracy [29]. However, it is not known when the prediction is correct or incorrect.For skill-based ML systems, it is pointless to ask for an explanation for a decision. Systems of the “associative memory” type, for example, store all cases and their diagnoses in a memory (database). The diagnosis of a new case is determined by searching for the most similar cases and assigning the majority of the diagnoses of the most similar case. An example of this type of algorithm is the k-nearest neighbor classification algorithm [30]. Attempts to analyze skill-based algorithms in detail only lead to an understanding of the mathematical model used in such a system. For example, patient A’s diagnosis is D because A is most similar to patient X, who had a diagnosis of D in the past. Moreover, fairness and nondiscrimination against minorities, as well as other ethical requirements, such as not harming people, cannot be guaranteed or enforced in skill-based systems (see https://digital-strategy.ec.europa.eu/en/library/communication-artificial-intelligence-europe, accessed on 15 December 2021).The main types of classifiers used in machine learning are symbolic [31] or sub-symbolic [32] classifiers. For symbolic classifiers, the decision of how a classification is arrived at can be interpreted by a domain expert as a combination of conditions on the features. For example, a symbolic classifier may consist of a set of rules that are hierarchical in a decision tree or non-hierarchical. This is consistent with what is currently required for an XAI. In contrast, subsymbolic algorithms do not make transparent the exact criteria used to assign a subject to a particular class, e.g., healthy or sick. An example of subsymbolic classifiers are random forest classifiers [33,34], which are based on a number of different, uncorrelated and simple decision trees. Class assignment is done by majority voting on many well-behaved trees.However, this research focuses on combining the advantages of both types of classifiers, i.e., the high performance of the subsymbolic algorithms and the transparency and, hence, trustworthiness of the symbolic classifiers. It is readily possible to narrow down the number and type of features of the subject on which the decision is based and to rank their importance, but the exact process remains a black box. Further analysis is needed to uncover the exact decision process. For example, one method developed for random forests is to analyze representative trees in the forest [35] (Figure 2). The result is a small selection of well-functioning prototypic trees out of a total of 1500 trees in the random forest on which possible decision processes can be traced. An alternative method for extracting non-hierarchical decision rules from the decision process in random forests, and also in other subsymbolic classifiers, is the so-called LIME method (local interpretable model-agnostic explanations) [36]. This learns an interpretable model locally around the single prediction of a trained AI, e.g., a random forest. This is achieved by changing the assignment rules for a single data instance, e.g., a single patient, by changing the feature values and then observing the resulting impact on the classification. The result of LIME is a set of rules representing the contribution of each feature to a prediction for a single data instance, which is a form of local interpretability.The detailed understanding of the mathematical details of an AI algorithm may be possible for experts in statistics or computer science. However, when it comes to the fate of humans, this “developer’s explanation” is not enough. For example, the World Bank requires of AI systems for credit scoring “the ability of humans to interpret, understand, explain, and justify decisions made with methods that use a large number of variables” [41]. Ultimately, a human must be able to take responsibility for the consequences of an AI system’s decision.Linear models, especially structurally simple ones, are assumed to be understood by mathematicians, statisticians, or computer scientists (i.e., the developers). However, these models are limited in what they can do. The development of “parallel distributed processing” models has attempted to overcome the limitations of linear systems [42]. Such models consist of a very large number of nonlinear functions, often referred to as neural networks or forests of decision trees. By adjusting many parameters (e.g., “synaptic weights”), such a model can “learn” to reproduce given input–output situations. Due to the large number of interacting elementary processes (neurons), understanding the details of such a system in finite time is neither intended nor possible. Such systems are referred to as “black boxes”. The comprehensibility of the system is sacrificed in favor of efficiency and simplicity of development. For example, it took many hundreds of man years of acoustic and statistical specialists to develop the first speech recognition program [43]. Modern so-called deep learning neural networks require only computational power to optimize a standard algorithm to achieve even better quality [44]. The comprehensibility of such systems was traded for the capability (accuracy) of their performance [41].XAI requires knowledge discovery methods that are machine-usable and explainable to a domain expert or even a layperson. The most precise definitions of XAI [6] go back to research on knowledge-based AI [45,46]. A truly explainable AI (XAI) system is one that draws its conclusions based on a model that is understood and accepted in depth by a human expert in the field in which the XAI is used (domain expert). This understanding and acceptance of the AI inference model must be such that the expert is willing to ultimately assume legal responsibility for the AI’s decisions. Such XAI systems cannot rely on their skills alone. Instead, they must make their decisions using scientific logical reasoning based on recognized expertise. XAI systems must be able to explain each decision and its derivations in a way that can be understood and comprehended by the domain expert (domain intelligibility).Consequently, XAI systems must be based on (machine-processable) knowledge oriented to human language (symbolic systems). The so-called “expert” or “knowledge-based” systems [47] fulfill this requirement. They are typically based on a representation of the concepts, facts, rules, relationships, and theories in a given domain [48]. The GeneOntology knowledge base [24] is an example of such a knowledge representation in the field of cell biology and genetics. An XAI system arrives at conclusions (decisions/diagnoses) by applying formal methods of scientific reasoning, e.g., predicate calculus [49]. There are machine learning AI systems that use this type of reasoning, i.e., the individual decision steps are provided directly by so-called “symbolic” machine learning methods that base the class assignment of a case on a set of hierarchically or non-hierarchically organized rules [50,51,52,53]. Examples include hierarchical classification and regression trees (CART [13]) or non-hierarchical repeated incremental clipping for error reduction (RIPPER [54]). Among the symbolic tree-based algorithms, the so-called “Fast and Frugal Trees” (FFTs [55,56]) provide particularly simple decision trees, usually consisting of 1 to 5 pieces of information, which makes them particularly suitable for biomedical problems, as they mimic the processes of making a clinical diagnosis [57].However, one important requirement for machine-learned AI systems is often overlooked: the comprehensibility of the knowledge used in the system to a domain expert. An important requirement for comprehensibility is simplicity. Machine-learned symbolic systems often lack this property. For example, decision trees may consist of hundreds of conditions. Identical subtrees may be used repeatedly in different branches of such a decision tree (Figure 2). It is acceptable for a computer algorithm to base a decision on hundreds of conditions. Humans, on the other hand, have a limited capacity in terms of the complexity and redundancy of models or explanations. According to Miller’s law, the typical limit of human information processing capacity is 7 ± 2 elements [58]. XAI explanations must, therefore, be as simple as possible (Occam’s razor) and use abstractions (generalizations) from example situations.Approaches to explain the decisions of deep learning algorithms for biomedical tasks have their main focus on visualizing the elements that contributed to each decision [59]. For example, one of these methods is interactive heat maps [60]. There are several ways in which such mechanical explanations can highlight which input is relevant to an output obtained, using gradients as a multivariable generalization of the byproduct, where the neural network is viewed as a function and the explanation is based on the gradient of the function available from the backpropagation algorithm [59,61]. The volume of studies on machine learning interpretability methods in recent years demonstrates the growing interest in this research area. However, despite the rapid growth, the goal of understandability for experts (statisticians) is sacrificed for the understandability for professionals [62].An attempt was made to define the position of XAI in the biomedical context in general. The main goal of achieving explainability and traceability of machine-learning-based decisions is inherent, and a further breakdown was proposed based on a study of the terms frequently used in the XAI context [6]. Accordingly, XAI should serve the following goals, namely (i) trustworthiness, (ii) causality, (iii) transferability, (iv) informativeness, (v) trust, (vi) fairness, (vii) accessibility, (viii) interactivity, and (ix) privacy awareness.Machine learning (ML) methods for classification tasks decide which class (e.g., a clinical diagnosis) is appropriate for a given case. When a person’s fate depends on the outcome of such a decision made by an algorithm, the trustworthiness of the ML system is of particular importance.The term trustworthy AI is increasingly used as an alternative to the term XAI in clinical research and AI-assisted decision making when the concept of XAI is used in the context of patient–physician interaction. The idea behind trustworthiness of an AI is to gain the trust of individuals or organizations in the AI model by explaining the characteristics and reasons for the AI output, which helps to achieve the full potential of the AI. For example, if neither physicians nor patients trust an AI-based recommendation for a clinical diagnosis, it is unlikely that any of them will follow the recommendation [63]. A solution is provided by the above-mentioned LIME method. A recent example of its use was the transparent assignment to specific pain phenotype clusters based on random forest [37]. However, in the same report, it was also shown that the LIME method is not perfect, and rules can only be expected for a subset of the instances in the data set.The following explanations and examples attempt to capture more of the main goals identified in [6], with a focus on biomedical research and, in particular, clinical decision making based on AI or machine-learned algorithms (Figure 3).Transparency is also grouped under the terms black box to white box, intrinsic explanations, understandability, or comprehensibility. All of these terms refer to a precise description of the mathematical/statistical/algorithmic details of how the AI model works internally [6]. This type of explanation may be understandable to statisticians, mathematicians, and/or computer scientists. However, it is usually useless to the physician or the patient. In banking, for example, the “transparency” of an AI deciding whether a customer is eligible for a loan is based on the equation lnp(x)1−p(x)=β0+β1x1 [64]. The success of today’s “subsymbolic” AI is based on trading the understandability of the model for its performance in terms of the accuracy of the AI’s predictions or class assignments, as described in more detail below in the History section. Such models, e.g., artificial neural networks (ANN [65]) or random forests [33,34], use a large number of nonlinear functions as neurons or decision trees coupled by thousands of coupling factors, such as synapses and weights, respectively, which in deep learning can be organized in many layers. Thanks to the power of modern computers, up to current PCs, the many thousands or millions of internal parameters can be optimized for these models so that those functions can be approximated that map the high-dimensional multivariate input data to multivalued outputs, such as different diagnostic classes. In this way, the individual elements (neurons, trees) can be accurately described, but the resulting collective behavior of the system cannot be understood. This can be compared to the impossibility of explaining thoughts or ideas by the firing of neurons in the brain. In systems theory, this is considered one of the central properties of emergent systems [66,67,68].Being also referred to as interpretability, the comprehensibility of an explanation means the provision of a causal and logical deduction of the results (decisions) from the given facts (input data) using the terms, formulations and methods of decision making in the respective subject area. Comprehensible explanations are formulated using deductive logic, considering approximations (fuzziness) and risks. This is the very meaning of explainability of an AI [6] and is used as explainable AI (XAI) in the rest of this paper. A common definition of XAI is that it is AI in which the results and the derivation of the solution can be explained in a way that is understandable to humans. The term “explainable” is often used and defined very differently by researchers, as there is no concrete mathematical definition [69]. Most importantly, the human who needs to understand the AI’s decision is controversial: is it the patient, the doctor, or a computer scientist or statistician? Another problem is the interchangeable misuse of interpretability and explainability in the literature, as there are significant differences between these concepts, but in all existing definitions, the term “understandability” emerges as the most essential concept in XAI [6].However, there is a consensus that XAI must ensure that computational decisions are transparent so that they can be communicated to affected patients in an understandable way when it comes to biomedical and especially clinical decisions. In this regard, the goal of XAI research is to define the specific interests, goals, expectations, needs, and requirements for artificial systems and to drive their implementation.The informativeness of AI is required, for example, in clinical decision support systems that assist physicians in diagnostic or therapeutic tasks. Such systems are based on AI and are increasingly used in clinical practice, with a current focus on medical imaging. Alzheimer’s disease can be diagnosed from magnetic resonance images by training deep neural networks to identify abnormal brain regions [70]. Similarly, deep neural networks were introduced in clinical imaging to facilitate decision making using these data [71]. However, while the results of these analyses appear reasonable to a medical expert because they are consistent with medical knowledge and, as such, could be communicated to the patient, the exact mechanisms of how a diagnosis is made for a particular patient remain vague. Deep neural networks are subsymbolic classifiers in the above sense, as are random forests, whose hundreds or thousands of decision trees also cannot be grasped in full detail by the physician and communicated in an understandable way to the patient. Informativeness aims at simpler models of what an AI does internally such that this abstraction provides more information to a user [6].Accessibility can mean involving end users in the process of improving and developing an AI algorithm, as previously suggested [6]. Furthermore, accessibility can be considered the ability to make machine-learning-based decisions without deep programming and AI knowledge. This can be done via interactive pre-packaged software, such as the R package “rattle” used in the second chapter of this report or other interactive tools such as the R libraries “AdaptGauss” https://cran.r-project.org/package=AdaptGauss, which provides interactive fitting of Gaussian mixture models [72], “pguIMP” for preprocessing biomedical laboratory data sets, including imputations of missing values by machine learning (https://cran.r-project.org/package=pguIMP [73], all accessed on 15 December 2021), and many others that would require separate review exceeding the present scope.However, accessibility can also be understood in terms of intellectual accessibility. This goal is often realized in symbolic forms of AI, often in simple machine learning algorithms that take the form of hierarchical or non-hierarchical rules. A symbolic rule-based classifier that included 21 individual or aggregate parameters, including demographic characteristics, psychological, and pain-related parameters recorded early after breast cancer surgery, predicted the subsequent development of persistent pain with a cross-validated accuracy of 86% and a negative predictive value of about 95% when most non-hierarchical rules were used [74]. Another example from pain research is how subsymbolic AI can be subjected to further analysis to extract the individual decision process for each case, or how it can be complemented by symbolic AI that provides understandable explanations for group assignment, although the exact decision process may differ from that of subsymbolic AI [37]. This has been elaborated in more detail in a visual analysis system for multi-model comparison of predictions for clinical data [75]. The system allows comparison and evaluation of different AI models based on their interpretable information, with the goal of assisting clinicians in decision making. The different models are compared in terms of the predictive criteria used, and the consistency of their application is evaluated.Transferability was cited as the second most common reason for using XAI in research [76]. Transferability means that explaining how an AI model works serves to better understand the underlying problem so that the solution can be more easily applied to a different application or problem. In the breast cancer cohort mentioned above [74], the initial set of rules for predicting pain persistence included information collected from patients through repeated use of comprehensive psychological questionnaires. Once this proved informative, supervised machine learning could be used to reduce the questionnaires to items relevant to the pain context. This was accomplished by creating a shorter form of questionnaires that contained only seven items, representing 10% of the original psychological questions, but yielded the same predictive performance for pain persistence as the full questionnaires [77]. Certainly, this short questionnaire is much easier for clinicians and patients to understand than the more general full questionnaires.Computer science is described as a rapidly growing multidisciplinary field that uses advanced computing capabilities to understand and solve complex problems [78]. It inherently requires collaboration that involves sharing and collaboration on information and methods between professionals of different domains, such as physicians and machine learning experts [79]. This sets XAI, i.e., explaining AI to non-mathematicians, into the center of the classifier development workflow rather than placing it at its end. Concerted model building, involving a variety of experts from different fields, is necessary to identify and eliminate machine learning pitfalls, such as confounding variables and surrogate markers, commonly referred to as data-leaking covariates. An example of this is the identification of a protective effect of the 5-HT3 serotonin receptor antagonist ondansetron, an antiemetic routinely used to treat nausea and vomiting, against hospitalization-related venous thromboembolism [79]. Whereas an initial classifier achieved a ROC-AUC for risk prediction of 0.92, after a concerted effort by biomedical and computer scientists to exclude data-leaking covariates such as specific pharmacological prophylaxis or treatment of thromboembolism, the ROC-AUC decreased to 0.87, which seems to more realistically capture the benefit of ondansetron in this context. A purely statistical approach without consulting biomedical expertise may not be sufficient here, as it has been emphasized that it is often difficult to distinguish between confounding and mediating variables in statistical analyses [80]. It appears that expertise is required to deal with confounding variables so that an expert can decide which variable can potentially be considered a confounding variable (rather than a mediating variable) or a surrogate marker [81]. As stated elsewhere, prediction only requires correlation, but understanding requires significant knowledge underlying the causal mechanisms [79,82]. Other problems in machine learning model building include the shortage of data points relative to the number of available variables to select from and sparse data sets where many of the labels are missing. Again, one way to address these shortcomings is proposed to consist of the involvement of domain human experts in various steps of data set construction, model training and evaluation and, in particular, the integration of prior medical knowledge [83]. An example for these effects is the identification of olfactory effects of various drugs from a data set with many candidate drugs applied to a limited number of patients, which has also been assessed by both biomedical and computer science experts [84]. It is noteworthy that this report is also the result of a collaborative project between authors whose original fields of study are medicine/data science, biology, or computer science.Artificial intelligence and its most popular application, machine learning, increasingly permeate many areas of daily life and science, including biomedical research. An automated search of the PubMed database on 25 September 2021, using the R library “RISmed” (https://cran.r-project.org/package=RISmed [85], accessed on 15 December 2021) yielded 166,938 hits with the search terms (“machine-learning OR artificial intelligence OR explainable artificial intelligence”) and 138,556 hits with the search terms (“artificial intelligence OR explainable artificial intelligence”). When excluding reviews by adding “NOT review[PT]” to the search terms, 153,868 and 127,438 hits were obtained, respectively. The earliest hit using the MeSH term “artificial intelligence” was a 1951 report on a neurological research robot [86]. Publications per year were infrequent until the 1980s and did not exceed 100 per year until 1986 (Figure 4A). Since then, publication activity has accelerated and reached a temporary peak in 2020, when the above searches, which included all types of publications, yielded 25,622 and 19,302 hits, respectively.In biomedical research, the concept of XAI was only recently mentioned in publications. XAI accounts for only a small portion of the hits in the second search above. An automated search of the PubMed database as above, using only the term (“explainable artificial intelligence”), yielded 340 hits on 25 September 2021, with the earliest publication dating from 1990 [87]. However, XAI is increasingly included in publications, and most publications are from the last three years, with 113 articles from 2020 and 172 articles already from 2021 (Figure 4B), which fits well with the publication dates of the seminal articles mentioned in the above chapters on concepts of AI and XAI.The purpose of artificial skill-based (AS) algorithms is to use examples to learn how to classify (diagnose) cases in such a way that this can be generalized to unseen cases. This is akin to teaching a child to ride a bicycle in a parking lot with the expectation that he or she will later be able to ride on the street. This is ideal for application areas such as drug repurposing or protein secondary or tertiary structure prediction (for a summary, see, for example,  [88]). Deep learning neural networks are the prototypical example of this type of AI. Skill-based algorithms can surpass the current state of the art in in patient categorization. However, they do so by intentionally sacrificing explainability. For those application domains that target performance, this may be appropriate, e.g., for purely technical applications such as AI-based detection and separation of cell types, which are often implemented in close proximity to the laboratory equipment used to collect or generate these biomedical data. The literature does not cover this aspect of different types of application domains: aiming for skill and performance versus aiming at knowledge and explainability.However, where the decisions made by AI are relevant to people’s lives, knowledge-based AI should be used. For this type of application, the AI’s decisions must be understandable to the medical or other professional, and the application of AI methods in the medical field should be limited to user-understandable systems. These are models that provide a causal and logical derivation of their decisions from the given multivariate data using the terms, formulations, and methods of medical decision making. This means that such systems should use a formal, i.e., understandable to humans (subject matter experts), knowledge representation. This is the viewpoint taken here for XAI. So-called white-box explanations, which provide intrinsic explanations, are left to mathematicians, statisticians, or computer scientists who deal with the internal workings of the models. XAI should focus on ensuring that computational decisions are made transparently and in a form that can be communicated to medical staff and patients in an understandable way. This approach is likely to accelerate the adoption of XAI in biomedical research and subsequently in clinical practice. This type of XAI can be implemented in a variety of ways, such as transforming subsymbolic AI into symbolic systems using knowledge discovery methods. XAI is an active research topic in computer science. Because of their direct impact on the realization of patients’ right to informational self-determination, their results have a direct impact on biomedical research and clinical practice and are rapidly being transferred from the field of theoretical computer science to practical applications in clinical work.The present XAI approach may differ from alternative approaches in that we explicitly define XAI as an algorithm that makes the decision as to why a particular individual should be assigned to a particular class (diagnosis) in a manner that is accurate and logically comprehensible to those involved. The steps of the decision-making process should be accessible and understandable at least to the expert in the field, who can then explain these rules to the affected individual. Ultimately, it should also be possible for the person affected to directly understand the decisions made by the system. This explicitly goes beyond making the decision-making process understandable to the data scientist who knows, for example, the mathematical background of a regression-based classifier. This background, expressed in equations, can hardly help most patients understand why an algorithm assigned them a particular diagnosis. It also goes beyond the mere plausibility of the features used to make the decision, which may allow a vague association with the classification but not the precise reasons for an individual. It also goes beyond the claim that it is sufficient for an XAI to meaningfully connect the input space, i.e., the available biomedical information, with the output space (clinical diagnosis). In contrast, the exact decision-making process must be made transparent to a medical professional for an algorithm to qualify as an XAI. As outlined above, XAI enables humans to get “into the loop” of machine learned systems, for example, to decide which variables can potentially be considered “surrogate markers”. Such markers are effective in predicting the diagnosis because their values are a consequence of the diagnosis. Such XAI can than contribute to (i) the understanding of the mechanisms of a particular disease and (ii) building trust that the results found by machine learning are not spurious [82].Although the present report emphasized the need for comprehensibility of AI-based biomedical decisions, it should not be ignored that it falls short to require interpretability only from statisticians involved in the medical decision-making process. Biomedical terms and methods may be similarly incomprehensible to a nonbiomedical expert as AI-specific terms and methods often are to the biomedical expert. Although the medical environment is the medical expert’s home professional field, with the increasing use of AI in the field, it is not enough to ask incoming disciplines to explain their methods without viewing this task as reciprocal, including the need for both informaticians and medical professionals to learn about each other’s disciplines. It is, therefore, the joint responsibility of biomedical and informatics experts to establish a common basis of terms and concepts for discussion, which each expert can then explain to the other expert and both experts to the patient. To return to the introductory example, CD19 is probably as unfamiliar to a computer science expert as SVM is to a medical professional. Both have the task of making themselves understood by the other expert and passing on their mutual understanding to the patient.A.U.—Conceptualization of the project, literature recherche, and writing of the manuscript. D.K.—Writing of the manuscript, literature recherche. J.L.—Conceptualization of the project, literature recherche, programming, writing of the manuscript, data analyses and creation of the figures. All authors have read and agreed to the published version of the manuscript.This work has received no external funding.Not applicable.Not applicable.Not applicable.The authors have declared that no competing interests exist.Classification performance of two different types of classifiers, comprising hierarchical decision rules implanted as classification and regression trees (“rpart”) and hyperplanes as used in support vector machines “ksvm”. Panel (A): Receiver operator characteristic of the two classifiers for the classification of cell samples as obtained from healthy subjects (class #1) or subjects with leukemia (class #2). The figure corresponds to the original output of the “rattle” R package with the curve for rpart (= CART) composed of only 3 points since a single decision rule was used in just one iteration for the present demonstration purpose. (B): Decision rule by which the hierarchical classifier made the assignment to class #1 or #2. (C): Schematic drawing of an SVM decision hyperplane between healthy and diseased samples. For illustrative purposes, the number of data points is reduced to n = 200, and the figure is purely schematic, without performing real calculations and SVM training. In contrast to panels A and B, which show results of computations, this is a schematic drawing. The plots were created using the R software package (version 4.1.2 for Linux; https://CRAN.R-project.org/ [17]) and the R library “rattle” (https://cran.r-project.org/package=rattle [11,12], and the vector drawing software “Inkscape” (https://inkscape.org/de/ [18], all accessed on 15 December 2021).Example of an attempt to make subsymbolic classifiers transparent in terms of the decision structure along which class assignment occurs by extracting well-behaved and representative trees from a random forest classifier. A subsymbolic random forest classifier with a size of 1500 trees was created that contained up to d = 3 pain-related variables by setting hyperparameters, using the R library “randomForest” (https://cran.r-project.org/package=randomForest [26], accessed on 15 December 2021). The pain-related variables are from [37] and consist of thresholds for different stimuli recorded in quantitative sensory tests in a study of experimentally induced pain in humans, namely, pain thresholds for noxious heat and cold. The variables used included heat pain thresholds (HPT) and cold pain thresholds (CPT). The pain data included the z-transformed pain thresholds for heat or cold stimuli recorded under control conditions and after UV-B irradiation, and the UV-B effects recorded as the difference between the z-transformed thresholds (zHPTbaseline, zHPTUVB, zCPTbaseline, zCPTUVB, UVBEffHeat, and UVBEffcold) acquired from 84 healthy subjects. Analysis of representative trees in the forest resulted in the four trees shown in panel C. This analysis used the trained random forest and the data to run predictions while identifying representative trees based on the d2 metric [35] using the Euclidean distance. The result was trees number 732, 905, 913 and 1070 of the 1500 trees in the forest. These calculations and plots were performed using the R libraries ”reprtree” (https://github.com/araastat/reprtree/blob/master/R/ReprTree.R [38], accessed on 15 December 2021) and “ggraph” (https://cran.r-project.org/package=ggraph [39], accessed on 15 December 2021). The figure shows the representative trees, with the class assignments as colored leaves at the respective bottoms. The figure was created using the R software package (version 4.1.2 for Linux; https://CRAN.R-project.org/ [17]) and the R package “ggplot2” (https://cran.r-project.org/package=ggplot2 [40], all accessed on 15 December 2021).Schematic representation of achieving trustworthy AI in biomedicine, including clinical decision making, and the requirements for such XAI in this environment. The left part shows paths to trustworthy AI. The AI-based decisions can be implemented as symbolic algorithms, which often use rules or small rule sets for classification, or a sub-symbolic and often more powerful type of machine learning algorithm is used, to which further methods are subsequently applied, such as local interpretable model-agnostic explanations (LIME [36]) to extract comprehensible rules for class assignment. The right part shows the main objectives assigned to an XAI in the biomedical and clinical context, as proposed in [6], with the main goal of making AI-based clinical decisions trustworthy by being comprehensible to both the physician and the patient. The figure was created using Microsoft PowerPoint® 365 (Redmond, WA, USA) on Microsoft Windows 11 running in a virtual machine powered by VirtualBox 6.1 (Oracle Corporation, Austin, TX, USA) on a computer running Ubuntu Linux 20.04.03 LTS 64-bit (Canonical, London, UK).Stacked bar chart of publications listed in PubMed per year, with particular emphasis on publications found with the search term ((“machine-learning OR artificial intelligence OR explainable artificial intelligence”) NOT (review[PT])), with the proportion of publications found with the search term (“explainable artificial intelligence” NOT (review[PT])) separated as blue parts of each bar. (A) Publications per year were infrequent until the 1980s and did not exceed 100 per year until 1986. (B) Enlarged view of the latter search, i.e., for “XAI” only. The figure was created using the R software package (version 4.1.2 for Linux; http://CRAN.R-project.org/ [17]) and the library “ggplot2” (https://cran.r-project.org/package=ggplot2 [40], all accessed on 15 December 2021).Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00002.txt

@@ -0,0 +1,56 @@
+The aim of this study is to develop a reliable 5D (x, y, z, time, flow dimension) model for medical decision making. Sophisticated techniques for the assessment of serious stenosis were developed using time-dependent instantaneous pressure gradients through the aorta (flow rate, Reynolds number, velocity, etc.). A 74 cardiac MRI scan and 3057 scans were performed on a 10-year-old patient with congenital valve and valvular aortic stenosis on sensitive MRI and coarctation (operated and then dilated) in the sense of shone syndrome. The occlusion rate was estimated to be 80.5%. The stenosis area was approximately 15 mm long and 10 mm high. The fluid solver (NS) exhibited a significant shear stress of −3.735 × 10−5 Pa within the first 10 iterations. There was a significant drop in the flux mass of −0.0050 (kg/s), as well as high blood turbulence in vortex field lines and low geometry Reynolds cells. The fifth dimension was used for negative velocity prediction (−81.4 cm/s). The discoveries of the 5D aortic simulation are convincing based on the evaluation of its physical and biomedical features.Visualization is crucial for the display and understanding of medical image data. For diagnostic and surgical planning, radiologists and surgeons must be able to evaluate the data appropriately. Many imaging systems’ data can incorporate both functional and structural information, resulting in 4D datasets. When the image contains spectral information, it can be extended to 5D in some circumstances. Overall, 5D imaging reveals more information than 4D imaging. However, there are various ways to visualize 4D medical data, visualizing 5D medical data. The inability to properly portray 5D medical datasets on a 2D screen has drawbacks. Currently, there are five generations of visualization techniques [1] in medicine: 1D waveform display, 2D sequence display, 3D dataset visualization, multidimensional dataset visualization, and virtual reality visualization. Dynamic volume datasets, such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) images, are multidimensional medical images that incorporate four dimensions: three for space and one for time. Five-dimensional medical images were created as a result of advancements in medical imaging methods; incorporating spectral information into an image collection is a good example. The representation of 5D medical data on a 2D screen poses significant hurdles. This study [2] used 5D brain electrical impedance tomography (EIT) images as a case study. However, the approach presented in this study can be applied to different imaging technologies. EIT is a relatively recent approach to medical imaging. EIT is based on the fact that different tissues have varied impedances. By injecting a tiny current via sensors encircling the region to be photographed, EIT imaging uses this attribute (i.e., impedance). EIT imaging is a low-cost, safe, and portable option. EIT offers great temporal resolution, but poor spatial resolution compared to other functional imaging methods. EIT is not currently used in ordinary biomedical practice for any reason; however, it was used in investigations to measure heart function, pulmonary hypertension, regional lung function, brain function, breast cancer, and gastrointestinal tract function. The initial generation of EIT imaging used a current supplied at a single frequency of approximately 50 kHz to monitor impedance changes over a few seconds or minutes. EIT may be performed at many frequencies simultaneously because various tissues have different spectral qualities. This has the advantage of allowing for more accurate tissue identification.The advanced EIT hardware can measure up to 30 frequencies at the same time [3]. When multi-frequency EIT hardware is used to track impedance changes over time, 5D EIT scan data are produced: three for space, one for time, and one for frequency [4]. The dataset’s dimension (S) is the sum of two portions to better comprehend the relationship and differences across multi-dimensional data visualization in various fields. Variables corresponding to these m dimensions are location variables, which are independent of one another, in which m dimensions constitute an m-dimensional space where samples are taken. The additional n dimensions are either n-dependent variables (depending on location variables) measured in the m-dimensional space, or n-independent variables if there are no location variables. All five dimensions (space + time + frequency) are location variables in a 5D EIT dataset and are utilized to locate the measured property impedance. An animation of a tomographic picture matrix can be used to show this type of dataset. However, given the difficulties of reading 4D data shown in this manner, doctors are likely to have much greater trouble interpreting 5D data using this strategy. For a dataset with more than three location variables, however, it is always preferable to provide three space dimensions together, as they would be in the actual world, and then present further dimensions based on this space.A revolution in dimensional processing was recently observed in medical imaging. The basics of 5D imaging vary depending on the scientific field [5]. Table 1 shows the published 5D imaging works based on various techniques as well as related work that explains 5D modeling in various imaging types. The fifth is yet to be defined and united in its meaning. The coordinate system and the notion of the 3D model essentially agree upon (x, y, z). The fourth dimension is described in a variety of ways, depending on the topic of study and how it affects the outcomes, but the time dimension is the most common. As demonstrated in Table 1, the complexity of their characterization is reflected in the diversity of the fifth dimension.In this study, we developed the concept and modeling of the 5D cardiac system (3D model + temporal dimension functional dimension of the flow) detailed in our research work mentioned in [13,14]. This strategy consists of reconstructing a 3D geometry of the descending aorta and diffusing a 2D viscous laminar fluid to detect the zones of narrowing. Blood flow is obstructed at the location of the stenosis in the restricted aorta, restricting the movement of blood cells and creating turbulence in the internal aortic wall. However, blood movement is affected by several other factors, including the thermophysical properties, including viscosity, surface tension, and wettability [15]. Over the last few decades, numerous statistical models of blood rheology have been researched, and various secret features, such as fluid behavior and flow behavior, have been uncovered as a result of these experiments [16]. Consequently, assessing the significance of wall pressure and wall strain stress in arteries is crucial for medical researchers. Pulsed flow behavior was documented by several scholars, and its variations are constantly damped, which can be due to the elasticity of blood vessels. Hemodynamics are affected by blood vessel disease because they disrupt the flow cycle, resulting in a reduction in the wall pressure and shear tension in the arteries [17]. The biochemical mechanisms of blood-related illnesses are the subject of several studies. However, understanding the fundamental physics of the disease in order to understand the mechanism and hence pave the way for less invasive and more sustainable strategies for their prevention is an enormous development. The parietal strain and parietal tension exerted by blood on the inner periphery of the artery are not recognized by doctors using current imaging techniques [18]. The breakthrough in anatomy and cardiovascular physiology physiological simulation has created an opportunity to bridge this gap. Recent advances in the field of computational fluid dynamics (CFD) have made it possible to model blood flow in heart structure geometry. It is a less invasive method, and a device may interpret the blood flow pattern of a disease-related artery. As a consequence, in the areas of congenital heart valve, coronary, myocardial, and peripheral artery disorders, CFD has become a clinical testing instrument for medical practitioners. The CFD module, on either hand, includes non-clinical experience, technical software, efficient computer systems, and a vast number of calculations because it depends on the exact requirements of geometric and flow boundary conditions [19]. It is necessary to implement CFDs in standard clinical procedures, owing to these conditions and restrictions. Viscous dissipation can also be measured using the viscous term of the Navier–Stokes equation, which removes the need for friction and relies solely on internal blood flow velocities, which can be measured noninvasively using 4D flow MRI [20].Our main contribution is to present a numerical simulation of laminar blood flow in 3D aortic modeling in the presence of a left subclavian aortic coarctation, and an analytical study is conducted to study the impact of a solver of dynamic fluid on the detection of aortic stenosis.Fluid flow simulations are based on Newtonian and fluid property physics. Physical rules, such as the maintenance of kinetic energy, which contribute to the equation of motion, must be satisfied by these features. The viscous stress tensor is linearly related to the rate of the strain tensor in a Newtonian fluid. Assume Stokes flows (low Reynolds number) and only slight spatial variations in hydrostatic pressure [21].Let Ω ⊂ R2 be an open domain bounded and connected to the Lipschitz boundary Γ. Consider the Navier–Stokes equation
+        (1)∂v∂t−νdiv ∇v+∇v T+v · ∇v+∇ p=f in Ω
+        (2)div v=0 in ΩWith the initial condition v (0) = v0, where v is the velocity, p is the pressure, ν is the viscosity (or the inverse of the Reynolds number, i.e., ν = 1/Re), it is an external time-dependent body force when the Reynolds number approaches a critical value or minor fluctuations are implemented into the flow. It is well understood that the flow transitions from laminar to turbulent [22].The fluid was intended to be incompressible in this simulation, which meant that the density had to be steady and Newtonian. The vessel wall had to be impervious to deformation. A volumetric flow waveform was defined at the inlet of the pulsatile flow. For the entry of this model, a fully defined laminar velocity profile was used. The blood fluid was set to have a density of 1056 (kg/m3) and a viscosity of 0.06 (Pa). These fluid properties set the Womersley parameter (α), which measures the frequency of the pulsation, defined as [23]:(3)α=Dωυ−0.52
+        where ω is the angular frequency, and υ is the kinematic viscosity of the fluid. Properties of the fluids are mentioned for this model as well as the boundary conditions.In laminar flow regimes, the related term of the Navier–Stokes equation can be used to measure the volume of the viscous dissipation process [20]
+        (4)ΦvD=12 μ Σi Σj ∂v∂xi+∂v∂xj−23Δ·νδij 2
+        where ΦvD is the viscous dissipation per unit volume based on the viscous dissipation term, and μ is the dynamic viscosity. δij = 1 for i = j and δij = 0 for i ≠ j, where i and j are the principal directions (x, y, z) in Equation (4), which consists of the elastic viscosity of the velocity field and the spatial derivatives. If the speed field is defined, such as from 4D flux MRI measurements, Equation (4) can be used to quantify the viscous dissipation per unit volume. The integral of the viscous dissipation of the unit was used to measure the overall viscous dissipation.
+        (5)∫ ΦvD dv=Σi−1numvoxels ΦvDviThe condition was extended to the fluid model, and the vessel walls were considered to have a slip-resistant boundary. The porosity of vessels is often underestimated. As a result, the fluid is not supposed to move through the vessel walls as it passes through it. The boundary conditions applied to the vessel walls are [23]:(6)ui=uj=0There must be no increase in velocity in the steering radial along the line’s equatorial plane, essentially reducing the radial portion of the velocity to zero. Consequently, the applied boundary condition is
+        (7)ui=∂ui ∂xi=0The incompressible Newton flow rate through the established geometry is depicted by a parabolic velocity profile. As the Dirichlet boundary state at the output, a variable and spatially uniform pressure boundary condition is applied, and for the limit condition, the velocity Neumann (zero gradient of velocity in the axial direction at the output) is applied [24].The wall of the aorta was rigid. Consequently, the walls of the flow domains were static and stationary. Boundary conditions for speed were imposed on the wall with no slip or flow. The non-flow condition means that the velocity in the normal direction of the wall is zero because the walls are impervious. The pressure gradient perpendicular to the wall was considered negligible using the Navier–Stokes equations [25].Blood is a vigorous medium because its viscosity varies as a feature of the shear deformation rate, making it behave like a non-Newtonian fluid. Blood is biologically smoother in the systolic crest than when it moves slowly, as in the diastolic crest. This occurs as red blood cells clump together. The impact of thinning blood shear is as follows [16,23]:(8)μ γ=μx+μ0−μx1+mγa In this equation, μ0 is the viscosity of blood at zero strain rates, and μ∞ is the viscosity of the blood at an infinite or very high strain rate. The constants m, n, and a were experimentally determined.The sensitivity of a fluid to deformation under shear stress is known as viscosity. This is a type of fluid “friction”, which explains the internal resistance of the fluid to the flow. The viscosity of a fluid is strongly influenced by the bonds between the molecules. The viscosity is mathematically defined as the ratio of the shear stress to the velocity gradient. The majority of fluids are Newtonian fluids, which have a steady viscosity. Plasma, blood cells, and other substances carried in the blood make up the blood. The number of particles in the plasma induces non-Newtonian activity in the blood, which means that the viscosity varies with the flow shear rate. The blood flow exhibits Newtonian flow behavior when the shear rate is sufficiently high. In normal situations, however, it is not possible to disregard the fluid’s non-Newtonian behavior [26].In the “no fall” condition placed at the boundary, fluid in contact with the vascular wall can pass at the same pace as the wall. This is because the shear forces imposed by the wall on the fluid will ultimately cause the boundary flow to have the same velocity as the downstream surface, which is a rational statement in this analysis. Meanwhile, to preserve mass equilibrium, the flow away from the vessel wall continues to speed up until it reaches an established profile, where no further velocity profile changes occur in the flow direction. This causes a non-zero speed gradient, ∂u ∂y, where u and y are the velocity component in the direction of flow and the coordinate of the space perpendicular to the direction of flow, as shown in Figure 1.Mathematical relationship between shear stress of the wall, τ, and viscosity, μ blood, as explained in Section 2.3.1, is indicated in Equation (8) [23]:(9)τ=μ∂u ∂yIt was shown that achieving this quantity using experimental methods and relying on the experience of the fluid’s viscosity and velocity profile near the vessel wall is challenging. This sum is easier to approximate using CFD techniques, but it is based on mesh consistency.The governing equations for simulating the hydrodynamic flow of blood through a stenosing aorta with sufficient boundary conditions are as follows [16]:(10)∂ui ∂xi=0Laminar pulsed flows describe the boundary conditions at the entrance of the stenosing aorta. Because all of the geometries’ inputs are circular, a speed limit is enforced to introduce spatial and temporal differences in the pulsed flow. Patients with aortic stenosis are most likely to have valvular stenosis. This study aimed to determine the effect of physiological conditions on the rupture wall of the stenosing artery.Laminar and turbulent flows are the two main types of flows that arise. Turbulence is a measure of the degree of oscillation of the surrounding fluid in the direction normal to the fluid flow, and it can be made on the basis of both the energy (kinetic energy of turbulence) and energy dissipation (dissipation turbulence rate) included in this motion. There is also little distinction between laminar and turbulent flows in terms of fluid particles. In chaotic flows, the particles appear to behave spontaneously, which is similar to the initial flow direction on average. The average fluid velocity, density, viscosity, and vessel diameter all play a role in determining when the fluids become turbulent (for internal flows). The Reynolds number Re is the name given to this characteristic value, as shown in Equation (12):(11)Re=UDv=ρUDμ
+          where U is the average velocity of the fluid, D is the diameter of the vessel, ρ is the density of the fluid, μ is the fluid viscosity (dynamic), and ν is the kinematic viscosity of the fluid. The fluid tends to become turbulent at a Reynolds number of approximately 2000 for internal flows [27].The laminar and turbulent flows have different profiles, as shown in Figure 2.It is extremely unlikely that blood can enter turbulent flow under natural circulation conditions. Only severely stenosed vessels, extremely abnormal flow paths, and even certain pathological disorders that impair blood flow properties cause this. Non-Newtonian fluids have a flatter profile toward the axes, so the formed profiles are different. Blood appears to have a uniform velocity profile rather than a developed profile because of the diverse flow paths and irregular sizes of arteries [28,29].The amount of cross section that is “blocked” by occlusion determines the degree of stenosis. The narrowest segment (called “groove”) of the blockage is usually considered the cross section for stenoses with variable geometry. Figure 3 depicts a common model of a tube with a single stenosis in which the original radius of the vessel R0 is obstructed by a stenosis with a normalized maximum height of value = h/R0, where h is the maximum height of the stenosis. To obtain the unit length, the available radius, R0 = 1. Because the standardized height of the stenosis must be 0 ≤ δ <1, we can deduce the following [23].The normalized cross section of light through CD = πThe normalized cross-section of light on AB = π (1 − δ)2Thus, the percentage of blocked area can be derived; therefore, the severity of the stenosis is obtained as follows:(12)π−π 1−δ2 π×100% =(1−1−δ2 )×100%In this section, we apply our solution’s architecture and general technique, which includes the use of the Navier–Stokes equation in 2D to determine the action of blood fluid in a 3D model. Figure 4 shows the steps of each component of the solution, as well as an overview of the advantages of using CFD modules and the methods used. Our benchmark is composed of three parts, as indicated in Figure 4:1—3D Medical Image Processing Software and Mimics Innovation Suite and Materialize for export, import, and management of 3D modeling of the stenosing aorta format;2—ANSYS Fluent for computational fluid dynamics (CFD) manipulation and fifth dimensional “flux” modeling;3—MR 4D Flow data for the visualization of aortic blood flow as streamlines, pathlines, and color-coded vectors with Pie Medical Imaging stands for expertise in cardiovascular quantitative analysis software (confirmed to have stenosis by human experts).Benchmark of 5D solution.The main purpose of CFD is to solve the problem by discretizing the paradigm into small cells. The CFD process is divided into three stages. The first phase is pre-processed. The domain of the study is defined and generated during this step, and the flux regions are discretized to form a mesh of cells. Better findings would be obtained with a finer mesh, but this would require more processing capacity. Several codes now have methods for adapting meshes after multiple iterations to obtain more precise meshes, if desired. The properties of the fluid, as well as any other phenomenon to be investigated, are evaluated at this level.It is essential to develop the required boundary conditions for the model until the medium begins, and the equations that need to be solved are determined. This is achieved by adding known or controllable values to the nodes or cells along the domain boundary. These limit values can be used to evaluate the solution for the remainder of the fluid velocity. It is projected that defining the domain geometry and creating the grid would take up more than half of the time spent on a CFD project. Although current software has made this move simpler, rational skills and knowledge are needed to build real-world meshes and provide correct answers without using more computing power than is appropriate. In the second step, the mesh generated in the previous step is solved. Assigning an initial value to cells first initializes the model and performs preliminary assessments of solution techniques. The final stage is post-processing, which involves extracting and analyzing the data produced in the previous step. Rather than exporting raw data, today’s commercially available applications aim to improve the accessibility of the findings by optimizing visual appearance and data manipulation. This provides a graphical user interface that can depict geometry and grid, vector displacements, contour and surface tracing, particle detection, and other features based on the needs of the user. This move often requires fundamental knowledge to understand these findings.To meet the increasing need for CFD in industrial applications, several businesses have created a variety of applications. To satisfy the existing demands of search and processing resources, the simulation capabilities as well as the user interface have been greatly enhanced.ANSYS (https://www.ansys.com/products/fluids/ansys-fluent, accessed on 21 November 2021 (Academic and Professional version of Enterprise)), which is still one of the most widely used commercial codes for structural research, has also played a key role in the establishment of a package in the market. The benefit of this program is that it allows multiple supported programs to run on a single platform and be connected to one another. This was initially designed to provide developers with a single board on which they could construct a product from the ground up, validate it, and refine it. This includes constructing the target model, discretizing it into finite elements, running the requisite models, supplying the required data, and optimizing the model based on the available data. The number of applications capable of running in this area grew in tandem with the market demand. Finally, in addition to basic structural models, a wide variety of mechanics are included, including fluid flow, electromagnetic, and thermal [30,31,32].We used the product of Pie Medical Imaging “Caas MR 4D Flow” (https://www.piemedicalimaging.com/product/mr-solutions/caas-mr-4d-flow/, accessed on 21 November 2021 (Private access to the user with a procuration of agreement for the use of the product)) to analyze the valve behavior, considering the difficulty of the reconstruction of the valve structure from the flow sections. This cardiovascular solution aims to analyze blood flow in cardiovascular structures. Blood flow can be evaluated using three-way phase-contrast MRI images and the corresponding MRI images. In addition, blood flow can be analyzed by retrospective reformatting of flow planes. Caas MR 4D Flow is software that allows a cardiologist or radiologist to imagine and analyze blood flow in cardiovascular systems using multi-slice and multi-MRI images. It uses MRI and speed-coded images to examine blood flow in the heart and large arteries, and it contains the following features: analysis of quantitative cardiovascular results, segmentation of cardiovascular structures, and simulation of the intensity and direction of blood flow.Whenever Caas MR 4D Flow findings are used in a clinical setting to sustain diagnosis, they should not be considered as the only compelling basis for clinical decision making.For a 10-year-old patient with a history of congenital valve and valvular aortic stenosis on close MRI and coarctation (operated and then dilated) in the sense of shone syndrome, 74 cardiac MRI scans and 3057 images were obtained. Technically, reconstruction and segmentation of the descending aorta were performed using the following:✓44 TRICKS angiographic slices in dynamic acquisition on the thoracic aorta;✓Injection perfusion sequences after injection;✓Sequences ciné-fiesta T2 short-axis 4 cavities;✓Subsequent infusion sequences short-axis.44 TRICKS angiographic slices in dynamic acquisition on the thoracic aorta;Injection perfusion sequences after injection;Sequences ciné-fiesta T2 short-axis 4 cavities;Subsequent infusion sequences short-axis.
+        Clinical Diagnostic Report of LV:
+      For the ascending thoracic aorta without parietal abnormality, its main measurements are as follows:✓Tricuspid valve. Aortic ring with 8.5 mm diameter;✓Aortic stenosis at 0.42 cm2 with reduction of sigmoid opening at 5 mm;✓At the sino-tubular junction: 25 mm;✓1/3 medium of the ascending aorta: 18 mm;✓Horizontal aorta: 16 mm;✓Size disparity with aortic stenosis at the isthmic level extended over 10 mm, reducing approximately 65% of its lumen by 6 mm in diameter;✓Mitral valve of normal diameter 2.5–4.3 cm2;✓The mass of VG tele-diastolic 90 g and tele-diastolic 70 g;✓The systolic ejection function, estimated according to the 82% contour method.Tricuspid valve. Aortic ring with 8.5 mm diameter;Aortic stenosis at 0.42 cm2 with reduction of sigmoid opening at 5 mm;At the sino-tubular junction: 25 mm;1/3 medium of the ascending aorta: 18 mm;Horizontal aorta: 16 mm;Size disparity with aortic stenosis at the isthmic level extended over 10 mm, reducing approximately 65% of its lumen by 6 mm in diameter;Mitral valve of normal diameter 2.5–4.3 cm2;The mass of VG tele-diastolic 90 g and tele-diastolic 70 g;The systolic ejection function, estimated according to the 82% contour method.
+        Features of TRICKS cuts and acquisition protocols:
+        
+          
+            ✓
+            Magnetic field: 1.5 Tesla;
+          
+          
+            ✓
+            Acquisition time: 1.2 s/Repetition time: 3.3 s;
+          
+          
+            ✓
+            Diameter of reconstruction of the cuts: 370 cm2;
+          
+          
+            ✓
+            Angle of acquisition = 30 degrees;
+          
+          
+            ✓
+            Acquisition matrix: 0/300/224/0;
+          
+          
+            ✓
+            Sections orientation matrix: −0.0393457\0.99917\−0.0105243\0.283665\0.00106985–0.958923;
+          
+          
+            ✓
+            Cutting position matrix: −1.15748\−152.042\358.794;
+          
+          
+            ✓
+            Number of time positions: 12;
+          
+          
+            ✓
+            Scanning Rentals: −43.95322037;
+          
+          
+            ✓
+            Space between pixels: 0.7227\0.7227 with allocation of 16 bits of memory.
+          
+        
+      Magnetic field: 1.5 Tesla;Acquisition time: 1.2 s/Repetition time: 3.3 s;Diameter of reconstruction of the cuts: 370 cm2;Angle of acquisition = 30 degrees;Acquisition matrix: 0/300/224/0;Sections orientation matrix: −0.0393457\0.99917\−0.0105243\0.283665\0.00106985–0.958923;Cutting position matrix: −1.15748\−152.042\358.794;Number of time positions: 12;Scanning Rentals: −43.95322037;Space between pixels: 0.7227\0.7227 with allocation of 16 bits of memory.In this section, we proceed to the first phase of our 5D model, which consists of defining the geometry of the descending aorta in 3D. This step states the reconstruction of the aortic model from the TRICKS sections, as shown in Figure 5.The multiplanar reconstruction in Figure 6 is a crucial step in the 3D reconstruction of the descending aorta presented in the axial plane in Figure 7.We applied a threshold to the areas of interest to define the active mask. It was applied based on a low limit and a high limit. The mask contained pixels with a value between two threshold limits. The upper and lower thresholds were limited to the maximum and minimum intensities, respectively. We present the concept of the generic algorithm in Table 2 as follows.Based on the detailed mathematical concept of [33], the thresholding is typically used in scans where a certain anatomy has a very distinct set of values (high contrast). Keeping only certain values of the medical image allows us to visualize the border of the aorta; in our case, the minimum and maximum threshold values were determined automatically. The threshold value for Mimi was approximately 270 (Hounsfield scale). A green mask was created after thresholding. The minimum threshold value for this case was 1235 (grey values) or 211 (Hounsfield values). Thresholding needs to be done before region growth since all previous work is lost after changing the threshold value.The updates according to the defined limits give a reconstruction of the model of the aorta in 3D based on the multiplanar reconstruction of the TRICKS cuts in Figure 8.The second step was to divide the single mask for the entire model into two separate masks. This tool allows easy and fast separation of anatomical parts, such as the anatomical model of the heart and aorta. Two masks with the region of interest must be created and used as inputs to divide the selected mask, as shown in Table 3.We chose for the geometry calculation the “optimal” quality, partly corresponding to the input mask and custom parameters. The last step of this section was to clean the waste generated by the 3D reconstruction, as well as the smoothing and refining of the aorta model presented in Figure 9.The best is that the accuracy of the three-dimensional reconstructions obtained depends essentially on the conditions and quality of acquisition of cuts, thickness of the cuts, distance between two successive acquisition slices, and the quality of the segmentation and the reconstruction algorithm used. The final geometry is shown in Figure 10.Based on a few statistical measures in Table 4, it can be seen that the thickness analysis presented in Figure 11 of the aorta can identify the stenosing portion of the aorta.Note that the local curvature, as shown in Figure 12, of the segment that contains the shrinkage takes negative values. At this stage, the surface deformation as well as the measure of local curvature can be useful indices for aid in diagnosis.Based on the study in Section 2.3.4, the height of the obstacle, h, is calculated in Table 5 with respect to the unobstructed radius, which gives δ in Equation (13) a value of the severity of the stenosis to each part of the segment of the aorta.In this section, it is crucial to prepare a proper format for the aortic geometry to proceed to the mesh operation. The 3D model was transformed into a geometric smoothing (IGS) format to divide the entire volume into facets that contain triangles, edges, and intersection nodes. This IGS export approach is used to process surface-based functional data, primarily to recreate functions from noisy observations. The proposed IGS method has a broad range of applications, especially in industrial settings. Based on the simulations, the IGS appears to be similar to other commonly used CFD processes. The mathematical modeling and simulation environment presentation is detailed by [34] with precision analysis for IGS models described by [35]. The second phase consists of decomposing the IGS model into three elements, the body, inlet, and outlet, as shown in Figure 13.The geometry of the stenosing aorta has the following characteristics: length along the x-axis = 3.6339 × 10−2 m length along the y-axis = 5.66596 × 10−2 m, length along the z-axis = 8.4559 × 10−2 m with a total volume = 2.5915 × 105 m3 number of nodes = 593,223, and number of elements = 3,168,417. The inlet edge field contains 32 faces with an area of 2.6563 × 10−4 m2.The outlet field contains 41 facets with an area of 4.0045 × 10−4 m2. The number of faces of the body (wall) is 449, with a total area of 5.6453 × 10−4 m2. The results of the mesh are shown in Figure 14.To solve the 2D Navier–Stokes equation, the configuration of this fluid solver is based on the pressure with absolute velocity formation in a stable time. Our fluid is blood with a density of 1056 (kg/m3) and a viscosity of 6 (kg/m3). The boundary condition must be zero for the aortic model. The execution of the laminar flow model defines the velocities and pressure in the viscous flow field (the Navier–Stokes equation) to solve the displacement of the internal facet of the 3D model of the aorta. During the resolution during the first 10 iterations, the velocity residuals along the three axes (x, y, z) are estimated in Figure 15.The dimensions of velocity and pressure are combined in the momentum that governs the fluid equations. The coupled and independent methods are the two primary methods for solving discrete time equation algebraic equations. The simultaneous solution of the velocity and pressure parameters characterizes a coupled system. However, it is not commonly used in technical problems, owing to its poor processing power and high memory requirements. Coupled methods are commonly used for calculating compressible flows, whereas separate methods are favored for calculating incompressible flows. Unlike a coupled solution, a single approach solves the velocity and pressure fields independently or sequentially. It has the benefit of reducing computer memory and processing time, making it more effective for studying incompressible environment simulation fluids, as in our case for aortic stenosis modeling [36]. The pressure analysis of the proposed model is shown in Figure 16.From this illustration, it can be concluded that the pressure at the entrance and exit is negligible compared to the body outside and inside the model of the stenotic aorta. The pressure intersection (Pa) between these two zones (inside and outside) of the wall at a critical point at a position of 15 mm shows a remarkable overlap between the red and blue parts, which explains why turbulence can occur because of the strength of the parietal wall exerted by the external and internal body. The location of blood flow in a case of aortic dissection with a complicated geometric feature, qualitatively and quantitatively, based on the evolution of vortex structures and their interaction in the narrowing region throughout a cardiac cycle can provide an index of the presence of stenosis from a 15 mm length of the aortic segment. A vortex field simulation in the wall of the internal aorta is shown in Figure 17 as well as the flow rate mass in Figure 18 to evaluate the vortex behavior.We deduced from these two results that at the position of 15 mm, there is a very high vortex magnitude that reaches (800,000 (1/s)) as well as a mass of negative fluxes during the first six iterations during the solver calculation. This reflux shows that it occludes a segment of the aorta. These indices support the results reported in the clinical assessment with isthmic extensor aortic stenosis extending over 10 mm in length, reducing approximately 65% of its lumen by 6 mm in diameter, compared to what we estimated to be a 15 mm wide aortic stenosis; a 5 mm error rate was detected.Owing to the very small size of the geometry, the Reynolds number is small enough for the laminar flow that appears throughout the aorta model, as shown in Figure 19. In our case of aortic coarctation, the highly disturbed flow areas exhibited Stokes flow characteristics. The friction factor, vortex length, shear stress, leakage flow strain, and turbulence results obtained with the low-Re turbulence model were compared with experimental data and findings achieved in terms of velocity profile, vortex length, shear stress, and turbulence.Because of the unfavorable pressure gradient produced in the expanding segment of the stenosis tube, downstream flow separation was observed, and its size increased with the Reynolds number. The wall shear stress along the stenosis aorta peaked at the stenotic throat, and the tip was notably lower for a longer stenosis. The critical Reynolds number at which the blood flow becomes transient or turbulent distal stenosis was accurately predicted using this model. More surprisingly, the vortex length measured with the formula (Low-Re) nearly matched the vortex length expected by laminar flow modeling on the Re spectrum of the laminar flow. The constant flux was caused by the low Reynolds number.The flow becomes unstable at a relatively high Reynolds number. Finally, the analysis clearly shows that the proposed model is appropriate for studying blood flow in specific occlusion zones [37].The average size of the stagnation and shunt zones was predicted by CFD, but the length of the current line and the variations in speed due to aorta flow were underestimated. However, for quantitative confirmation of CFD findings and the quest for flow effects, such as tortuosity and laminar flow behavior, the measurement accuracy must be increased. At almost the same time, the CFD simulation effects are represented not only in terms of pressure drop, but also in terms of vector orientation and velocity, current lines, and regions of blood flow stagnation [38]. This time-based flow concept is used in the 3D modeling of the stenosing aorta to generate the 5D format shown in Figure 20.Owing to the fiber of this model, the fluid flow is deflected, resulting in a sinuous orientation of the flow lines, which shows a three-dimensional flow around the wall of the aorta, mainly directed from the inlet to the outlet. The area of stenosis in the middle, indicated by lower speeds, is shown in Figure 21.Focusing on the velocity in the stenosing part, we noticed that the flux field lines were delayed and decreased from 6.034 to 1.207 × 101 to 1.810× 101 ms−1. Interestingly, the detection of negative pressure in the value narrowing zone was −3.735 × 105 Pa, as shown in Figure 22.Our clinical case of aortic coarctation led to narrowing of the aortic valve with shone syndrome (two single-pillar valvular leaflet). In this case, the 5D model for the valve structure is corrosive, given the morphological dissection of the valve leaflet. The definition of the inputs and outputs, as well as the conditions of the limits of the aortic valve, are complicated because the valve acts as a blood pump. For this reason, the use of this research product that enhances the fifth dimension of 2D blood flow is used to predict and simulate the occlusion rate at the aortic valve level. Caas MR 4D Flow allows examining the 2D blood flow by reformatting planes in a 3D volume retrospectively. Additional plans can be placed on the 3D volume’s middle side. Standard 2D flow parameters, such as the flow rate, forward flow, and back flow, can be measured using 2D flow analysis. In addition, an extreme eccentric flow measurement in a plane was calculated based on the method of moving the flow [39]. The contour is redefined on its initial contour, initially shown by the Caas MR 4D flow in Figure 23.The displacement of the flow is defined as the distance between the center of the light and the “center of velocity” of the flow, normalized with respect to the diameter of the light. The center of velocity (Cvel) was calculated as the average position of the pixels of the light (ri, where i = x, y, z), weighted by the velocity information (vi) as follows:(13)Cvel,j=∑iri,jvi∑ivi
+        where i represents the pixels within the vessel’s outline, and j represents the spatial orientation of the pixels relative to the location of the center of the vessel. This method is described in more detail in the literature [39,40]. The flow displacement was calculated for the maximum systolic phase only, as shown in Figure 24. The maximum systolic phase was determined for each transmitter plane and determined using the time slot at the top of the 2D graph. The maximum systolic phase was indicated as a marker of phase slider. A description of the measurements is presented in Table 6.Antegrade flow is used to quantify the blood pumped in the positive direction measured on the plane in a cardiac cycle, while retrograde flow is used to measure the amount of blood pumped in the positive direction on the plane in a cardiac cycle. The pump function of the valve was measured using forward flow to the back. The regurgitation fraction is the ratio of the backward flow (mL) to the forward flow (mL) and the area under the curve of the negative (backward flow) portion of the cardiac cycle. The gap between the vessel’s axis and the center of the eccentric flow was normalized to the total size of the vessel to enhance the heat transfer displacement [39]. It is possible to draw from this section the interest of studying the fifth dimension of blood flow to quantify and estimate both the regurgitation rate as well as a simulation of the minimum negative sign velocity of −81.4 (cm/s) to predict stenosis at the level of the aortic valve.A medical decision support method for cardiac imaging in MRI focused on 5D modeling (3D anatomical structure, temporal dimension, and blood flow dimension) for the study of a promising case of aortic coarctation with extreme valve narrowing was described. Our object of interest was a segment of the descending aorta for prediction, identifying the presence of aortic stenosis. The results show that in the first stage, the 3D modeling provides a very interesting index for the experts, which makes it possible to estimate the occlusion rate of 80.5% compared to what was manifested in the clinical assessment with a rate of 82%. The measurements extracted for the thickness and the local curvature with respect to the geometry of the aorta mark the zone of stenosis and the degree of deflection of the narrowing portion. In the second stage, the presentation of the 5D approach was performed through a combination of the 3D model and the size of the circulating blood flow as a function of time. The experiments that were performed for the fifth dimension provided high accuracy for the location of the stenosis zone of 15 mm in length compared to the clinical prognosis, indicating that aortic stenosis is extended over 10 mm, which is deduced from the solver of fluid (NS). During the first 10 iterations, a significant decrease in the flux mass was reported with −0.0050 (kg/s), as well as high blood turbulence in vortex field lines and low geometry Reynolds cells, which is based on the understanding and observation of a negative pressure value of −3.735 × 105Pa. For negative velocity recognition (−81.4 cm/s), the fifth dimension was managed separately to assess velocity at the aortic valve with shone syndrome.An approach for diagnosing medical decisions using cardiac imaging with MRI was described previously. This method necessitates the creation of a 5D model, which is composed of five dimensions: the anatomical structure of the heart in 3D, temporal dimensions, and a functional dimension of blood flow for the diagnosis of valve stenosis. A comparative study of practical technologies that leverage the fifth dimension of flow for the derivation of medical inference in clinical routine was also included. This contribution was studied for an aortic stenosis, and it comprised creating a 3D model and solving the Navier–Stokes equations for laminar and viscous blood fluid to arrive at the proposed 5D model (3D + time + flow). The region and degree of stenosis can be classified by extracting measures (vortex field, flow masses, static pressure, and Reynolds number) based on the fifth dimension. With the increasing need for high-resolution simulations, it is critical to investigate the cost and reaction time of digital solvers that could benefit from recent architectures, including multicore processors, in the future.All authors have read and agreed to the published version of the manuscript.This research received no external funding.Not applicable.Not applicable.The data are private to the Carthage International Medical Center https://www.carthagemedical.com.tn/en/accueil/, accessed on 20 December 2021.We would like to acknowledge the Carthage International Medical Center that supported this work and the medical staff for providing the blind data used in this study.The authors declare that they have no conflict of interest.Modeling of the stenosing aorta.Velocity profiles for laminar (a), turbulent (b) and uniform (c) flow [9].The geometry showing the height of the stenosis, δ, initial radius of the vessel, R0 [9].Sample of TRICKS cuts for 2500 s of acquisition.Multiplanar reconstruction of TRICKS cuts for the aorta.Model of the descending aorta in axial plane.Reconstruction of the 3D model of the heart with the aorta. (a) Detection of the wall of the Aorta. (b) Detection of the inlet outlet of the Aorta. (c) Applying the Mask. (d) Reconstruction of the whole format of the AortaA 3D aorta model with reconstruction waste.Final reconstruction of the descending aorta model in 3D.Thickness analysis of the aorta.Analysis of the curvature of the aorta.(a) Geometry of the aorta (a) in inlet (b), outlet (c) and wall (d).Mesh of 3D aorta.Calculation of velocity residues during the first 10 iterations.Calculation of static pressure for input of aortic model in 3D (inlet), outlet (outlet), model of aorta outside (wall), model of aorta inside (interior-face5).Magnitude field of the vortex in the aorta.Mass flow rate.Reynolds cell number presentation for 3D aortic model (internal wall).Model 5D of the aorta.Flow line visualization in the 5D aorta model (without the 3D inner wall). (a)velocity streamline (frame1). (b) velocity streamline (frame2). (c) velocity streamline (frame3). (d) velocity streamline (frame4).Pressure of the 5D model of the aorta (without the inner wall in 3D).Contour accuracy of the aortic valve.Simulation of flow within the aortic valve.Five-dimensional description in different fields of imaging.Algorithm of thresholding.Algorithm of the mask application.Statistical analysis of the thickness and local curvature of the aorta.Estimation of the degree of aortic stenosis.Extraction of the measurements for the quantification of the fifth dimension of flow.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00003.txt

@@ -0,0 +1,29 @@
+These authors contributed equally to this work.Single-cell RNA-sequencing (scRNA-seq) technology provides an excellent platform for measuring the expression profiles of genes in heterogeneous cell populations. Multiple tools for the analysis of scRNA-seq data have been developed over the years. The tools require complicated commands and steps to analyze the underlying data, which are not easy to follow by genome researchers and experimental biologists. Therefore, we describe a step-by-step workflow for processing and analyzing the scRNA-seq unique molecular identifier (UMI) data from Human Lung Adenocarcinoma cell lines. We demonstrate the basic analyses including quality check, mapping and quantification of transcript abundance through suitable real data example to obtain UMI count data. Further, we performed basic statistical analyses, such as zero-inflation, differential expression and clustering analyses on the obtained count data. We studied the effects of excess zero-inflation present in scRNA-seq data on the downstream analyses. Our findings indicate that the zero-inflation associated with UMI data had no or minimal role in clustering, while it had significant effect on identifying differentially expressed genes. We also provide an insight into the comparative analysis for differential expression analysis tools based on zero-inflated negative binomial and negative binomial models on scRNA-seq data. The sensitivity analysis enhanced our findings in that the negative binomial model-based tool did not provide an accurate and efficient way to analyze the scRNA-seq data. This study provides a set of guidelines for the users to handle and analyze real scRNA-seq data more easily.The single-cell RNA sequencing (scRNA-seq) technique allows researchers to perform genome-wide gene profiling at the individual cell level [1]. This technology has led to a new beginning in transcriptomics by observing the expression dynamics of genes at the single-cell level, elucidating the complex biological systems, such as cancer, embryogenesis, etc. [2]. One of the recent studies highlights the use of single-cell technology in designing immunotherapy strategy for patients with early-stage lung cancer [3]. The single cell technology was used to create a complete immune cell atlas and track changes in the immune response to lung cancer.The study of scRNA-seq started with the characterization of cells from early developmental stages way back in 2009 [4]. The scRNA-seq requires the isolation and lysis of single cells, converting their RNA into cDNA, and the amplification of cDNA to generate high-throughput sequencing libraries. The outlines of the procedures involved in single-cell sequencing are shown in Figure 1. There are many protocols of scRNA-seq that exist in the literature, such as Fluidigm (C1 platform) [5], SMART-seq2 [6], CEL-seq [7], CEL-seq2 [8], Drop-seq [9], In-Drop [10], MARS-seq [11], SMART-seq [12], etc. The protocols vary in terms of coverage, sensitivity of mRNA capture, technical variability and costs involved [13].Biological processes are often dynamic and bulk RNA-sequencing (RNA-seq) techniques cannot capture the cellular heterogeneity and stochastic transcriptional processes [4]. Thus, the advent of scRNA-seq has brought radical changes and a new perspective to explore the biological processes at individual cells sampled from the cell populations (i.e., tissue samples) [6,7]. The main difference between scRNA-seq and the bulk RNA-seq lies mainly in the goal of the experiment in terms of what question is being addressed and a more complex analysis workflow [14]. Bulk RNA-seq is typically used to compare conditions and scRNA-seq is used to compare differences between cell types or identification of cell types. In scRNA-seq, each sequencing library represents a single cell instead of a population of cells, compared to bulk RNA-seq [14]. In addition to the usual analysis, a scRNA-seq data analysis involves handling of CBs (i.e., unique bar codes attached to each cell) and unique molecular identifiers (UMIs; i.e., unique tags attached to each transcript) [15].The main objectives of scRNA-seq include the identification of all kinds of cell types present in a tissue, estimation of the changes that occur during cell differentiation representing different stages or across time points and identification of differentially expressed (DE) genes across cell types [15]. In addition, the scRNA-seq has unique features, such as low library sizes of cells, stochasticity of gene expression, high-level noises, lower capturing of mRNA molecules, high dropouts, amplification bias, multi-modality of data, zero-inflation, etc. [16]. These make the analysis of scRNA-seq data more complicated compared to bulk RNA-seq.With advances in scRNA-seq, there are two key challenges, (a) noisy and excess over-dispersed data and (b) missing values [17]. There are a lot of technical and biological noises that leads to excess overdispersion in data. Because of the low amount of RNA and limited efficiency in mRNA-capturing from cells, there are many zeros in the data. These are called dropout events [18]. The efficiency of mRNA capture by oligo-dT primer depends on the length of the poly-A tail and so, the mRNAs with short poly-A tails are captured inefficiently [19]. Due to the low capture efficiency and dropout events, the output data are highly inflated with zeros. Moreover, a ‘zero’ count can be a low expression of a gene, i.e., structural zero or dropout/false zero, i.e., RNA in the cell was not detected due to limitations of current experimental protocols [20,21]. The dropout events increase the cell-to-cell variability and can reduce the detection of gene–gene relationships [22]. Therefore, dropout events can affect the downstream analyses.There are many tools available in the literature to perform individual analyses of raw FASTQ scRNA-seq data, quality control, preprocessing, mapping, zero-inflation and other downstream analyses [2,23,24,25]. These tools require complicated commands. In other words, these existing applications may not be too handy and easy to use for the users from non-bioinformatics backgrounds. Further, there is no optimal pipeline available for a variety of applications and analysis of scRNA-seq data. Scientists and genome researchers need to plan experiments and adopt different analysis strategies depending on the organism being studied and their research goals. This requires an easy to implement set of guidelines for the analysis and their application to real raw scRNA-seq experimental data.Therefore, we demonstrate here the steps involved in scRNA-seq data analysis including data collection, pre-processing and quality check, mapping to reference genome and other downstream analyses along with their application to a real raw experimental data. The first component of scRNA-seq analysis is the generation of a gene expression data matrix, after a thorough quality check. The second component is the major downstream analyses of the obtained single-cell expression data. The downstream analyses include cell clustering, zero-inflation and DE analysis. As the scRNA-seq data was zero-inflated, we studied the effects of various proportions of zeros on various downstream analyses of the data. Here, we also present a comparative performance assessment of two popular tools for DE analysis, i.e., DESeq2 [26] and DEsingle [27], of scRNA-seq data. This step-wise guide will help the experimental biologists and genome researchers in handling and performing various analyses of raw scRNA-seq experimental data.In this study, we used a real experimental dataset from the experiment “Single-cell profiling of 3 Human Lung Adenocarcinoma cell lines” to demonstrate the workflow of scRNA-seq UMI data [28,29]. This dataset comes from an equal mixture of cells from the three human Lung Adenocarcinoma cell lines, such as H2228, NCI-H1975 and HCC827. Here, 120,000 live cells were sorted using FACS (Fluorescence-Activated Cell Sorter) to derive an equal mixture from these three cell lines [29]. The chromium 3′ single-cell platform (10X Genomics) was used for processing and the Illumina NextSeq 500 sequencer was utilized for sequencing [29]. The considered dataset is available at National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database. The GEO accession of the study is GSE111108 and the Sequence Read Archive (SRA) ID is SRP133476. The run ID is SRR6782109 with the BioProject number PRJNA435946.We downloaded the raw scRNA-seq files (in .fastq format [30]) from the SRA database. The FASTQ files of a typical UMI-based dataset contain the reads, CBs and the UMI files. This dataset has a read1 (R1) file that contains a CB and the UMIs and a read2 (R2) file containing the sequence reads. We used the SRA toolkit (https://trace.ncbi.nlm.nih.gov/Traces/sra/sra.cgi?view=software, accessed on 10 December 2021) [31] to download the raw data using the FTP (file transfer protocol) function on a local computer from the SRA repository [32]. The Linux code used to download the data through SRA toolkit is given in the  Supplementary Document S1.FASTQ files are compressed files with the extension *.fastq.gz and can be unzipped using the command gunzip file.fastq.gz. They are intermediate output files generated by the sequencer and used for further analyses, including mapping. A FASTQ file consists of four lines: (i) sequence identifier, (ii) sequence, (iii) separator (contains only “+”) and (iv) Phred quality score (ASCII characters for each base call in the sequence). An example of .fastq file is shown in Supplementary Figure S1.The raw data files are in the FASTQ format, a text format consisting of quality scores calculated for the sequence reads, and need to be processed to proceed with the downstream analysis. The outline of the steps involved in the preprocessing of the data is shown in Figure 1B. There are few tools required to process the raw data to derive a count matrix, which is necessary for further analysis. The details of software/tools used in this study are summarized in Table 1.The critical challenge in a scRNA-seq study is to ensure that only single and live cells are included in the downstream analysis. The inclusion of dead cells or doublets would bias the results of the data analysis. Low-quality libraries in scRNA-seq data can stem from damaged or stressed cells or errors in library preparation. FastQC [33] is one of the most commonly used quality control tools, initially developed for bulk RNA-seq and later extended to scRNA-seq. The output from FastQC is an HTML file viewed in a browser after analyzing a read file in FASTQ format. It gives quality statistics from sequencing data and contains information about the input FASTQ file, type of quality score encoding, total number of reads, read length and GC content (Supplementary Figures S2–S4). For instance, the first plot, Per base sequence quality, gives quality score distribution over all the sequences. The red line in the center of the box and whisker plot gives the median value, the yellow box represents the inter-quartile range (25–75%), the upper and lower whiskers of the plot represent the 10% and 90% percentile scores and the blue line represents the mean quality of the read (Supplementary Figure S3).The sequence qualities are stored in the form of a Phred score [42]. The Phred score is an error probability belonging to each base and is calculated by
+        (1)Q=−10log10P
+		where Q is the quality score and P is the error rate. For instance, if Q = 30 is assigned to a base, this means the chances that a base is called incorrectly is 1 in 1000 (p = 10−3; the error rate is 0.1%). A high-quality score implies that a base call is more reliable and less likely to be incorrect; for example, p = 10−10, Q = 100 (1 in 10 billion), which is unrealistic and unlikely.In most of the scRNA-Seq library preparation protocols, the first 10–15 nucleotides are not uniformly distributed. So, the Per-base sequence content plot shows non-uniform base composition. The following plot is the Sequence GC content that gives the distribution of mean GC content. The Per base N content gives the percentage of bases at each position. The Sequence length distribution plot shows the distribution of fragment sizes and has one peak depicting the length of the read but flags a warning if multiple fragment lengths are there in the file. The Sequence duplication plot gives the percentage of reads of any given sequence in the file, which occurs several times in the file. The Overrepresented sequence lists the sequence which appears more than expected in the file. For RNA-Seq data, it is usual that few transcripts may be counted as overrepresented sequences due to the high abundance. A sequence is considered overrepresented if it is accounted for ≥0.1% of the total reads, for which a warning is raised, and a failure if it is >1%. The Adapter content plots the fraction of reads where the sequence library adapter sequence is identified. These quality indices are demonstrated in Figures S4–S10.CBs are short nucleotide sequences, such as the UMIs that are used for the identification of independent cells. The nucleotide sequence of CBs is known and serves as a unique identifier for a single cell in the gel droplet. Here, the read1 contains 26 nucleotides, out of which 16 bases correspond to the cell barcode (CB) and the other 10 bases correspond to the UMI. Each CB marks a partition from where DNA originates, although it may not contain a cell. The CB would differentiate between cells, but the UMI distinguishes between the RNA fragments. A total of 737,000 different sequences make a comprehensive whitelist to which any CB belongs [43]. It ensures there are no errors if the observed barcode does not match any barcode on the whitelist. The UMI-tools is a repository of a set of tools or functions that handle and process UMI-based data. UMI-tools’ whitelist command produces a list of CBs that are used in the sequencing of the dataset. To achieve this, the UMI-tool command given in Supplementary Document S1 can be used.The read1 or R1 file that contains the barcodes is specified as the input file. The --bc-pattern is the part of the command that finds and marks the CB and UMI in the read sequence (Supplementary Document S1). The default location of the barcodes is at 5′ end. The Cs denotes the 16 bases of the CB followed by 10 Ns representing the 10 bases of the UMI. There are many variations and several options that can be used to obtain the desired kind of output file, i.e., a .txt format.The next step is to extract the CB and UMI from read1 and add it to the read2 name. Afterward, the reads that do not match one of the accepted CBs are filtered out using the filter-cell-barcode function. The command for this is given in the Supplementary Document S1.Mapping the sequencing reads to the reference genome to obtain read counts is one of the critical, necessary steps in scRNA-seq data analysis. For this purpose, the common mapping tool used is STAR [44]. Though STAR is memory intensive, it is shown to have better accuracy and efficiency. Any other aligner that can identify splice junctions can be used as well. Multimapping reads are not allowed in this process. To begin with STAR, one needs to create a genome index. The genome index is created by STAR using a function by providing a reference genome in FASTA format and an annotation file in GTF format. STAR aligns and maps the reads data to the genome index creating the mapped BAM files. We expect >80–85% reads to align to the genome, assuming that there is no significant contaminant in the sample. The command used to generate the BAM files is given in the Supplementary Document S1.The reference genome with the GTF file GRCh38 was used in this study. It has annotations for 229,580 transcripts and 60,656 genes (https://www.gencodegenes.org/human/stats.html, accessed on 10 December 2021) [45]. Two reads from the same gene may be mapped to different locations and be counted as duplicates even though they belong to the same gene. First, reads are aligned; then, they are assigned to genes using the featureCounts tool from the subread package. This was achieved through running the command given in the Supplementary Document S1. The output file created by the featureCounts function is not sorted initially and needs to be sorted. Alignments are then indexed using the samtools through the code Counting reads given in the Supplementary Document S1. The count function processes the UMIs aligned with every gene in each cell to find the number of distinct and unique UMIs mapping to each gene and generate a count data file. The command used to generate a count data file (i.e., gene expression data matrix) is given in the Supplementary Document S1.The key steps involved in our analysis of the scRNA-seq dataset are outlined in Figure 1C. The first step in the analysis process requires the count data generated in the previous steps, followed by clustering and differential gene expression analysis.The negative binomial (NB) model is mostly used for fitting over-dispersed count data, such as RNA-seq data, that is, when the conditional variance exceeds the conditional mean. It has been implemented in most of the widely used tools for downstream analysis, such as DESeq2 [26], edgeR [46] and baySeq [47]. Let Yij be a random variable (rv) representing the read counts for the ith gene (I = 1, 2, …, N) in the jth cell (j = 1, 2, …, M), μij be the mean parameter for the ith gene in the jth cell and θij be the inverse of the dispersion parameter of the ith gene in the jth cell. The NB model used for scRNA-seq count data fitting can be expressed as
+          (2)fNB(y)=P[Yij=y]=G(y+θij)G(y+1)G(θij)(θijθij+μij)θij(μijθij+μij)y ∀ y=0, 1, 2, …
+          where μij≥0; θij>0 are the parameters of NB distribution and G(.) is a Gamma function.The expected value of the rv Yij can be given in Equation (3).
+          (3)E(Yij)=μij and V(Yij)=μij+μij2θij
+          If θij→∞ (No dispersion); NB(μij,θij) →Poisson(μij)ScRNA-seq count data are characterized by the presence of excess zeros due to low input mRNA materials, low capture rates for cells, etc. Therefore, the NB model in Equation (2) cannot give satisfactory results, as it does not account for excess zeros present in the data. The zero-inflated negative binomial (ZINB) model attempts to account for the extra zeros present in scRNA-seq data [27]. The ZINB model estimates two equations simultaneously, one for the count model (i.e., NB) and one for the excess zeros (i.e., Dirac delta function). For any π ∈ [0, 1], the UMI counts in the scRNA-seq study are assumed to follow a ZINB distribution. The PMF of the ZINB distribution can be expressed as
+          (4)fZINB(y)=P[Yij=y]=πijδ0(y)+(1−πij)fNB(y)∀ y=0, 1, 2, …
+          where fNB(.) is the PMF of the NB distribution given in Equation (2) and δ0(.) is a Dirac’s delta function. Here, δ0(.) is used to model the excess zeros in scRNA-seq data and its PMF is equal to zero for every non-zero counts except zero-counts and can be expressed as
+          (5)δ0(Yij=y):={1; y=00; y≠0Now, the PMF of the ZINB distribution to model the UMI counts from scRNA-seq data can be given as
+          (6)P[Yij=y]={πij+(1−πij)(θijθij+μij)θij when y=0(1−πij)G(y+θij)G(y+1)G(θij)(θijθij+μij)θij(μijθij+μij)y; y>0Now, Yij~ZINB(πij,μij,θij); then, the expected value and variance of Yij can be obtained as
+          (7)E(Yij)=(1−πij)μij and V(Yij)=(1−πij)μij(1+πijμij+μijθij)
+          If πij=0; ZINB(πij,μij,θij)→NB(μij,θij)
+          If θij→∞ (No dispersion); ZINB(πij,μij,θij) →ZIP(πij,μij)Zero-inflation and excess overdispersion are inherent problems in scRNA-seq data due to several reasons, such as technical noise, smaller input materials, low capture rates of protocols, etc. They affect the analysis, if not appropriately addressed during the data analysis. Further, zero-inflation stands for the proportion of zeros in the data, which is much higher than the proportion of the non-zero values. Mathematically, let Yij be any random variable having distribution function F(.), Yij ~ F(y); then, the expected value of zeros can be written as
+        (8)E(Yij=0)=SjP(Yij=0)
+        where Sj=∑iYij (library size for jth cell) and P(Y=0) is the probability of the scRNA-seq read count equal to zero. If the observed number of zeros in the data is higher than the theoretically expected value, we call the data zero-inflated. Moreover, when the observed variance is higher than the variance of the underlying theoretical model, overdispersion has occurred in the data. In other words, the observed variance is a function of the expected value. It is well established that the count data from bulk RNA-seq and scRNA-seq study are highly over-dispersed [18,20], as the variances of genes are the functions of their expected values (Equations (3) and (7)). So, we only focus on the testing of zero-inflation for the scRNA-seq data.To test the statistical significance of the zero-inflation parameter (πij) in Equation (4) of the ith gene in the jth cell (i.e., the proportion of zeros in the scRNA-seq data), we adopt the following generalized likelihood ratio test (GLRT) procedure. Here, for the testing purpose, we define the following null hypothesis:(9)H0: πij=0 vs. H1:πij≠0
+        where, H0 and H1 are the null and alternate hypotheses respectively. In other words, the null hypothesis tells us that the ith gene is not zero-inflated; subsequently, the scRNA-seq data structure is the same as RNA-seq data. Further, if we fail to reject H0, then the RNA-seq DE tools can be used for the DE analysis of scRNA-seq data with the expectation of satisfactory results. For simplicity, we assume that μi1=μi2=…=μiM=μi, θi1=θi2=…=θiM=θi and πi1=πi2=…=πiM=πi.The test, as mentioned above, H0 vs. H1, can be tested through GLRT and the test statistic can be given as:(10)−2lnα=−2{l(Ωi=Ω^i0; Yij)−l(Ωi=Ω^i; Yij)}
+        where Ω^i0 is the maximum likelihood estimator (MLE) of the parametric space, Ωi, for the ith gene under the constraint of H0; Ω^i: is the unconstrained MLE of Ωi for the ith gene and Ωi is the parametric space for the ith gene, i.e., Ωi={μi, θi, πi}. The test statistic in Equation (10) is asymptotically distributed as a chi-squared distribution with 1 degree of freedom under H0. Since droplet-based single-cell sequencing methods can capture approximately 1–10% of mRNA from the cells, ‘zero’ counts (for low expressed genes) or dropout events (due to stochasticity of expression) are observed in single-cell data [9,48,49].Clustering techniques start with all cells present in data, which are then grouped into sets or groups known as clusters. Clustering is performed in such a way that the cells present within the same cluster are homogenous with respect to cells in other clusters. The main rationale behind clustering is that cells in scRNA-seq data may be highly heterogeneous and we need to determine if the cells belong to same cell type or not. It also helps in identifying new genes and the marker genes for cell types [6,7,8,9,10]. Further, clustering is one of the essential tasks in exploratory data mining and is very often used in statistical data analysis. All clustering methods have the same approach of determining the similarity index and then grouping together similar objects into groups or clusters. In scRNA-seq data analysis, K-means clustering is extensively used and is described in the following section.K-means clustering is a type of unsupervised clustering method of vector quantization that partitions data points into k pre-defined clusters [41]. Each observation of data belongs to the cluster with the nearest mean. Each centroid of the cluster contains feature values which define the resulting groups. K-means clustering minimizes within-cluster variances. Each observed data point is assigned to its nearest centroid, based on the squared Euclidean distance.The main challenge in the cell cluster analysis of scRNA-seq data is determining the number of optimum cell clusters in which the cells need to be grouped [50]. This analysis is essential to determine the optimum number of cell types. Hence, we used the statistical approach developed by Das and Rai (2021) [51,52] for determining the optimum number of cell clusters for scRNA-seq count data. This is briefly presented below.Let Yijk be the expression value of the ith gene in the jth cell of the kth cluster (k = 1, 2,…, K); Yjk. be the mean expression value of the jth cell in kth cluster; Y..k be the mean expression value of the kth cluster; and Y… be the overall mean across M cells. Then, the total sum of squares (TSS) can be expressed as
+        TSS=∑k=1K∑j=1Mk(Y.jk−Y…)2
+        =∑k=1K∑j=1Mk(Y.jk−Y..k+Y..k−Y…)2
+        =∑k=1K∑j=1Mk(Y.jk−Y..k)2+∑k=1KMk(Y..k−Y…)2
+        (11)=WSS+BSS
+        where WSS is the within-cluster sum of squares and BSS is the between-cluster sum of squares. Now, the index can be given as
+        (12)rk=WSSWSS+BSSInitially, the values of k are taken as 2, 3, …, 50. For each value of k, the total cells present in scRNA-seq data are divided into that number of cell clusters and, subsequently, rk (Equation (12)) are computed for each k. The k value, which provides the maximum value of rk, can be chosen as the empirical number of optimum cell clusters for the observed scRNA-seq data. This is performed through plotting the values of k against rk values and, from the graph, the optimum value of k (optimum number of cell clusters) is determined. Here, we used k-means clustering as it is a non-parametric method and does not depend on the distributional nature of scRNA-seq data, as well as being flexible in selecting the k.The DE analysis is necessary for identifying key gene markers for novel cell type detection and studying the stochastic gene expression process [16]. There are a lot of tools publicly available for DE analysis of scRNA-seq data; an excellent review for this can be found in [16]. The DE analysis of scRNA-seq data plays a vital role in understanding the intrinsic and extrinsic biological processes in a cell [51,52,53]. The scRNA-seq data is highly heterogeneous and comprises a vast number of zero counts, which introduces challenges in detecting DE genes, one of the main applications of scRNA-seq. In this study, we considered two tools for performing the DE analysis of scRNA-seq data from adenocarcinoma cell lines.DESeq2 [26] is a method initially developed for DE analysis of bulk RNA-seq data which assumes the read counts follow an NB distribution. The input for the DESeq2 package is the raw count data matrix from the RNA-Seq or scRNA-seq. The read count Y ij for the ith gene in the jth cell is described with the NB generalized linear model by the following expression:(13)Yij~NB(μij,αi)
+        (14)μij=sjqij
+        (15)log2qij=xj.βi
+        where the mean = µij and the gene-specific dispersion factor = αi. The fitted mean comprises a sample-specific size factor sj and a parameter qij, the expected count of fragments for the jth cell. The coefficients βi give the log2 fold changes for the ith gene for each column of the model matrix. The DESeq2 first estimates the size factors that account for the differences in the library size, then estimates the dispersion for each gene and, lastly, fits a generalized linear model [26]. The DESeq2 uses the Wald statistic to calculate the p-value and size effect estimate for the log2 fold change.DEsingle [27] is an R package for DE analysis specifically for scRNA-seq data. It implements the ZINB model, given in Equation (4) and (5), to discriminate the observed zero values into two parts, i.e., constant zeros and zeros from the NB distribution. With the model, DEsingle is designed to overcome the issue of the excessive zero values observed in scRNA-seq data. To detect DE isoforms between two groups, DEsingle first calculates the maximum likelihood estimates (MLE) of two ZINB populations’ mean parameters (μ1 and μ2), then computes the constrained MLE of the two models’ parameters under the null hypothesis (H0: μ1=μ2) and, finally, uses the GLRT for testing H0. The normalization step is usually conducted before DE analysis to correct the amplification bias. We used the median normalization method, as implemented in DEsingle and DESeq2, to normalize the scRNA-seq count data (Supplementary Document S2).The performance of two methods for identifying genuine DE genes is evaluated using the area under receiver operating characteristic (AUROC) curve (i.e., true positive rate (TPR) vs. false positive rate (FPR)). These metrics are defined as
+        (16)TPR=Sensitivity=TPTP+FN
+        (17)FPR=1−Specificity=FPFP+TNWe computed the performance metrics including true positives (TPs), false positives (FPs), true negatives (TNs) and false negatives (FNs) through comparing the genes selected through each method (i.e., DESeq2 and DEsingle) with the reference genes. It is very difficult to obtain true reference genes for Adenocarcinoma cell lines; therefore, we selected the reference genes from the data itself using the fold change criterion [51]. Then, we computed these indices, i.e., TPs (Equation (16)) as the selected DE genes which are matched with the reference genes and FPs (Equation (17)) as the genes which were found to be significant but were not reference genes. Similarly, TNs (Equation (17)) were defined as genes that were not reference genes and were not found to be significant and FNs (Equation (16)) were defined as genes that were reference genes but were not found to be significant.The FastQC generates several reports on different quality parameters, such as summary statistics, distribution of per-base sequence quality, distribution of quality scores per sequence, distribution of sequence content, distribution of GC content, distribution of per-base N content, sequence length distribution, sequence duplication and distribution of over-represented sequences. Initially, we checked the quality of the raw data through FASTQC and the results are shown in Supplementary Figures S2–S11. For instance, Supplementary Figure S2 gives the basic statistics of our input FASTQ file and details regarding the file name, which is SRR6782109_2. fastq, and a type of base call file encoded by Sanger/Illumina. It was observed that the total number of reads in this file is 109,178,700, with read length for each sequence 98, and the percentage of GC content is 48% (Supplementary Figure S2).The per sequence quality plot, shown in Supplementary Figure S3, exhibits the blue line for the median quality score in the green-colored encoded portion for the plot. It is observed that the quality scores for most of the reads are above 30, which indicates better quality (Supplementary Figure S3). The sequence quality plot shows the distribution of average read quality in our dataset (Supplementary Figure S4). We found that the observed mean quality score was approximately 31 for our dataset, which indicates better quality of reads, as this value exceeds the threshold value (Supplementary Figure S3). In other words, we could not trace any universally low-quality reads in our dataset; therefore, the raw datasets could be used for further analyses. Further, similar interpretations of other quality parameters can be made from Supplementary Figure S5–S11. Overall, the data were compliant with the quality control standards values (Supplementary Figures S2–S11); hence, we proceeded to further process the raw data without trimming.The UMI-tools process the data downloaded in the FASTQ format after completing quality checks to generate the count matrix through mapping to the reference genome. Further, a whitelist.txt file, which comprises the accepted CBs that meet the default threshold, is generated in the first step. The file contains a table with four columns: the accepted CB, a list of other CBs within a default threshold distance, read count of the accepted CBs and counts of the other accepted in the list. This list is used in the second part of the step. First, the read1 file of the dataset containing the CBs and UMIs is extracted to a file read1_extracted in.fastq format. This step adds the CBs and UMIs removed from the read1 and adds them to the names of read2. This is an important step that makes the file ready to be used for mapping after filtering unique CBs (Supplementary Figure S12).We mapped the reads in fastq files to the human reference genome with the STAR aligner [44]. Out of the 81802319 total input reads, 81.77% of the reads were found to be mapped uniquely to the human reference genome with an average mapped length of 96.73bp. No reads mapped to multiple loci and 7.25% of reads remained unmapped. This indicates that a significant portion of the read sequences was mapped to the reference genome. The mapped files are in the SAM/BAM format. Moreover, the SAM format is human-readable version, while the BAM file format stores mapped reads in a standard memory efficient and compressed format. These files begin with a header section that includes details on the sample preparation, sequencing run and mapping details, quality, etc., followed by the tab-separated alignment section.After the mapping was complete, reads were assigned to the genes using the featureCounts function of the UMI-tools. It attaches a new tag and outputs a BAM file containing the identity of the gene that the read maps to. The counts function uses this file to output the error-corrected UMIs mapping to each gene. The output file contains a table with three columns: the gene_id, the CB and the count of UMIs. The count data generated for our dataset had 972 distinct and unique UMI counts contributing to the 972 cells detected and 42,406 transcripts. This count matrix was used for downstream data analysis. The read count data matrix gives a finite number of reads mapped to the reference genome. A glimpse of the output file for the Adenocarcinoma single-cell experiment is shown in Supplementary Figure S13.After the count data matrix was generated, we determined the percentage of zeroes in the dataset, since we were aware that there was a higher proportion of zeros present in the scRNA-seq datasets, i.e., most of the reads marked as zeros. In other words, counting the zeroes gives an idea of the presence of dropout events present in scRNA-seq data. The percentages of zeros present in each of 42,406 transcripts, as well as the fitting of the models to the data, are shown in Figure 2. Out of 42,406 transcripts, almost 35,000 had zeros over all the cells (Figure 2C). There were only fewer genes with fewer zeros across some cells (Figure 2C). Figure 2D shows the relation between observed zero proportions and estimated zero-inflation from the ZINB model. It was found that the observed zero proportions were more significant than the estimated zero-inflation parameter for each transcript. This is due to the fact that the observed zero proportions in scRNA-seq data were a mixture of the dropout zeros (i.e., zero-inflation parameter) estimated through the Dirac’s delta function and true zeros from the estimated NB model. Further, results from the statistical test of zero-inflation are shown in Figure 2F. It was found that the zero-inflation p-values for most of the genes were statistically significant (Figure 2F). This observation validated that scRNA-seq data was indeed zero-inflated due to the presence of dropout events or other experimental artifacts.With an increase in the heterogeneity of a biological sample, a larger sample size is needed to identify and define the cell population. Determining the sequencing depth at which the majority of human transcripts are expressed in a cell and which has sufficient coverage. This has always been a debatable topic. One study showed that estimated expression levels from one million reads per cell might be adequate [49], while another study stated that a shallow sequencing depth of only 20,000 reads per cell was also sufficient [50]. Therefore, it is pertinent to study the distribution library sizes of cells in scRNA-seq data. The distribution of cell library sizes over the cells and their ranks are shown in Figure 3. The graph indicates that, out of the 972 cells, about 60% of them had a library size greater than the mean library size of 6000 (Figure 3A). Further, there existed a sigmoid-type relation among the library sizes and ranks of the cells, as depicted by the s-shaped curve (Figure 3B).The most popular downstream analysis for scRNA-seq data is clustering, which is usually practiced to identify the cell types that exist among the cell population. However, this study remains subjective in deciding the optimum number of cell clusters that the cells present in scRNA-seq data can be divided. Here, we discussed an algorithm to determine the optimum number of cell clusters. We set the values of k as 2, 3, 4,… 50 and computed the clustering index for each k. The distribution of clustering indices over different cell cluster numbers is shown in Figure 4. Here, for lower k, we observed a higher clustering index value and this value gradually decreased with the increase in cell cluster numbers. We observed the point of inflection for this plot at k = 10 (Figure 4A). The inflection point is the point where the curve changes its direction and becomes parallel to the x-axis. In other words, the 972 cells present in the Adenocarcinoma scRNA-seq data were optimally clustered into 10 cell clusters (Figure 4A). Further, the optimal number of clusters depends on the total number of cells and the clustering index value.Single-cell experiments are often performed on mixtures of multiple cell types with increased heterogeneity [53]. All genes can be analyzed, but we may add noise by including all genes that are not expressed at an adequate level to provide a meaningful result [54]. This may hinder the analysis. We can filter genes based on the average gene expression level and select genes that are unusually variable across cells.We sought to test the effect of the missing values or the zeroes on the optimal number of cell clusters. For this analysis, the dataset was reduced at various levels depending on the percentage of zeroes in the dataset. We used the complete dataset with all genes included and no reduction of any sort for this case, to determine the optimum cell clusters. The results for this entire data case are shown in Figure 4A. Here, the different number of cell clusters was plotted against their corresponding computed cluster indices. It was observed that the curve flattened at the point x = 10, which means the point of inflection for this curve was 10. So, we can say that, for the no-reduction case, the cells present in the data were optimally divided into 10 cell clusters. These observed cell clusters could be mapped to different cell types. In other words, with all 42,406 genes included in the scRNA-seq data, the 972 cells were clustered into 10 cell clusters.In the second case, we reduced the number of genes based on the number of zeroes present, to find an optimal number of cell clusters. It was achieved by data reduction, whereby a certain percentage of genes whose expressions were ‘0’ in a specific percentage of cells were deleted. To be more precise, in this setting, we deleted the genes which had zero expressions in 80% cells and tried to determine the optimum number of cell clusters. This reduction process retained count expression data for 2415 genes over 972 cells. These data were used to determine the optimum number of cell clusters. The results are shown in Figure 4B. For this case, we found that the curve flattened at point 10 (i.e., point of inflection), which means the 972 cells were clustered into 10 cell clusters. In other words, the optimum number of cell clusters was 10 for 80% gene reduction. Here, we can say that gene reduction had no effect on the optimum number of cell clusters determination. This claim was further validated with other reduction scenarios and the results are shown in Table 2. Similarly, we reduced the number of genes based on 60%, 50% and 30% reduction to study the effect of gene reduction on clustering and optimal cell clusters. For a 60% reduction (i.e., number of genes reduced to 1201), the optimum cell cluster number was 10 (Table 2, Figure 4B). Similarly, for the 60% reduction case (number of genes = 1201), 50% reduction case (number of genes = 879) and 30% reduction (number of genes reduced to 454), the number of optimum cell clusters number was observed to be 10 (Table 2, Figure 4C,D, Supplementary Figure S14). From the above observations, it can be inferred that gene reduction did not affect the clustering of genes and the optimal number of clusters remained the same for all reductions. This implies that the zero counts in the data did not affect the optimal number of cell cluster determination.At the preliminary stage, we removed the cells whose library size was less than 1800 and further removed the genes which had non-zero expressions in ≤5 cells. Through this process, we selected the complete dataset having expression counts of 42,406 genes over 972 cells for further analyses. Prior to DE testing, we used the NB and ZINB models to study their suitability for fitting scRNA-seq data. The results are shown in Figure 2. The results indicate that, for fitting over-dispersed and zero-inflated datasets such as scRNA-seq, the ZINB model provided better results than the NB model (Figure 2A,B). This implies better suitability of the ZINB model for modeling the scRNA-seq count data, as well as better estimates of the parameters than the NB model. The reason may be attributed to the fact that the NB model accommodates excess zeros by underestimating the mean and overestimating the dispersion parameters [16, 51]. This phenomenon jeopardizes the statistical power of NB-based RNA-seq DE tools on discovering DE genes in the presence of zero-inflation when applied to scRNA-seq data [16].DE testing is a well-documented problem that originates from bulk gene expression analysis [55]. Here, we compared the two methods, i.e., DESeq2 and DEsingle, which are based on two different models to identify the DE genes. At a 1% level of significance, DEsingle identified 634 genes and DESeq2 detected 79 genes with only 25 genes common between them. At a 0.1% level of significance, 401 genes were detected by DEsingle, while 75 were detected by DESeq2, with only 22 common genes detected by both methods. The results of this analysis are summarized in Table 3. Further, the list of the top 500 DE genes for the Adenocarcinoma cell lines is given in Supplementary Table S1.From the above table, it can be concluded that NB model-based tools are not efficient in handling zero-inflated datasets such as the scRNA-seq. So, methods specific to the scRNA-seq need to be used. To substantiate our findings, we conducted a sensitivity analysis through ROC curves.We evaluated the performances of the two DE analysis methods on these Adenocarcinoma scRNA-seq data and the results are shown in Figure 5. In other words, the ROC and AUROC of the two methods, i.e., DESeq2 and DEsingle, are shown in Figure 5. The AUROCs for DEsingle and DESeq2 were found to be 76.2% and 66.6%, respectively (Figure 5). It is observed that DEsingle has a higher AUROC value than DESeq2 (Figure 5). This indicates that DEsingle performed better than DESeq2 on these Adenocarcinoma scRNA-seq data. This is because the ZINB model implemented in DEsingle provides better estimates of mean and dispersion than the NB model [16]. Thus, it offers better suitability of the ZINB for modeling the zero-inflated and over-dispersed scRNA-seq count data (Figure 2A,B). Further, the poor performance of DESeq2 can be attributed to the fact that the NB model accommodated excess zeros in scRNA-seq data by underestimating the mean and overestimating the dispersion, which further jeopardizes the statistical power to detect DE genes [51].Here, we provide a comprehensive step-by-step guide for the analysis of raw scRNA-seq data. Since the noise of scRNA-seq data is high, it is crucial to use appropriate methods to overcome noises in scRNA-seq data. Quality control helps in excluding low-quality cells to avoid involving artifacts in data interpretation. The count data generated after pre-processing was zero-inflated. We observed that the number of zeroes in a dataset did not affect our clustering or cell type detection. In other words, our statistical results indicate that the zero-inflation had no or minimal role in clustering. We also provide an insight into the comparative analysis for two DE analysis tools based on the ZINB and NB models. The results indicate that the existing DE tools designed for the RNA-seq data are not capable of distinguishing the two types of zeros. Further, the sensitivity analysis-based findings suggest that bulk RNA-seq DE methods did not provide an accurate and efficient way to analyze zero-inflated scRNA-seq data.Although many methods have been specially designed to analyze the scRNA-seq data, new techniques that can effectively handle the technical noise and expression variability of cells are still required. The new bioinformatics approaches would significantly enhance biological and clinical research and provide deep insights into the gene expression heterogeneity and cell dynamics. The approach of determining the optimum number of cell clusters is graphical, which is qualitative. Hence, a statistically sound approach needs to be developed, where the number of cell clusters is determined based on statistical significance values. To study the effect of zero-inflation on the performance on DE analysis approaches, more comprehensive computational studies need to be designed. Since there are no pre-existing clusters (such as cases and controls), selecting the optimal number of clusters may have an effect on significant gene signatures that we plan to study somewhere else.The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/biomedinformatics2010003/s1, Figure S1: Glimpse of the Fastq file. Figure S2: Basic summary statistics of raw sequence data (fastq file). Figure S3: Per base sequence quality plot (depicting the quality of the reads in the fastq file). Figure S4: Distribution of per sequence quality scores. Figure S5: Distribution of per base sequence content.  Figure S6: Distribution of per sequence GC content.  Figure S7: Distribution of per base N content. Figure S8: Sequence length distribution plot.  Figure S9: Sequence duplication levels plot. Figure S10: Chart of the overrepresented sequences. Figure S11: Distribution of percentage of Adapter content. Figure S12:  Part of the whitelist file. Figure S13: Glimpse of resulting count matrix. Figure S14: Number of optimal clusters 60% for reduction in genes. Document S1: Linux commands for various analyses. Document S2: Data normalization. Table S1: List of top 500 differentially expressed genes.S.D. conceived and designed the study; S.D. and A.M. developed the methodologies; S.D. and A.M. contributed the R-codes and Linux codes; S.D. and A.M. contributed materials; A.M. and S.D. drafted the manuscript; S.D. and S.N.R. corrected the manuscript; S.N.R. acquired the funding. All authors have read and agreed to the published version of the manuscript.This study was supported by Netaji Subhas-ICAR International Fellowship, OM No. 18(02)/2016-EQR/Edn. (S.D.) of Indian Council of Agricultural Research (ICAR), New Delhi, India. It was also partly supported by Wendell Cherry Chair (S.N.R.) in Clinical Trial Research, University of Louisville, USA., and multiple National Institutes of Health (NIH) grants (5P20GM113226, PI: McClain; 1P42ES023716, PI: Srivastava; 5P30GM127607-02, PI: Jones; 1P20GM125504-01, PI: Lamont; 2U54HL120163, PI: Bhatnagar/Robertson; 1P20GM135004, PI: Yan; 1R35ES0238373-01, PI: Cave; 1R01ES029846, PI: Bhatnagar; 1R01ES027778-01A1, PI: States) and Kentucky Council on Postsecondary Education grant (PON2 415 1900002934, PI: Chesney). The content is solely the responsibility of the authors and does not necessarily represent the views of NIH or ICAR.Not applicable.Not applicable.All the datasets used in this study are publicly available at the NCBI GEO database.Authors duly acknowledge the help and support obtained from Education Division, ICAR, New Delhi, India and ICAR-Indian Agricultural Statistics Research Institute, New Delhi, India. Authors are extremely thankful to anonymous reviewers for their thorough suggestions and critical comments for improving the manuscript.The authors declare no conflict of interest.Outlines of the workflow for various steps in scRNA-seq data analysis. (A) Key steps involved in a typical single-cell RNA-seq experiment starting from the sample preparation by the isolation and lysis of single cells up to the data analysis. (B) Data preprocessing steps beginning from the .fastq files up to the generation of count matrix and the tools required at each stage. (C) Significant data analysis steps with the input count data matrix undertaken in the scRNA-seq study.Data structure model, distributions and estimated parameters. (A) Different cumulative distribution function (CDF) fitted to the single adenocarcinoma cells RNA-seq data. The x-axis corresponds to the cumulative densities and the y-axis represents the read counts. The red color corresponds to the observed CDF, the pink color to NB and the blue color to ZINB. (B) Fitting of count data models to the given adenocarcinoma single cells RNA-seq data. In this plot, the x-axis represents the scRNA-seq read counts and the y-axis represents the densities. The red color corresponds to the observed density, the pink color to NB density and the blue color to ZINB density. (C) Distribution of zeroes. The x-axis is the number of genes and the y-axis shows the percentage of zeroes. (D) Plotting of observed zero proportions vs. estimated zero-inflation. The x-axis represents the estimated zero-inflation and the y-axis represents the observed zero proportion. Here, the plot shows that the observed zero proportions are greater than the estimated zero-inflation. (E) Relation between mean and dispersion. The log(mean) is shown on the y-axis vs. log(dispersion) on the x-axis. (F) The plot shows zero-inflation in data. The y-axis corresponds to the p-value and the x-axis represents the genes.Distribution of cell sizes for Adenocarcinoma scRNA-seq data. (A) Distribution of library sizes across the total number of cells. (B) Plot for cell ranks vs. cell sizes—distribution of cell library sizes over the cell ranks. Here, the y-axis represents the cells’ rank and the x-axis represents the sequencing depth.Effects of zero reductions on the determination of an optimum number of cell clusters. The figures are shown for (A) no reduction, (B) 80% reduction, (C) 50% reduction and (D) 30% reduction. The y-axis represents the values of clustering indices and the x-axis represents the values of optimum cell clusters. The blue line indicates the value of the optimum number of cell clusters in which the cells in the data can be clustered. For (A), we observe the optimal number of clusters is approximately 10 for the 972 cells with all genes included. (B) 80% reduction (80% reduction of zeros with 2415 reduced genes): the number of cell clusters was found to be 10.. (C) 50% reduction (50% reduction of zeros with 879 reduced genes): the number of cell clusters was found to be 10. (D) 30% reduction (30% reduction of zeros and with 454 genes): the number of cell clusters was found to be 10.Comparative analysis of DEsingle and DESeq2 in terms of AUROC. The figure shows the ROC curves of the two DE analysis methods, DESeq2 and DEsingle. The red color indicates the DEsingle and the blue color represents DESeq2. DEsingle has better performance in terms of AUC value as compared to DESeq2.List of the tools used in this study.Lists the optimal number of clusters and the number of genes in each reduction.Summary of results from a comparative analysis of DEsingle and DESeq2 using case I clustering method.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00004.txt

@@ -0,0 +1 @@
+These authors contributed equally to this work.The self-organizing maps portraying has been proven to be a powerful approach for analysis of transcriptomic, genomic, epigenetic, single-cell, and pathway-level data as well as for “multi-omic” integrative analyses. However, the SOM method has a major disadvantage: it requires the retraining of the entire dataset once a new sample is added, which can be resource- and time-demanding. It also shifts the gene landscape, thus complicating the interpretation and comparison of results. To overcome this issue, we have developed two approaches of transfer learning that allow for extending SOM space with new samples, meanwhile preserving its intrinsic structure. The extension SOM (exSOM) approach is based on adding secondary data to the existing SOM space by “meta-gene adaptation”, while supervised SOM portrayal (supSOM) adds support vector machine regression model on top of the original SOM algorithm to “predict” the portrait of a new sample. Both methods have been shown to accurately combine existing and new data. With simulated data, exSOM outperforms supSOM for accuracy, while supSOM significantly reduces the computing time and outperforms exSOM for this parameter. Analysis of real datasets demonstrated the validity of the projection methods with independent datasets mapped on existing SOM space. Moreover, both methods well handle the projection of samples with new characteristics that were not present in training datasets.The high-dimensional low sample size nature of modern -omics datasets necessitates application of dimensionality reduction and clustering approaches for their efficient analysis. The self-organizing maps (SOM) portrayal method implemented in the oposSOM package [1] has been proven to be a powerful approach for analysis of differential expression [2], molecular subtyping [3], and sample stratification [4]. SOM clusters gene expression profiles (vectors of gene expression values across samples) into miniclusters called meta-genes and projects high-dimensional data into two-dimensional maps. SOM clustering coupled with extensive downstream functional analyses allows for comprehensive annotation of transcriptome landscape, identification of co-expressed gene clusters, and linking them to biological processes using curated sets of genes with known functional background [5]. In contrast to other clustering and dimension reduction approaches, the SOM method allows for feature extraction and offers a mechanistic interpretation of underlying biological mechanisms in terms of molecular “portraits”, and spot modules of co-overexpressed genes [6]. Moreover, the SOM method was extended to the analysis of genomic [7], epigenetic [8], single-cell [9], and pathway-level [10] data as well as for “multi-omic” integrative analyses [11].However, the SOM method has a major disadvantage: it requires the retraining of the entire dataset once a new sample is added. In the case of few samples in the new dataset, this is a valid approach [12,13], but once the number of samples is large, or individual samples are being added consecutively, it could be time- and computing resource-consuming. Moreover, retraining causes the change of gene arrangements, thus making the results hard to compare.To address this issue, we have developed two new approaches that allow for extending SOM space with new samples, meanwhile preserving its intrinsic structure. The extension SOM (exSOM) approach is based on adding secondary data to the existing SOM space, while supervised SOM portrayal (supSOM) adds support vector machine regression model on top of the original SOM algorithm and allows “predicting” the portrait of a new sample. Both methods reuse information of the primary SOM for improved sampling efficiency of the secondary data and as such refers to transfer learning in SOM space. They have been shown to accurately combine existing and new data. exSOM is characterized by higher accuracy compared to supSOM, while the latter is useful when the sample size to secondary data is large, or samples are obtained sequentially.The general workflow of the algorithms is presented in Figure 1. In both cases, the “primary” dataset is trained with self-organizing maps (SOM), followed by clustering and downstream analysis [1] (Figure 1A). In exSOM, “secondary” data is added to the existing SOM space by passive training (Figure 1B). For supSOM, the support vector machine regression model (SVMR) is trained that maps input expression dataset to SOM “portraits” generated from “primary” data. Finally, a “secondary” dataset is supplied to the model for projection into the SOM space (Figure 1C). Below, the details of each algorithm are addressed in detail.The SOM algorithm realizes three main analysis tasks (Figure 1A): (1) dimension reduction of the single gene expression profiles into a reduced set of meta-gene profiles, (2) thereby clustering of similar gene profiles, and (3) multidimensional scaling represented by the mapping of each gene into the two-dimensional SOM grid. We used the parallelized SOM training algorithm implemented in Bioconductor R-package “oposSOM” [1]. The method projects high-dimensional gene expression data into a two-dimensional space: N (genes) × M (samples) gene expression matrix is translated into K (meta-genes) × M (samples) matrix of reduced dimensionality [14,15]. Genes are assigned to meta-genes based on the similarity of expression profiles across the samples. Each meta-gene profile can be interpreted as the mean profile averaged over all gene profiles of the respective meta-gene cluster. During the SOM training phase, the algorithm distributes the genes over the meta-genes using the Euclidean distance between the gene and meta-gene profiles as a similarity measure. Meta-genes are arranged in a k × k = K two-dimensional grid coordinate system and colored according to their expression level for each sample, providing the so-called “expression portraits” [1]. Group-specific mean portraits are generated by averaging the portraits of all cases belonging to a given group or subtype.In the SOM space, genes with similar profiles are located in adjacent meta-genes, which form “spot-like” areas of up- and downregulated expression meta-gene clusters on the map due to the self-organizing properties of the SOM. These spots represent clusters of co-regulated genes, termed expression modules, and their patterns are a characteristic fingerprint of each particular sample/group of samples. Lists of genes included in each of the spot modules provide a functional context of the spot and were evaluated with gene-set enrichment analysis approaches [14].The SOM extension method (exSOM) aims at adding new, secondary data (e.g., independent data on the same system obtained from follow-up studies or web repositories) to an already existing SOM space (e.g., that of the primary data portraying analysis). For this, the original SOM algorithm was adapted to realize standard meta-gene training for the samples already contained in the primary SOM training, and a passive, “piggyback” training of the meta-genes for the extension data (Figure 1B). In brief, the exSOM training algorithms comprise three steps analogous to the SOM training, which are iteratively repeated until a convergence criterion (e.g., predefined absolute number of iterations) is achieved:Training profile selection: A gene profile (i.e., a vector of expression values for all samples) is selected, usually by sequential order.Determination of best-matching meta-gene: The meta-gene profile, which is the most similar to the training profile, is determined using the Euclidean distance metric. Importantly, only data points corresponding to the original samples contribute to the similarity metric; data points of the extension samples are not considered in this step. This ensures that gene to meta-gene assignment is not altered by adding the extension samples when compared to the primary SOM training.Meta-gene adaptation: The expression values of the meta-genes are adapted according to the Hebbian learning rule according to the original SOM training algorithm [16]. It combines the difference between the training and the meta-gene profiles with a learning rate and a neighborhood factor, both incrementally decreasing as the training proceeds. In this step, samples from the original and the extension set are considered, resulting in iteratively optimized meta-gene expression values for all samples.This training algorithm eventually provides unchanged meta-gene values for the primary data and new, adapted meta-gene data for the secondary data, allowing for direct comparison and integrated downstream analyses.Supervised SOM (supSOM) portrayal is based on support vector machine regression (SVMR) and provides an alternative approach for extending an existing SOM space. In supSOM, one SVMR model is trained for each meta-gene individually, using the genes’ expression profiles of the primary data as independent variable, and the corresponding meta-gene profile obtained from the initial SOM training as dependent variable. Thereby, only genes associated with the particular meta-gene or one of the adjacent meta-genes are considered as predictors. Once a model is trained, gene profiles in new samples can be used to predict the corresponding meta-genes (Figure 1C). We applied SVM regression model with Gaussian kernel and evaluated supSOM performance for varying neighborhood radii.Performance and accuracy for exSOM and supSOM were assessed based on evaluation of correlation and root-mean-square deviation (RMSD) between metadata of the extension samples (i.e., the portraits) generated by SOM as reference vs. exSOM or vs. supSOM. For benchmarking runtime of the SOM initialization and training phases, we generated artificial expression matrices for the primary and secondary (extension) data (m1 = m2 = 50, 100, 200, 500, and 1000 arrays per class) using the “madsim” R package [17] (for parameters, see Text S1).We used the supSOM and exSOM approaches to evaluate the effect of infliximab treatment on transcriptome landscapes in ulcerative colitis and Crohn’s disease (GSE23597 and GSE16879) and to study disease grade-associated transcriptome changes in breast cancer (GSE42568, GSE10810, and GSE29431), respectively. Microarray raw intensity data were downloaded from the Gene Expression Omnibus repository [18]. Before proceeding with analyses, the data were converted to log2 expression, quantile normalized, and annotated using the “affy” package for R.The complete analysis results were deposited as supplementary data in the open-access repository Zenodo [19].We generated simulated microarray data for two classes with 10,000 or 30,000 genes and 50, 100, 200, and 500 samples per class, respectively, and used this data in the original SOM algorithm [14,15]. As was expected, the increase of the sample size, as well as the number of genes, caused a considerable extension of SOM initialization and training times (Table 1). In particular, time for training increases linearly with both the number of genes and the number of samples in the input data, as well as with total number of meta-genes in the map which was kept constant in this benchmark (K = 1600).We compared the accuracy of exSOM and supSOM using the “self-portraying” approach, which is equivalent to “resubstitution” error estimation in SVM classifiers [20]. For this, we generated another dataset consisting of 50 cases and 50 controls and 10,000 genes, trained the SOM, and then used exSOM and supSOM with the same dataset to evaluate the accuracy of the “portraying” of secondary data. The accuracy was calculated based on correlation and RMSD between meta-genes in primary and secondary data (exSOM) or SOM trained and SVMR predicted meta-genes (supSOM). The results showed that exSOM generates secondary “portraits” exactly identical to primary “portraits” with correlation equal to 1 and RMSD equal to 0 (Figure 2; for full portraits, see Figure S1).The supSOM performed slightly poorer compared to exSOM. The accuracy of supSOM portrayal depended on the neighborhood radius. The correlation values varied between 0.90 and 0.99, depending on the neighborhood parameter. A steep decrease of RMSD values was observed when increasing neighborhood radius from 1 to 4, while the selection of larger radii caused an increase in RMSD, presumably because of the inclusion of gene profiles from distant meta-genes (Figure 3). Based on the RMSD curve, we chose a radius value equal to 4 on 40 × 40 SOM grid for further analyses. Finally, we evaluated whether supSOM portrayal has an advantage over exSOM in terms of computing time. For this purpose, we generated simulated two-class microarray datasets (200 samples in each class with 30,000 genes). We used 50 random samples per class for SOM training and performed extSOM or SVMR portrayal on the rest of the samples and compared the times spent in each case (Table 2).The results obtained with simulated data indicate that both methods can be used for accurate “projection” of new datasets to the existing SOM space without perturbing the intrinsic structure of the latter. exSOM outperforms supSOM for accuracy, while supSOM significantly reduces the computing time and outperforms exSOM for this parameter. exSOM might be the method of choice when accuracy is important; however, one has to consider that self-portrayal used as a simulation model is based on SOM-training and is, thus, method-consistent for trained and verified extension data, while supSOM is not. Advantages of supSOM of faster computation may become more pronounced if the size of new samples is large or they become available not at once, but sequentially.In this section, we used two publicly available datasets from the context of inflammatory colon diseases as an exemplary use case: GSE23597 (title: “Expression data from colonic biopsy samples of infliximab treated UC patients”) and GSE16879 (title: “Mucosal expression profiling in patients with inflammatory bowel disease before and after first infliximab treatment”). The GSE23597 dataset contains samples from patients with baseline ulcerative colitis (UC) disease, as well as patients treated with infliximab or placebo (54,613 genes × 113 samples). This dataset was used as a reference (primary data) for SOM training. The GSE16879 dataset contains samples from patients with UC and Crohn’s disease (CD) before and after treatment with infliximab (54,613 genes × 90 samples). This dataset was used as secondary data for the extension approaches in a second step. The samples in both datasets were additionally stratified to responders and nonresponders. SOM portrayal of the primary dataset demonstrated that infliximab responders and nonresponders showed distinct patterns of deregulation of functional spots on the SOM transcriptome landscapes (Figure 4). The SOM portraits of disease baseline (untreated), as well as nonresponder patients, were characterized by upregulated spots on the upper right corner of SOM maps (spot H and F), while responder patients were characterized by an overexpressed spot in the left bottom corner of the map (spot P). The functional analysis of deregulated modules suggests the upregulation of inflammatory response, particularly tumor necrosis factor (TNF) signaling pathway in baseline nontreated patients and nonresponders, in agreement with previous studies [21,22]. In contrast, patients who responded to infliximab showed marked downregulation of inflammation and upregulation of functional gene sets associated with tissue restoration and cell metabolism. Interestingly, the gene expression landscape in the placebo group was similar to that of patients receiving infliximab; however, the magnitude of spot expression was considerably lower. However, compared to the drug, the placebo group was still characterized by upregulation of immune/inflammatory gene signatures (Figure 4 and Figure S2), suggesting that infliximab possesses strong anti-inflammatory effects in responders, and, in parallel, induces injured tissue restoration by activating growth factor signaling and metabolic pathways.Based on the SOM landscape obtained, we performed a supSOM and exSOM portrayal of gene expression in an independent dataset (GSE16879), which contained biopsy samples from patients with ulcerative colitis (UC) and Crohn’s disease (CD) before and after treatment with infliximab as well as normal colonic mucosa samples. In addition, patients were retrospectively stratified into infliximab responder and nonresponder groups [23]. supSOM (Figure 5) as well as exSOM (see Figure S3) portraits of UC patients after treatment showed perfect matching to the corresponding SOM portraits. Additionally, obtained results allowed for gaining additional insights into mechanisms of inflammatory bowel diseases and infliximab treatment efficacy. First, we observed considerable differences in the spot patterns observed in responder vs. nonresponder IBD patients before treatment (Figure 5). Both UC and CD nonresponder patients showed marked upregulation of immunity and inflammation-related signatures localized on the top right corner of the SOM portraits (corresponds to spots F, H, and I in primary SOM landscape, see Figure 4), particularly TNF signaling via TNFR2, pattern-recognition receptor signaling, nitric oxide synthesis, neutrophil activation, etc. (Supplementary Figure S4). Interestingly, baseline (before treatment) molecular portraits of UC and CD responder groups showed distinct patterns of up- and downregulated functional modules. The molecular portraits of CD responders were more similar to the healthy subjects (Pearson’s r = 0.74), compared to the UC responders (Pearson’s r = 0.09). Further analysis indicated that the UC and CD nonresponders were characterized by the increased baseline levels of TNF-a compared to the responders, however, with a similar tendency of expression decrease after treatment (Figure S5). This can indicate that not responding to the drug can be at least partially attributed to inadequate dosing of infliximab [24].As a second use case, we used the exSOM approach to perform disease grade-associated molecular portrayal of breast cancers. The GSE42568 dataset contains gene expression profiles measured in 121 healthy and breast cancer tissue samples. Samples were stratified by breast cancer histologic grading (17 normal, 11 Grade I, 40 Grade II, 53 Grade III) [25]. Using this dataset as primary, we performed 40 × 40 SOM training to cluster co-expressed genes and characterize transcriptome portraits of cancer grades. The results indicate that normal breast tissue expression signatures substantially differ from diseased ones (Figure 6A,B). Breast cancers were generally characterized by the loss of normal tissue gene expression (spot A), including response to hypoxia, lipid metabolism process, cell adhesion, and extracellular matrix organization. Moreover, we observed a grade-dependent increase in the number of differentially expressed genes (Figure S6). Furthermore, we also noticed switching cancer gene expression signatures from luminal to basal type (Figure 6C). Grade I cancers were characterized by upregulation of spot B associated with luminal type, response to estrogen, and immune response. Grade II cancers largely share gene expression signatures with Grade I and Grade III, representing a transition type without having a characteristic spot. The Grade III cancers were additionally characterized by upregulation of functional modules involved in cell proliferation, cell–cell adhesion, cell migration, and epithelial–mesenchymal transition (spots C) (Figure 6C).Next, we used exSOM portrayal to map samples from two different secondary datasets (GSE29431 and GSE10810) to the primary SOM landscape. The GSE29431 dataset contained 51 samples (12 normal, 3 Grade I, 11 Grade II, and 25 Grade III); the GSE10810 dataset contained 47 samples (27 normal, 2 Grade I, 10 Grade II, and 10 Grade III). The exSOM, as well as supSOM portraits for both secondary datasets, showed a good correlation with primary SOM counterparts (Figure 7 and Figure S7). Moreover, exSOM portraits further emphasized the “indiscrete” pattern of Grade II breast cancers. In line with previous data, our results suggest that grades are not discrete but rather form a continuum with uncertain boundaries, which complicate classification and assignment [26,27,28] of this important prognostic marker.In this paper, we described options for extending SOM-based high-dimensional transcriptomic data portraying with additional, independent samples. The two extension approaches presented enable overcoming the main limitation of SOM machine learning, namely, that adding samples or complete datasets changes the intrinsic primary structure of primary SOM. Both exSOM and supSOM demonstrated their utility in overcoming this drawback. Both methods have their advantages and disadvantages: while exSOM seems more accurate, supSOM is time-efficient. From the methodical side, the novelty of the study is provided by the combination of previous SOM portrayal neural network machine learning with extrapolation of metagene values for novel samples using additive transfer learning approaches which transfer novel data into a multidimensional space obtained from previously collected data. The novel methods considerably widen the application range of SOM portrayal because they not only make computations more effective but, especially, because they enable usage of always analyzed data space for novel samples.Analysis of inflammatory disease and cancer datasets demonstrated the validity of the projection methods with independent datasets mapped on existing SOM space. Moreover, we showed that the methods well handle the projection of samples with new characteristics that were not present in training datasets (see the “inflammatory bowel disease response to infliximab” section of the Results).Thus, we demonstrated that SOM extension methods (exSOM and supSOM) can remarkably extend the usage scenarios of SOM “molecular data portrayal” approaches.The following are available online at https://www.mdpi.com/article/10.3390/biomedinformatics2010004/s1 Text S1: gives the codes for the generation of a simulated data set. Figure S1: Complete portrayal of simulated dataset with SOM, exSOM, and supSOM, Figure S2: Pairwise differential gene expression in primary SOM IBD dataset (GSE23597), Figure S3: Comparison of supSOM and exSOM portraits in secondary IBD dataset (GSE16879), Figure S4: Biological processes associated with upregulated spots F, H, and I on the primary SOM IBD dataset. Baseline disease and nonresponders were characterized with upregulated stops F, H, and I (see Figure 4) related to inflammatory response, cytokine-mediated signaling, neutrophil activation, reorganization, etc., Figure S5: Differential expression landscape in IDB responders vs. nonresponders in the secondary dataset (GSE16879). Orange color indicates upregulation, blue color indicates downregulation, white indicates the region of invariant gene expression, Figure S6: Grade-dependent change of differential expression genes in breast cancer, Figure S7: Comparison of exSOM and supSOM portraits in secondary breast cancer datasets. (A) GSE10810, (B) GSE29431.Conceptualization, A.A., H.L.-W., M.N., and H.B.; methodology, A.A. and H.L.-W.; formal analysis, S.D. and M.N.; data curation, A.A.; writing—original draft preparation, A.A., H.B., H.L.-W., M.N., and S.D.; visualization, M.N.; supervision, A.A.; funding acquisition, A.A. and M.N. All authors have read and agreed to the published version of the manuscript.The work was supported by the Science Committee of RA in the frames of the projects 21AG-1F021 and 21SC-BRFFR-1F020 (to A.A.), and Armenian National Science and Education Fund in the frames of the project ANSEF compsci-2324 (to M.N.). The paper is supported by the State Target Program of the Government of the Republic of Armenia under grant agreement № 1-8/20TB project “Creating a Cloud Computing Environment for Solving Scientific and Applied Problems”.Not applicable.Not applicable.Raw data and scripts are available as supplementary datasets in the open-access repository Zenodo (https://zenodo.org/record/5736510, accessed on 1 December 2021) as well as supplementary materials of this article.The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.General workflow of exSOM and supSOM algorithms. (A) For both algorithms, the first step is SOM training with the primary dataset. (B) In exSOM, “secondary” data is added to the existing SOM space by passive training. (C) In supSOM, the SVMR model is trained to map the primary dataset to its SOM “portraits”. (D) During supSOM testing, the “secondary” dataset is supplied to the model for projection into the SOM space. Single arrows indicate the order in the pipeline, while double arrows the dimensions of samples/features in the matrix.Performance of exSOM transfer learning using “self-portraying”. (A) exSOM adapts secondary data to an existing SOM via passive training of the meta-genes. First, SOM arranges primary data (black lines in the top-right pane) on the grid, while secondary data (grey line in the top-right pane) do not contribute to gene clustering. During the extension phase, the secondary data is mapped to the existing SOM grid (black line on the bottom right pane). (B) The meta-gene adaptation of secondary data results in identical images compared to the corresponding sample in the primary data. (C) The correlation between paired samples from primary and secondary datasets showed perfect matching (Pearson’s correlation coefficients equal 1 (red diagonal in the heatmap) and RMSD equal to 0, not shown).Performance of supSOM transfer learning using “self-portraying”. (A) supSOM utilizes support vector machine regression to train the primary dataset with the corresponding meta-genes. Single arrows indicate the order in the pipeline, while double arrows the dimensions of sam-ples/features in the matrix. (B) The trained models are then used for the prediction of meta-gene values of the secondary data. (C) The performance of supSOM portrayal depends on the neighborhood radius. The optimal radius value was selected equal to 4 (red arrow) based on RMSD and Pearson’s correlation. (D) supSOM portrayal shows slight differences compared to original SOM. The white areas on the maps represent meta-genes where the prediction failed. (E) The correlation heatmap between paired samples from primary and secondary datasets showed good matching (Pearson’s correlation coefficients close to 1 (red diagonal in the heatmap).Transcriptome landscape (primary SOM space) of response to infliximab in ulcerative colitis. (A) The overview map is segmented into 19 modules, from which 5 modules (F, H, I, J, and P) were deregulated in a group-specific manner. Group-specific mean transcriptome portraits of studied groups (see [15]). (B) Module (spot)-specific expression profiles in groups. The results show that modules F, H, and I were upregulated in the baseline disease and nonresponder group, while the expression of modules P and J were upregulated in responders. (C) The heatmap of module-specific gene-set enrichment scores. The results indicate that drug responders and nonresponders show differential deregulation of gene modules that are associated with inflammation, TNF-alpha signaling (spots F, H, and I), tissue restoration, and cell metabolism (spot P).Projection of GSE16879 (mucosal expression profiling in patients with inflammatory bowel disease before and after first infliximab treatment) dataset onto primary SOM space. supSOM portraits of UC patients after treatment showed perfect matching to the corresponding portraits of the original SOM. supSOM portraits highlight the differences in deregulated spots between responder and nonresponder IBD patients before treatment. Both UC and CD nonresponders are characterized by an overexpressed spot on the top-right corner of their corresponding group portraits, which remains unchanged after treatment. In contrast, US and CD responders showed a different distribution of upregulated spots before treatment, while their corresponding portraits after treatment resemble transcriptome portraits of healthy mucosa.Primary SOM transcriptome landscape of disease grade-stratified breast cancers. (A) Overview map of deregulated functional gene modules in primary SOM. (B) Group-specific module expression. (C) Heatmap of enrichment analysis scores of functional modules.Comparison of exSOM portrayal of breast cancer transcriptome landscapes. (A) Portrayal of GSE42568 (primary) and GSE10810 (secondary/extension). (B) Portrayal of GSE42568 (primary) and GSE29431 (extension).SOM training times for gene expression matrices of different sizes.Comparison of computational time of exSOM and supSOM.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00005.txt

@@ -0,0 +1 @@
+Typhoid fever caused by the bacteria Salmonella typhi gained resistance through multidrug-resistant S. typhi strains. One of the reasons behind β-lactam antibiotic resistance is -lactamase. L, D-Transpeptidases is responsible for typhoid fever as it is involved in toxin release that results in typhoid fever in humans. A molecular modeling study of these targeted proteins was carried out by various methods, such as homology modeling, active site prediction, prediction of disease-causing regions, and by analyzing the potential inhibitory activities of curcumin analogs by targeting these proteins to overcome the antibiotic resistance. The five potent drug candidate compounds were identified to be natural ligands that can inhibit those enzymes compared to controls in our research. The binding affinity of both the Go-Y032 and NSC-43319 were found against β-lactamase was −7.8 Kcal/mol in AutoDock, whereas, in SwissDock, the binding energy was −8.15 and −8.04 Kcal/mol, respectively. On the other hand, the Cyclovalone and NSC-43319 had an equal energy of −7.60 Kcal/mol in AutoDock, whereas −7.90 and −8.01 Kcal/mol in SwissDock against L, D-Transpeptidases. After the identification of proteins, the determination of primary and secondary structures, as well as the gene producing area and homology modeling, was accomplished. The screened drug candidates were further evaluated in ADMET, and pharmacological properties along with positive drug-likeness properties were observed for these ligand molecules. However, further in vitro and in vivo experiments are required to validate these in silico data to develop novel therapeutics against antibiotic resistance.Typhoid is a usual illness in economically handicapped countries where public health settings are very poor. A globally estimated 12–27 million people get stricken with typhoid fever each year, whereas the overall yearly estimated incidence lies between 292 and 395 cases per 100,000 people in Bangladesh. This infection-causing agent is an anaerobic Gram-negative rod, namely Salmonella enterica serotype Typhi (S. typhi), a highly conserved serovar subspecies of S. enterica, which is transmitted by the fecal-oral route and can infect the intestinal tract and blood [1,2,3,4,5]. S. typhi can provoke many health issues, such as fever, abdominal discomfort, and several gastrointestinal complications, such as nausea, vomiting, constipation, diarrhea, etc. The first approved antibiotics for the prevention of typhoid fever were chloramphenicol, ampicillin and cotrimoxazole [6], which have already started showing resistance and evolved multidrug resistance (MDR) S. Typhi strains over the last two decades. Due to the ever-increasing pattern of MDR in many parts of the world, combating typhoid is becoming more difficult, creating a major public health concern around the world [7].In the early 1970s, the first MDR S. typhi strains displaying concurrent resistance to the first-line antibiotics, such as ampicillin, chloramphenicol and co-trimoxazole, were demonstrated, followed by the emergence of ciprofloxacin-resistant strains in the 1990s [8,9]. Currently, the latter is observed in more than 90% of clinical isolates from endemic areas [10,11,12]. A 15-year (1993–2013) genome-wide study on S. typhi conducted in Bangladesh using 536 medical isolates reported that these bacterial strains show resistance to ampicillin (amp), co-trimoxazole (sxt), chloramphenicol (chl), ciprofloxacin (cip), and ceftriaxone (cro) where 37.69% strains displayed co-occurring resistance towards amp, sxt, chl, and cip followed by only cip-R (R = resistant) strains to comprise 31.53% of the total. Some of the resistance genes detected in the isolates of that study were blaTEM-1B in 50.28% of amp-R, qnrS1 in 10.2% of cip-R, and tet (A, B) in 9.46% and 8.53% tet-R (tet = tetracycline) strains, respectively [13]. The presence of extended-spectrum β-lactamase (ESBL) resistance Salmonella prevalence in poultry sourced recently from super shops of five divisional megacities of Bangladesh implies its possible human transmission through contaminated foods of poultry origin and the potential health risk of the people [14,15]. Additionally, as recorded in various parts of the world, S. typhi is now increasingly developing resistance to ciprofloxacin and fluoroquinolone and has emerged as a new threat to the treatment of typhoid fever [16,17,18,19,20,21,22,23,24].S. typhi acquires a ciprofloxacin-resistance (cip-R) property through the point mutations in quinolone resistance-determining regions (QRDR) with several positions corresponding to the genes, topoisomerase IV (parC and parE) and DNA gyrase (gyrA and gyrB) of S. typhi [25,26,27,28], whereas the acquisition of the blaTEM gene is responsible for the resistance property of S. typhi against β-lactam antibiotics through encoding the β-lactamase enzyme that hydrolyzes the peptide bond of the four-membered β-lactam ring and thus prevents β-lactam antibiotics from exerting their effect [29]. Moreover, derivatives of TEM, along with those of SHV- and CTX-M-type β-lactamase genes, comprise the family called extended-spectrum β-lactamases (ESBLs), which leads to the development of multidrug-resistant S. typhi, limiting the current treatment practices and thus posing an alarming situation in public health [30].Typhoid toxin is a prominent feature of S. typhi that contributes potential virulence to the bacterial infection causing typhoid fever by exclusively targeting the immune system and central nervous system of the host. The presence of one type of sialic acid is necessary for its binding to the host, which is abundantly available in humans. Thus, S. typhi cannot cause typhoid in hosts other than humans [31]. The export of the toxin to the outer membrane of the bacteria begins with the secretion of the individual subunits to the periplasm via Sec machinery and their assembly into the holotoxin complex. This holotoxin is then translocated across the PG (peptidoglycan) layer from the cis side to the trans side by the action of a special type of muramidase, TtsA, which is located at the bacterial poles and requires the PG editing by the L, D-Transpeptidases, namely YcbB, for its activity. After being translocated to the trans side of the PG layer, it becomes compartmentalized in an S. typhi-containing vacuole, from where its eventual release takes place upon exposure to the antimicrobial peptide or bile salts, and in this way, the transmission continues from one infected cell to the other [32].The function of L, D-Transpeptidases, YcbB, which is exclusively present in the bacterium S. typhi, is well-understood in the edition of PG. Glycan strands are the building blocks of the bacterial PG that are composed of N-acetylglucosamine (GlcNac) and N-acetylmuramic acid (MurNac) [33,34,35]. These building blocks make the PG by being connected by small peptides. Here, the enzyme L, D-Transpeptidases plays its role in introducing cross-links within L- and D-amino acids that comprise the peptides (Figure 1). This PG remodeling by L, D-Transpeptidases is necessary for TtsA to position the typhoid toxin for its proper release [32]. Here, L, D-Transpeptidases can be a major target for in silico studies as it plays a vital role in the secretion of typhoid toxin, and there is no effective drug available to inhibit it without exhibiting any side effects. For example, drug carbapenem and copper can inactivate L, D-Transpeptidases, yet they are associated with diarrhea, nausea, vomiting, skin rash, low blood pressure, anemia, heart problems, etc. [36]. Moreover, several antibiotics are working alone or coupled with β-lactamase inhibitors (Avibactam, Clavulanic acid (clavulanate), Relebactam, Sulbactum, Tozobactum, etc.), which have many adverse effects such as gastrointestinal complications, impairment of nervous system, hematological effects, and dermatological abnormalities, including Stevens-Johnson syndrome, toxic epidermal necrolysis, and drug-induced eosinophilia, etc. [37,38,39,40,41,42,43].Curcumin, the main bioactive component of turmeric (Curcuma longa L.), has been shown to be a powerful antioxidant, anti-inflammatory, antibacterial, antifungal, and antiviral agent in many studies [44]. Curcumin has been shown to be antibacterial against Staphylococcus aureus (S. aureus). Curcumin has significantly more effective antibacterial properties when combined with other antibacterial drugs, as revealed by in vitro experiments [45]. Curcumin inhibits bacterial growth due to its structural properties and the production of anti-oxidative chemicals. Through the bacterial quorum sensing regulatory system, curcumin can decrease bacterial virulence factors, reduce bacterial biofilm formation, and restrict bacterial adherence to host receptors [46]. Curcumin’s potential antibacterial action makes it a viable option for enhancing the inhibitory impact of current antimicrobial drugs through synergism [47]. It decreased Salmonella enterica serovar Typhimurium’s motility by reducing the length of the flagellar filament (from 8 m to 5 m) and lowering its density (4 or 5 flagella/bacterium instead of 8 or 9 flagella/bacterium). Curcumin therapy reduced the proportion of flagellated bacteria from 84 percent to 59 percent [48].As curcumin has antibacterial properties, including antioxidant and anti-inflammatory properties, we selected 70 curcumin analogues as ligands in this study. The schematic representation of the methodology applied in the present study is displayed in Figure 2.About 70 curcumin molecules were selected as ligands from literature to target the proteins β-lactamase and L, D-Transpeptidases of S. typhi. Each of the molecules was assessed for Lipinski’s rule of five or not [49,50]. Molinspiration Cheminformatics server (https://www.molinspi-ration.com/cgi-bin/properties accessed on 8 October 2020) was applied to experiment various drug-like parameters of the ligand molecules [51,52] (Table S2). Molinspiration allows for the prediction of significant molecular parameters (logP, polar surface area, number of hydrogen bond donors and acceptors, and so on), as well as the prediction of bioactivity scores for the most relevant therapeutic targets (GPCR ligands, kinase inhibitors, ion channel modulators, nuclear receptors). Compounds that did not comply with the rule were excluded from further study.The complete protein sequences of β-lactamase and L, D-Transpeptidases were retrieved from NCBI (https://www.ncbi.nlm.nih.gov/ accessed on 12 October 2020) in the standard FASTA format.The physical and chemical parameters of the proteins, including molecular weight (MW), theoretical pI, amino acid composition, estimated half-life, instability index, aliphatic index, etc., were computed using the ProtParam tool of ExPasy server as it evaluates physicochemical data (molecular weight, theoretical pI, amino acid composition, atomic composition, extinction coefficient, estimated half-life, instability index, aliphatic index, and grand average of hydropathicity (GRAVY)) from a protein sequence. (http://web.expasy.org/program/ accessed on 12 October 2020) [53].The online tool SOPMA [54] (https://npsa-prabi.ibcp.fr/NPSA/npsa_sopma.html accessed on 15 October 2020) was applied for the secondary structure prediction of the proteins. Homology Modeling was performed in Swiss-Model, as its workspace is a Web-based modeling expert system that is integrated. We search a library of experimental protein architectures for acceptable templates for a specified target protein. A three-dimensional model of the target protein is constructed based on a sequence alignment between the target protein and the template structure [55,56,57,58,59] (https://swissmodel.expasy.org/, accessed on 18 October 2020) to construct the tertiary structures of the target proteins using the three-dimensional structure of a related protein as a template. Homology modeling was also done by Phyre2. It is a web-based collection of tools for predicting and analyzing protein structure, function, and mutations. Phyre2’s goal is to provide biologists with a simple and intuitive interface to cutting-edge protein bioinformatics tools. Phyre2, which builds 3D models, predicts ligand binding sites and analyzes the influence of amino acid changes (e.g., no synonymous SNPs (nsSNPs)) for a user’s protein sequence using advanced distant homology detection algorithms [60].Disordered regions present in protein molecules remain unstable in the native state. To find out the disordered regions in proteins for which they lack a fixed tertiary structure, the Protein Disorder prediction System (PrDOS) server [61] (http://prdos.hgc.jp/cgi-bin/top.cgi, accessed on 18 October 2020) was exploited. This server predicts the disordered regions based on both local amino acid sequence and the template or homologous proteins through the SVM algorithm and PSI-BLAST, respectively. The prediction method consists of two predictors: one based on local amino acid sequence information and the other on template proteins. For each residue, the server aggregates the findings of the two predictors and delivers a two-state prediction (order/disorder) and a disorder probability.Model validation was carried out in PROCHECK [62,63] (http://www.ebi.ac.uk/thornton-srv/software/PROCHECK/, accessed on 20 December 2021), which assesses the stereo-chemical quality of a protein structure, i.e., how normal or unusual the pattern of the protein residues is, compared with a fined-tuned, high-resolution structure of a protein. PROCHECK evaluates the stereochemical quality of a protein structure by generating a series of PostScript graphs that analyze its overall and residue-by-residue geometry. It contains PROCHECK-NMR, which is used to check the quality of structures solved by NMR. For the prediction of active sites in the proteins where ligands will likely bind, two servers CASTP (Computed Atlas of Surface Topography of proteins) [64] (http://sts.bioe.uic.edu/castp/ accessed on 20 October 2020) and COACH [65,66] (http://zhanglab.ccmb.med.umich.edu/COACH/ accessed on 22 October 2020) were utilized. CASTP implements the theoretical and algorithmic results of computational geometry to predict the ligand-binding sites. It has several advantages: (1) pockets and cavities are recognized analytically, (2) the boundary between the bulk solvent and the pocket is accurately specified, and (3) all derived parameters are rotationally invariant, do not need discretization, and do not make use of dot surface or grid points. On the contrary, the COACH server applies two comparative methods, TM-SITE and S-SITE, to identify active sites in the protein.The ligand molecules were minimized in the Avogadro software using the mmff94 force field. Then, the protein structure was minimized in YASARA software using the AMBER14 force field. The docking program was carried out in the AutoDock Vina program. The ligands that will give the best results will be docked again using the SwissDock server. We used two servers to check the validity and to build a strong hypothesis. AutoDock Vina, a novel molecular docking and virtual screening application, has been introduced. Vina’s local optimization process employs a powerful gradient optimization algorithm. The gradient computation essentially provides the optimization algorithm with a “feeling of direction” from a single evaluation. Vina may speed up processing by making use of many CPUs or CPU cores by employing multithreading. SwissDock is a webserver dedicated to doing protein-ligand docking simulations in an easy and beautiful manner. SwissDock is protein-ligand docking software with a simple and integrated interface that is based on EADock DSS.AutoDock Vina PDB file of protein was converted into a pdbqt file using the Auto Dock Vina tool. Then, every selected ligand file was converted to a PDB file using pymol because the AutoDock tool can only recognize the PDB format. Then, the PDB file of the ligand was converted into a pdbqt file, which is the criterion for the AutoDock run. After docking, we visualized the docked file using pymol.SwissDock: At first, we set up the protein PDB file manipulating PDB code, which was retrieved from Uniprot or RCSB PDB. We specified chain A in β-lactamase. Then, ligands were selected from the ZINC database, but those that were not present in this database were uploaded in Mol2 format. The server provides very fast, fast, and accurate results.ADME/T describes the Absorption, Distribution, Metabolism, Excretion and Toxicity of a drug-like substance. These properties account for the success of a drug in clinical trials. Therefore, in silico ADME/T profile examination of the candidate drugs is a prerequisite for the fruitful measure of drug designing expenditure [67,68]. The best 8 ligands (based on the docking score) were utilized to speculate their drug-like potential by observing pharmacokinetic and pharmacodynamics features. ADME/T profile of all the chosen ligands were calculated utilizing admetSAR 2.0 (http://lmmd.ecust.edu.cn/admetsar2/ accessed on 23 October 2020) and pkCSM (http://biosig.unimelb.edu.au/pkcsm/ accessed on 23 October 2020) server [69,70]. admetSAR presents an easy-to-use interface for searching for ADME/T (Absorption, Distribution, Metabolism, Excretion, and Toxicity) attributes profiling by name, CASRN, and similarity search. With QSAR models, admetSAR can predict around 50 ADMET endpoints. The pkCSM signatures were effectively employed to create predictive regression and classification models across five different pharmacokinetic property classes.ERRAT software was used to validate the protein model. ERRAT is a software that verifies crystallographically determined protein structures.The elected ligands were employed to determine their pharmacological and biological activities accurately by using Prediction of Activity Spectra of Substances (PASS) Online (http://www.pharmaexpert.ru/passonline/ accessed on 26 October 2020) and Molinspiration Cheminformatics server [71]. These methods are used in conjunction with recognized compounds present in the database, depending on the structure-activity relationship (SAR). PASS Online predicts about 4000 different types of biological activity, such as pharmacological effects, mechanisms of action, toxic and unfavorable effects, interactions with metabolic enzymes and transporters, gene expression influence, and so on. Molinspiration Cheminformatics is also useful software to predict pharmacological and biological activities. In silico methods can contribute significantly to the prediction of drug metabolism sites focusing on the experimental view of the drug designing process. By conducting bioassay, these sites impart the knowledge of the molecules’ metabolic susceptibility and their fate inside the body [72]. RS-WebPredictor (http://reccr.chem.rpi.edu/Software/RS-WebPredictor/ accessed on 1 November 2020), an online server, was used to predict the best sites of drug metabolism mediated by CYP2C9, CYP2D6, and CYP3A4, three promiscuous isoforms of Cytochrome P450 (CYP) family. Predictions may be made for the promiscuous CYP isozymes 2C9, 2D6, and 3A4, as well as CYPs 1A2, 2A6, 2B6, 2C8, 2C19, and 2E1. The RS-WebPredictor service is the first publicly available server that predicts the regioselectivity of the last six isozymes.Among 70 types of ligands, all the compounds except DM1 were compatible with the drug-likeness features. Most of the selected compounds did not display any violations to Lipinski’s rules, which indicated the pharmacokinetic conformity of these compounds. Hence, all these compounds were taken into account for the next phases of the study (Tables S1 and S2).Two protein sequences of β-lactamase (EC 3.5.2.6) (Uniprot ID P62593; PDB ID 1ZG4) and L, D-Transpeptidases (Uniprot ID P22525; PDB ID 6NTW) were retrieved from NCBI (https://www.ncbi.nlm.nih.gov/ accessed on 12 October 2020) in standard FASTA format.Using the ProtParam tool of the ExPasy server, the physical and chemical parameters of the proteins were analyzed. β-lactamases is 286 amino acids long, weighted at 31515.20 grams, and has an instability index of 40.74. L, D-Transpeptidases is 615 amino acids long, 67812.49 (gm) weighted, with an instability index value of 43.79. The Instability Index is a metric for determining whether a protein will remain stable in a test tube. Table 1 shows the physicochemical properties of these two proteins. The SOPMA tool was used to predict the secondary structures of these two proteins. The values of alpha helix were 49.30% and 39.35% for β-lactamase and L, D-Transpeptidases, respectfully. The value of the extended strand was also greater for β-lactamase (12.94%) compared with the L, D-Transpeptidases (11.54%). L, D-Transpeptidases (43.58%) exceeds the β-lactamase (29.37%) in the case of a random coil. The values of 310 helix, pi helix, beta bridge, bend region, and ambiguous status were 0.00% for both proteins (Table 2).The three-dimensional structures of β-lactamase and L, D-Transpeptidases were predicted in Swiss-Model web tools. The biounit oligo state of both proteins was a monomer. The template displayed 0.61 sequence similarities, coverage score of 1.0, and 24–286 range for β-lactamase, whereas L, D-transpeptidases showed 0.62 sequence similarity, coverage score of 0.95, and 37–615 range (Table 3). In Swiss-Model, two models for β-lactamase and three models for L, D-Transpeptidases were predicted based on the top 31 and 50 templates, respectively. On the contrary, the top 20 models were predicted by Phyre 2 for both proteins each. The best fit built by the two servers for β-lactamase had a confidence score of 100 when modeling 263 amino acid residues at positions 24–286, and for L, D-transpeptidases, it modelled 505 residues at positions 37–615 with a confidence score of 100.In ERRAT, the overall quality factor of β-lactamase and L, D-Transpeptidases was is 97.2549 and 84.1141, respectively. The two proteins also passed the verified 3D in their respective prediction results. The β-lactamase (EC 3.5.2.6) has an 11% disease-causing region, and its active sites are acyl ester intermediate (position 70) and proton acceptor (position 168). The Ramachandran plot analysis showed that both proteins delineated more than 90% of the amino acid residues in the most favored regions. The number of non-glycine and non-proline residues is 228 among 263 residues. Twenty-one glycine residues and 12 proline residues are present (Ramachandran plot), and 93.4% of residues are in the favored region. The L, D-Transpeptidases is 615 amino acids long, and its molecular weight is 67812.49. The active site of this protein stays in the 528 position. Here, 91.8% residues are in the favored region with 427 residues excluding glycine and proline, 8 terminal residues other than Gly and Pro, 31 glycine (represented as a triangle), and 3 proline residues (Table 4 and Figure 3 and Figure 4).Using CASTP (Computed Atlas of Surface Topography of proteins) and COACH servers, the active sites of β-lactamase and L, D-Transpeptidases were predicted (Figure 3F and Figure 4F).The best five ligands were selected based on docking experiments among 70 ligands. Go-Y032, NSC-43319, Cyclovalone, Salsalate, and Cyclocurcumin showed the best docking results against the β-lactamase (1ZG4) enzyme (Table 5 and Table 6 and Figure 5), and NSC-43319, Cyclovalone, Cyclocurcumin, Difluorinated curcumin, and Go-Y032 showed the best docking results against the L, D-Transpeptidases (6NTW) enzyme (Table 5, Table 6 and Figure 5). Two controls were selected for each of the enzymes. Clavulanic acid and Tazobactum were docked against β-lactamase as they are existing drugs. These existing drugs showed low affinity to other ligands. For L, D-Transpeptidases, Carbapenem and Cephalosporin were selected as controls. These controls were also showed lower affinity. ADME/T profiling was carried out for the ligands that gave the best docking scores in the molecular docking study (Table S6) and control group (Tables S3 and S4). Intestinal absorption and oral bioavailability were high for all of the ligands. Caco-2 permeability was high for the ligands Go-Y032 and NSC-43319, while the remaining ligand showed low Caco-2 permeability. Furthermore, NSC-43319 and Cyclovalone were predicted as substrates of P-glycoproteins, while the remaining was non-substrates of membrane P-glycoproteins. All of the ligands were able to enter the blood–brain barrier, excluding Cyclocurcumin. All of them exhibited substrate specificity of CYP3A4 except Cyclovalone. The NSC-43319, Salsalate, and Cyclovalone showed no specificity as substrates of CYP3C9. All of them were negative for AMES toxicity, and compounds demonstrated the inhibitions of the Human ether-a-go-go related gene (hERG) channel and also a less toxic profile.The screened ligands were inspected for the pharmacology-related study (Table S5) involving their accordance with Antibacterial, Bacterial Efflux Pump Inhibitor, Antibiotic Anthracycline-like activity, β-lactamase Inhibitor, Anti-mycobacterial and antibiotic activities. Here, Cyclovalone showed all of these pharmacological activities, whereas Go-Y032 was observed to have all of these properties except Antibiotic activity. The NSC-43319 showed antibacterial, bacterial efflux pump inhibitor, β-lactamase inhibitor, and anti-mycobacterial activities. In this study, Go-Y032, NSC-43319, and Cyclovalone were found as the best-performing ligands (Table 7). Thereafter, these five ligands were analyzed to observe whether they function against G protein-coupled receptor (GPCR) ligand, protein kinase, ion channels, enzyme protease, nuclear receptor ligand, etc., (Table S6).The most potential ligands were investigated for the prediction of their possible sites of metabolism (SOMs) by three main Cytochrome P450 (CYP) isozymes, i.e., 3A4, 2C9, and 2D6, respectively (Table 8). For CYP2D6 and CYP2C9 isoforms, Go-Y032, and for CYP3A4 and CYP2C9 isoforms, NSC-43319 have shown identical metabolism sites. However, for all of the isozymes, cyclovalone exhibited different metabolism sites.Molecular docking is the most significant approach in in silico drug designing. It assesses the binding affinity of a protein–ligand complex in the form of binding energy using computer algorithms. The lower the binding energy, the higher the affinity of the ligand bound to the target [73]. Five ligands gave the best free binding energies in the docking experiment conducted by AutoDock Vina and SwissDock, which included Go-Y032, NSC-43319 with β-lactamase enzyme, and NSC-43319, Cyclovalone with L, D-Transpeptidases (Table 5 and Table 6). The docking results were compared with controls. For the β-lactamase enzyme, the docking results with other ligands were compared with Clavulanic acid and Tazobactum (Table 9). Clavulanic acid and tazobactam are all plasmid-mediated β-lactamase inhibitors. Several studies have concluded that Clavulanic acid inhibits extended-spectrum TEM and SHV β-lactamases. They expressed lower affinity in our study compared to other ligands. Several studies reported that β-lactam antibiotics could work against L, D-Transpeptidases. Carbapenem and Cephalosporin are antibiotics in the beta-lactam class that kill bacteria by attaching to penicillin-binding proteins and blocking bacterial cell wall formation. As these are existing drugs, we compared their activities with our selected ligands. These controls were showed very low binding affinities in docking. Therefore, the ligands we selected have a better chance of working against those enzymes or as antibacterial drugs. After docking, the ligands were introduced to ADME/T prediction tools admetSAR 2.0 and pkCSM software for in silico prediction of ADME/T properties. This time-saving and cost-effective approach helps establish a candidate molecule as a promising drug through in vitro experiments [74,75]. A leading concern for drugs that mainly target the central nervous system (CNS) is that they need to be able to permeate across the blood–brain barrier. The most common route for drug delivery is the oral delivery system, through which the delivered drug enters into the intestine; hence, an investigated drug needs to be highly absorbed in the intestine. Cell membrane proteins, such as P-glycoproteins, facilitate the movement of many drugs through the cell membrane. Caco2 permeability of a drug reflects the permeability across the intestinal lining of humans, as this tissue is widely used in in vitro drug permeability studies due to its small intestinal mucosa-like behavior when cultured [75,76,77]. The Cytochrome P450 enzyme family focuses on the regulation of drug interaction, biotransformation, and their elimination outside the body. Acute toxicity, delayed removal, and eventual drug compound failure within the human body result from these enzymes’ inhibitory activity of the drugs [78,79,80]. The purpose of in silico AMES toxicity test is to determine the toxicity and mutagenicity of the chemicals [81,82]. Voltage-gated potassium ion channels, namely hERG channels, are involved in the transport of potassium ions through the cell membrane, which may be subjected to off-target drug interaction resulting in inhibition. Therefore, proper screening is necessary to see whether the investigated drugs have inhibitory activity on these transporters [83]. Renal organic cation transporter 2 (OCT2) plays a major role in the removal of drugs and xenobiotics via the kidney. It is considered that the substrates of this transporter protein are quickly excreted by urine [84]. In the ADME/T test, all the selected ligands showed almost similar properties (Table S6).Afterwards, the pharmacological and biological activities of the ligands were carried out in the PASS online server and Molinspiration Cheminformatics server, respectively. Pharmacological activity (PASS prediction) is determined in terms of the likelihood of activity (Pa) and the likelihood of inactivity (Pi) of a drug, and the result of the prediction ranges between 0 and 1. The pharmacological activity of the drug is deemed possible if Pa > Pi [85]. The probability of anti-mycobacterial activity (Pa) for NSC-43319 and Cyclovalone was between 0.5–0.7, while for all ligands, Pa of all activities was <0.5, implying the unlikeliness of their activities [86]. However, Cyclovalone, Go-Y032, and NSC-43319 showed more satisfactory outcomes than others (Tables S5 and S6). Assessing biological activities against the most influential drug targets in the human body, such as G protein-coupled receptors (GPCRs), ion channels, enzymes, nuclear receptors, etc., is crucial because, when coupled with them, a drug mediates its therapeutic activity inside the body [87]. Probability scores for Petasiphenol representing activity against the targets were comparatively significant (Table S5).Finally, the ligands were assessed in the RS-WebPredictor server to predict the probable sites where their metabolism will be likely to occur. Almost similar metabolism sites were reported for Go-Y032, NSC-43319, Cyclovalone, Difluorinated curcumin, which exhibited multiple sites of metabolism except for Cyclocurcumin and Salsalate. These two compounds showed few sites of metabolism compared to others (Table 8). After a successive screening of all potent ligands, this study suggests Go-Y032, Cyclovalone, NSC-43319, Difluorinated curcumin, Salsalate, and Cyclocurcumin as the best inhibitors of the enzymes β-lactamase and L, D-Transpeptidases, respectively, while NSC-43319, Cyclovalone, Cyclocurcumin, and Go-Y032 are recommended for both. These compounds may not express better performances at all screening tests. As our study was based on a computational approach, further in vitro analysis is needed to prove their activities. These compounds may give better performance in an in vitro approach than the controls we chose. This study aimed to develop curcumin derivatives as potential drug candidates against S. typhi causing typhoid fever. The target proteins included two crucial proteins of this bacterium, β-lactamase and L, D-Transpeptidases, which are involved in acquiring antibiotic-resistance properties and toxin secretion, respectively. Although several antibiotics and β-lactam inhibitors are available to treat this fatal illness, none is effective enough to avoid side effects that eventually result in adverse complications. After surpassing multiple stages in assessing drug-like properties, among 70 ligands, three, including Go-Y032, Cyclovalone, and NSC-43319, were reported as the best-performing ligands. Further investigations based on in vivo and in vitro experiments are needed to ascertain the use of these ligands as drugs to treat S. typhi infection. Additionally, the other four ligands providing a satisfactory docking performance are also recommended for further wet laboratory investigation.The following are available online at https://www.mdpi.com/article/10.3390/biomedinformatics2010005/s1, Table S1: The selected ligands; Table S2: Drug likeliness properties of the ligands; Table S3: ADME/T result of controls; Table S4: ADME/T prediction result; Table S5: Pharmacological activities; Table S6: Biological activities.Conceptualization, T.A., M.C., A.Y.T., S.M. and T.B.E.; methodology, T.A., M.C., A.Y.T., M.H.R., M.S.S.S., S.M., M.A.S., S.A.S. and T.B.E.; software, M.S.S.S., S.M., M.A.S., S.A.S. and T.B.E.; validation, M.H.R., S.M., M.A.S. and T.B.E.; formal analysis, S.M., M.A.S., S.A.S. and T.B.E.; investigation, T.A., M.C., A.Y.T., M.H.R., M.S.S.S., S.M., M.A.S., S.A.S. and T.B.E.; resources, M.A.S. and T.B.E.; data curation, S.M., M.A.S., S.A.S. and T.B.E.; writing—original draft preparation, T.A., M.C., A.Y.T., M.H.R., M.S.S.S., S.M. and T.B.E.; writing—review and editing, M.S.S.S., S.M., M.A.S., S.A.S. and T.B.E.; visualization, M.S.S.S., S.M., M.A.S., S.A.S. and T.B.E.; supervision, M.A.S. and T.B.E.; project administration, M.A.S. and T.B.E.; funding acquisition, S.M., M.A.S., S.A.S. and T.B.E. All authors have read and agreed to the published version of the manuscript.This research received no external funding.Not applicable.Not applicable.Available data are presented in the manuscript.The authors declare no conflict of interest. A toxin produced from cell-secreted into periplasm through SEC machinery, then these toxin subunits assemble into holotoxin-translocated into trans side of bacteria mediated by TtsA. Peptidoglycan remodeling occurs through L, D-Transpeptidases, YcbB, enriched in the bacterial poles.Methodology/Overall study.Result of β-lactamase protein (A) secondary structure prediction result using SOPMA software, (B) disease-causing region prediction using PrDOS software, (C) Ramachandran plot using PROCHECK, (D) 3D structure predicted by Swiss-Model, (E) secondary structure and disordered region predicted by Phyre2, and (F) active site predicted by COACH Software.Result of protein L, D-Transpeptidases (A) secondary structure prediction result using SOPMA software, (B) disease-causing region prediction using PrDOS software, (C) Ramachandran plot using PROCHECK, (D) 3D structure predicted by Swiss-Model, (E) secondary structure and disordered region predicted by Phyre2, and (F) active site predicted by COACH Software.Molecular docking experiments by targeting β-lactamase and L, D-Transpeptidases. Hydrogen bonds are displayed as green balls and sticks, hydrophobic bonds (Pi-Pi/Pi-sigma/amide-Pi interaction) are displayed as violet balls and sticks, hydrophobic bonds (Pi-alkyl/alkyl interaction stacking) are displayed as pink balls and sticks, hydrophobic (Pi-sulfur) are displayed as gold balls and sticks, and carbon–hydrogen bonds are displayed as white balls and sticks.The physiochemical properties of β-lactamase and L, D-Transpeptidases.The secondary structures of the proteins: β-lactamase and L, D-Transpeptidases.The homology modeling parameters for β-lactamase and L, D-Transpeptidases.The quality of the hypothetical model protein.Docking results of β-lactamase.Bold text indicates the best docking scores.Docking Results of L, D-Transpeptidases.Bold text indicates the best docking scores.Best-performing ligands after ADME/T, pharmacological, and biological activities prediction.Results of P450 sites of metabolism prediction. (Best three vulnerable atoms are marked in encircled number.)Docking result of the best-performing ligands.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00006.txt

@@ -0,0 +1 @@
+Studies have shown that STK11 mutation plays a critical role in affecting the lung adenocarcinoma (LUAD) tumor immune environment. By training an Inception-Resnet-v2 deep convolutional neural network model, we were able to classify STK11-mutated and wild-type LUAD tumor histopathology images with a promising accuracy (per slide AUROC = 0.795). Dimensional reduction of the activation maps before the output layer of the test set images revealed that fewer immune cells were accumulated around cancer cells in STK11-mutation cases. Our study demonstrated that deep convolutional network model can automatically identify STK11 mutations based on histopathology slides and confirmed that the immune cell density was the main feature used by the model to distinguish STK11-mutated cases.Non-small cell lung cancer is the most common type of lung cancer accounting for more than 80% of lung tumor malignancy cases, among which 50% are adenocarcinoma (LUAD) [1]. STK11 is a critical cancer-related gene that provides instructions for making a tumor suppressor, serine/threonine kinase 11 [2]. About 24% of all adenocarcinoma cases are STK11-mutated, and molecular studies have shown that STK11-mutation plays an important role in influencing the tumor immune environment including the intratumoral immune cell densities [1]. As a result, many researchers have suggested that precision immuno-therapy approaches should take STK11 status of individual tumors into consideration [3,4,5]. In recent years, deep-learning-based methods have been proved to be able to capture morphological features on tumor images that are associated with molecular features such as mutations, subtypes, and immune infiltration. For example, a customized multi-resolution CNN model showed its power in classifying molecular subtypes in endometrial cancer [6]. An InceptionV3-based model was able to identify BRAF mutations in malignant melanoma tissue [7]. A similar architected model was also capable of predicting non-small-cell lung cancer subtypes with high accuracy [8]. In other cancer types that are more heterogeneous such as glioblastoma and colon cancer, CNN-based imaging model also showed its power in predicting critical morphological and molecular features such as G-CIMP and MSI [9,10]. Here, we trained a deep-learning model that can determine LUAD patients’ STK11 mutation status based on histopathology slides with high performance. Visualization of the key features learned by the model confirmed that STK11 mutation is associated with the density of immune cells near cancer cells. Practically, this model is capable of providing guidance to immunotherapy in a faster, more convenient, and less expensive way by examining histopathology images without doing sequencing analyses. Inception-Renet-v2, a modified version of Inception-v4 with residual connection derived from the original InceptionNet, was used as the architecture of the deep-learning model for this project [11,12,13]. Figure 1 and Figure S1 shows the general workflow. The nature of digital histopathology images is quite different from the images from ImageNet which these CNN architectures were designed for and pre-trained on. For example, the digital histopathology images are often much larger in size than ImageNet’s. Also, the features are quite different since features in histopathology are often textures rather than objects in ImageNet. Therefore, we believe training end-to-end is a better strategy than transfer learning for our task. The 541 scanned diagnostic histopathology slides from 478 patients with STK11 mutation status were downloaded from Genomic Data Commons (GDC) of the National Cancer Institute (NCI). The data were then separated into training (80%), validation (10%), and testing (10%) sets at per-patient level. Due to the large size of the slides, they were cut into 299-by-299-pixel tiles at 20× magnification level and background was omitted. The model was trained from scratch at per-tile level with batch size of 64 and dropout keep rate of 0.3. The training process stopped when either training or validation loss did not decrease for more than 10,000 iterations to avoid overfitting. When training loss reached minimum at some point, a 100-iteration validation was performed. The model was saved as the best performing one only when both training and validation losses were at minimum. The training time took about 3 days while the testing for one slide took less than 15 min. We used the NYU Langone Health BigPurple high performance computing (HPC) platform with a NVIDIA Tesla V100 GPU and the model is also possible to be trained and tested on other platforms such as Google Colab.The model achieved per-slide level area under ROC curve of 0.795 (95% CI: 0.601–0.988) and 0.696 (95% CI: 0.692–0.7) at per-tile level (Figure 2). The top-1 accuracy with cutoff at 0.5 was 0.855 (95% CI: 0.742–0.931) at per-slide level and 0.837 (95% CI: 0.835–0.839) at per-tile level. In addition, we also tried an InceptionV3-based model, but the performance was lower with a per-slide level area under ROC curve of 0.64. Considering this is a molecular feature prediction task and the labels are at per-slide level only, we believe that these results are quite decent and successful.The activation maps before the last fully connected layer of 30,000 randomly selected tiles in the test set were recorded. These activation maps were then projected onto a tSNE plot (Figure 3). To have a more straightforward visualization of the features, we put thresholds on prediction scores and randomly selected tiles to represent their corresponding local binned areas on the tSNE space (Figure 4). An experienced pathologist with no previous knowledge in machine learning interpreted patterns in Figure 4 that tiles in the positively predicted clusters (STK11-mutated) were generally showing plenty of cancer cells with very few immune cells, while a large number of immune cells were present around the cancer cells in the negatively predicted areas (wild-type). In addition, most cancer cells were observed in the areas with high positive or negative prediction scores, suggesting that cancer cells were the main focus of the model in making decisions. These findings validated the molecular studies that STK11 mutation decreases the immune response in LUAD patients.The model we trained showed capability in predicting STK11 mutation in LUAD patients based on histopathology images. It has great potential in providing guidance to immunotherapies in a faster, cheaper, and more convenient way without any sequencing analyses. Scientifically, it confirms the molecular level findings that STK11 mutation leads to less immune response in LUAD tumor from histopathology perspective and links a critical lung cancer molecular feature to a previously unknown morphological pattern. Moving forward, we will continue working on building the connection between cancer molecular features and morphological features using deep-learning techniques.The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/biomedinformatics2010006/s1, Figure S1: general workflow.Conceptualization, R.H., W.L. and D.F.; methodology, R.H.; software, R.H.; validation, R.H. and W.L.; formal analysis, R.H.; investigation, R.H.; resources, R.H.; data curation, R.H.; writing—original draft preparation, R.H.; writing—review and editing, R.H., W.L. and D.F.; visualization, R.H.; supervision, D.F.; project administration, D.F.; funding acquisition, D.F. All authors have read and agreed to the published version of the manuscript.This work was supported by NIH/NCI U24CA210972.Not applicable.Not applicable.Genomics data and digital histopathology data can be found at Genomic Data Commons of National Cancer Institute https://gdc.cancer.gov, accessed on 24 May 2019.We would like to thank the High Performance Computing administration team at NYU Langone Health for maintaining the computational resources of this project. The authors declare no conflict of interest.The general workflow of data preprocessing, model training and evaluation, and feature visualization.Per-slide level ROC curve (left) and per-tile level ROC curve (right) of the trained Inception-Resnet-v2 model applying to the test set.30,000 tiles were randomly sampled from the test set. The activation maps before the last fully connected layer of these tiles were represented in the tSNE plot. The color of labels indicates the positive prediction scores of the tiles. Clusters of predicted STK11-mutated and wild-type tiles can be observed.Randomly selected tiles represent binned areas on tSNE space (full resolution figure in supplement). Examples of STK11 mutated and wild-type tiles are shown. Cancer cells are the main focuses in these tiles. Predicted STK11 mutated tiles show no immune cells (smaller and darker cells) around cancer cells (larger, lighter, and irregular shape cells) while plenty of immune cells are present in predicted wild-type tiles.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00007.txt

@@ -0,0 +1,23 @@
+These authors contributed equally to this work.The computational electroencephalogram (EEG) is recently garnering significant attention in examining whether the quantitative EEG (qEEG) features can be used as new predictors for the prediction of recovery in moderate traumatic brain injury (TBI). However, the brain’s recorded electrical activity has always been contaminated with artifacts, which in turn further impede the subsequent processing steps. As a result, it is crucial to devise a strategy for meticulously flagging and extracting clean EEG data to retrieve high-quality discriminative features for successful model development. This work proposed the use of multiple artifact rejection algorithms (MARA), which is an independent component analysis (ICA)-based algorithm, to eliminate artifacts automatically, and explored their effects on the predictive performance of the random undersampling boosting (RUSBoost) model. Continuous EEG were acquired using 64 electrodes from 27 moderate TBI patients at four weeks to one-year post-accident. The MARA incorporates an artifact removal stage based on ICA prior to RUSBoost, SVM, DT, and k-NN classification. The area under the curve (AUC) of RUSBoost was higher in absolute power spectral density (PSD) in AUCδ = 0.75, AUC α = 0.73 and AUCθ = 0.71 bands than SVM, DT, and k-NN. The MARA has provided a good generalization performance of the RUSBoost prediction model.Traumatic brain injury (TBI) has a tremendous impact on neurological dysfunction and death in young people (i.e., younger than 45 years old) and children (1–15 years old) worldwide [1,2,3,4]. Most TBI is graded based on initial Glasgow Coma Scale (GCS) as mild (GCS score 13–15), meanwhile approximately 8–10% is graded as moderate (GCS score 9–12) or severe (GCS score 8 or less) [5,6] when recorded during the emergency room admission [7]. The effects of TBI on brain electrical activity, due to injury on a number of ionic channels, electrical generators and network dynamics involved in the distribution and coordination of electrical energy, can be easily measured using electroencephalography (EEG). EEG records the neuronal activities with non-invasive electrodes fitted on the scalp, allowing the analysis of neuronal activity in five canonical EEG frequency bands: delta δ (<4 Hz),theta θ (4 to 8 Hz), alpha α (8 to 12 Hz), beta β (12 to 30 Hz) and gamma γ (<30 Hz). The electroencephalogram can provide invaluable information regarding the instantaneous changes of brain electrical activity, specifically in aiding neuroprognostication of TBI [8,9,10,11].Numerous research groups have been working on developing a TBI-diagnostic model based on quantitative EEG (qEEG) features (i.e., differentiation between mild TBI (mTBI) and no mTBI) [12,13,14,15,16,17,18] but have yet to develop a TBI-prognostic (i.e., prediction TBI recovery) tool effectively that has gained widespread attention. Quantitative EEG analyses use computationally derived features that highlight specific components of EEG with numerical values [19]. Predictive models are statistic models that incorporate patient data to anticipate outcomes and are more robust than simple clinical judgments [2,20]. There have been numerous reports of prognostic models, but none are widely used. It may be because the validity and usefulness of predictive models in TBI have not been demonstrated with sufficient clarity and certainty to convince clinicians of their potential added value. A systematic review offers possible explanations [20].Modern advances have shown that EEG is a prospective neuroimaging modality for accurate prognostication of patients after moderate to severe brain injuries [8,21,22,23]. Advancements in computational EEG signal processing have significantly improved the reliability and validity of electrophysiological brain measurements. The preprocessing of EEG data is tedious and labor-intensive as the recorded EEG data is usually long, and analyzing raw data through visual inspection is time-consuming. Hence, there is a need for an automated system to perform the analysis (i.e., feature extraction, feature selection, and classification).Several surveys and studies have been conducted, and the automatic learning methods (i.e., machine learning (ML)) have proven their effectiveness in recognizing EEG wave patterns [23,24,25,26,27,28,29,30]. A key advantage of ML is manipulating multimodal objectively; and modeling hidden relationships in complex datasets with heterogeneous distribution using advanced mathematical techniques [31,32]. The learning strategy is particularly based on supervised learning (i.e., the algorithms learn from labeled training data to create a model that can generate predictions based on unknown data) and unsupervised learning (i.e., the algorithms analyze and cluster unlabeled data, and discover hidden patterns of data clusters without the need of human intervention). More in-depth descriptions of ML and its limitations can be found in references [19,25,32,33,34].The procedure for building a TBI predictive outcomes model using the continuous EEG data typically involves preprocessing the raw recordings to mitigate the low signal-to-noise (SNR) ratio in order to obtain a more accurate representation of the primary brain activity. The data confounded with noise or artifacts such as eye blinks, muscular movements, and other instrumentation noises may not correctly represent the underlying brain signals [19]. In the literature, independent component analysis (ICA) has been investigated as a chosen technique for artifact rejection to improve the quality of EEG signals. The ICA has been widely used in EEG signal analysis and brain-computer interface (BCI) [31,32,33]. Khoshnevis and Sankar [34] confirmed that the blind source separation in ICA allows estimation of independent components (ICs) from multiple mixed observations without prior knowledge about brain activity to remove correlation between the channels [35]. Lee et al. [36] argued that the gold standard for EEG review (i.e., traditional approach) is a manual inspection by experts, but ICA algorithms (i.e., automatic approach) could produce EEG with higher signal quality.With ICA, the signal sources are assumed to be instantaneous linear mixtures of cerebral and artifactual sources that can be decomposed into ICs. Once the ICs have been extracted from the original signals, the clean signal is reconstructed by discarding the ICs that contain artifacts. Vigário [37] tested the ICA method on simulated and experimental data and found that it performed well in the separation of signals from their linear mixtures and the extraction of eye information from electrooculography (EOG) signals [38]. Romero et al. [39] used ICA to reduce EEG artifacts at various sleep stages and discovered that the bidirectional property of EEG and EOG had little effect on ICA. Therefore, noise reduction is a compulsory technique as this method will influence the computation of qEEG features (e.g., power spectral density (PSD), coherence or connectivity). If the extracted features do not precisely designate the essential signals, a classification algorithm employing such features might have problems in identifying the classes of the features. Methodological differences of artifacts removal make it challenging to extract accurate qEEG features and they also pose a problem in assessing the reproducibility of the ML models across restricted datasets.Recent studies revealed that the ML based on qEEG characteristics yields superior performance in classifying the outcomes of TBI patients. The results highlighted that the ML algorithm (i.e., random under-sampling boosting decision trees (RUSBoosted Trees)) that uses qEEG features (i.e., PSD in specific frequency band (e.g., δ, θ, α and γ)) demonstrated promising results to predict outcomes of highly imbalanced moderate TBI dataset [40]. The present study extends our prior work [40] by including a modification of adding an automatic artifacts rejection method (i.e., multiple artifact rejection algorithm (MARA)), an independent component analysis (ICA)-based algorithm in EEG preprocessing steps and exploring their effects on the predictive performance of the RUSBoost prediction model.This paper consists four main sections: Section 1 introduces the present study, including some backgrounds and literature review. Section 2 presents the dataset, proposed methodology, and its performance evaluation. Section 3 includes the results and discussion. Finally, Section 4 concludes this paper.Patient recovery assessment was conducted through telephone calls by physicians from four weeks to one-year after the accident. The Glasgow Outcome Scale (GOS) was used as the primary outcome measure, which was dichotomized as a bad outcome (i.e., GOS score of 1–4) and a good outcome (i.e., GOS score at 5), in approximately 12-months after injury. In this study, an expert (i.e., neurosurgeon) in our team evaluated the neurological outcomes of moderate TBI patients based on GOS score (given in Table 1) that corresponded to the specific level of improvement of each patient [41,42].Continuous EEG eyes-closed data from 27 moderate TBI patients (B1–B27) were obtained from 64 EEG electrodes to record the brain’s signals from 64-sites on the scalp at a sampling rate 1 kHz. Electroencephalograms of 27 moderate TBI patients (n = 27) were collected at the Hospital Universiti Sains Malaysia, Kubang Kerian, Kelantan, Malaysia. Ethical clearance for this study was attained from the Human Research Ethics Committee, Universiti Sains Malaysia (USM), with a permission number USM/JEPeM/1511045.Nonsurgical moderate TBI patients aged 18 to 65 met the inclusion criteria. The first hit involved the left or right hemisphere, which was confirmed by a computerized tomography (CT) scan at the time of diagnosis. Criteria of ages under 18 years old, serious scalp and skull deformities, bone fractures, and drug use were all ruled out. All patients had given informed consent before participating in this study. All of the patients in this study were men who had sustained a TBI due to vehicle accidents. Table 2 provides a detailed explanation of the characteristics of moderate TBI patients for this study.The EEG signals were obtained using 64 electrodes to record brain signals from 64 different locations on the scalp. All electrodes were placed following an international standard of 10–10 electrode configuration [16,43]. The CPz (i.e., equivalent to Ch-32) was set as an EOG channel for tracking the eye movement and blinking artifacts. As a result, only 63 EEG channels were used as input data in our classification model.The patient ground electrode serving as reference electrode was placed at 10% frontward to Fz connected to earlobes. A programmable direct current (DC) broadband SynAmps amplifier was used to measure the brain signals by boosting up to 2500 gain and precision of 0.033/bit in the recording range of 55 millivolts (mV) at the DC 70 Hz frequency range. The 16-bit alternating current (AC)-DC converters were used for digitizing EEG signals to 1 kHz. During the recording, each patient was sat in a comfy seat in front of a computer screen in a dimly lit room and advised to remain motionless and close their eyes to remain task-free (i.e., no tasks or activities performed) for 350 s.The continuous EEG eyes-closed data were obtained from moderate TBI patients with follow-up visits. The first measurement (i.e., four–10 weeks post-accident) contributed to thirteen moderate TBI data. Eleven moderate TBI EEG data were contributed from the second measurement (i.e., six-months post-accident). The third measurement (i.e., one-year post-accident) contributed to three EEG data. The patients would be rejected if they failed to participate in the follow-up EEG measurements within the given time frame.The raw, unprocessed EEG data were exported for further analysis. The patient would be disqualified if they failed to participate in EEG measurement within the time frame. The raw, unprocessed EEG data were exported for further analysis. Figure 1 depicts a block diagram of the suggested methodology.The continuous eyes-closed EEG data were preprocessed in MATLAB R2020a (MathWorks, Natick, MA, USA) using open-source EEGLAB toolbox version 2019.0 [44,45] with the custom MATLAB script. The details about the preprocessing script are given as follows. We assigned the 63-channels continuous eyes-closed EEG signal to be presented as a vector x=[x1,x2,x3,x4,⋯,x61,x62,x63], where xm is the 60 segments (i.e., 1-s) from the m-channel of EEG signal (i.e., m = 63 channels).The power line noise was eliminated in the first step by applying a notch filter at 50 Hz (in Malaysia) on the signal x with EEGLAB function (i.e., pop_ eegfiltnew()). Next, the ICA was performed, and artifact-related components were rejected according to the MARA [46,47] to remove the EOG artifacts (e.g., eye blinks, eye movements), electromyogram (muscle) artifact, and electrocardiograph (EKG) artifactual activity components. MARA is a supervised ML algorithm that automates artifact removal by hand-labeling ICA components. For this purpose, MARA employs a classifier (i.e., a linear programming machine) that discriminates an ICA component that is not derived from brain activity. The MARA classification system is based using two linear classifiers for finding a separating hyperplane (P), which is mathematically solved by Equation (1).
+        (1)P=sign(wv+b){−1,1}ICA components are classed as neural or artifact w is a weight vector taken from labeled training data samples, v is a feature vector, and b is a bias factor [47,48].The MARA provides visualization of each component scalp map (see Figure 2a,b), its spectrum, respectively (see Figure 2c,d), and the current label of the component (i.e., artifact or neural) by presenting each component’s probability based on the six features (i.e., current density norm, range in pattern, mean local skewness, λ, fit error and 8–13 Hz) as feature selection procedure described in [47]. If the artifact probability is greater than 0.5 (p-artifact = 0.99, see Figure 2e), it is considered as an artifact, and if the artifact probability is less than 0.5 (p-artifact = 0.00), it is considered as a neuronal signal (see Figure 2f). Features that contribute to a component marked as an artifact are plotted as red bars, and features that indicate the component containing neuronal activity are marked in blue. These features are invaluable in understanding the MARA’s decision.In this study, IC17 is a typical eye blink artifact shown by intense frontal activity steep power spectrum; IC32 is a neuronal component showing the alpha peak around 10 Hz. The scalp map indicates the occipital brain source. We can choose to remove ICs automatically without inspection, but in this study, ICs flagged as artifacts will be rejected based on the artifacts probability and scalp map computed by MARA to produce a clean signal y. After that, a band-pass filter with cutoff values at 0.1 Hz and 100 Hz was applied to the clean signal y to remove any undesired peaks with extreme signal values. The output signal was represented as the signal z. In order to cover multiple informative frequency bands (i.e., δ, θ, α, β,and γ), the frequency analysis is limited to a range of 0.1 to 100 Hz as recommended by McNerney et al. and van den Brink et al. for TBI classification [12,18].The first 60 s of recording were deleted due to artifact contamination. The next 60 s of EEG data were then split into 60 fragments of one second. The fragmentation began at a 60,001 miliseconds because the first 60 s had been rejected. The input signal z was matrix-arranged (i.e., the amplitude of the EEG channel × time) following to the default arrangement of the 64-channel WaveGuard EEG helmet cap. The M×Fs is the signal z, where M is the number of EEG channels (i.e., M = 63 channels) and Fs (i.e., = 1 kHz) is the sampling rate. Therefore, there were 1620 segments of data (i.e., 27 recordings × 60 segments/recording).Feature extraction is critical stage in any EEG analysis that identifies common feature representations among EEG samples. The absolute frequency bands were determined by integration of the PSD inside each frequency band: δ (0.5–4 Hz), θ (4–7 Hz), α (7–13 Hz), β (13–30 Hz) and γ (30–100 Hz). Based on Equation (2), the Fourier transform (FT) of z^(ω) over an interval of [0,T] is calculated to assess the frequency content of the input signal z(t) based on Equation (2).
+        (2)z^(ω)=1T∫0Tz(t)exp−iωtdzThe average PSD is subsequently computed for each EEG frequency band of each channel. Equation (3) can be used to represent the generic PSD, Pzz(ω) of a given signal z(t), where E is the estimated value and T is the time range for the PSD, Pzz(ω).
+        (3)Pzz(ω)=limT→∞E[|z^T(ω)|]2PSD is a frequently extracted characteristic that quantifies the power contained within a frequency domain signal. It is similar to the FT of a signal’s auto-correlation function [19,49]. Therefore, each EEG sub-band had an average of 63-PSD in one segment. For each frequency band, the 63-PSD average was convolved to generate a feature vector (e.g., Pβ=[Pβ1,Pβ2,⋯,Pβ63]).RUSBoost is a hybrid sampling-boosting model that balances the characteristic of classes by minimizing instances from dominant classes [50,51]. Since the dataset distribution were heavily skewed, a robust classification algorithm was required which would work well with such a skewed dataset. The implementation of the RUSBoost as a classifier for predicting moderate TBI outcomes is feasible since the epochs reflecting poor and good outcomes are not uniformly distributed in both the training and testing datasets. A skewed dataset slows the classifier’s learning rate for the poor outcome, as most data corresponds to the good outcome. As a result of this imbalance, the model’s predictions are biased towards the majority class, resulting in a decrease in the model’s overall performance [52].A majority class is known as a negative class and constitutes the maximum of the dataset. With increasing samples in the negative class, learning becomes more complicated since ML classifiers for used learning purposes in such imbalance datasets may disregard positive class (i.e., minority class) samples as noise or outliers. In the majority class (i.e., good), the predictive model leans towards better accuracy, meanwhile poorly performing on the side of the minority class (i.e., poor). The examples of the minority class are misclassified at a higher rate than examples of the other classes. The steps for RUSBoost implementation is described in Algorithm 1 [50].The present study implemented the RUSBoost algorithm with decision tree (DT) as the weak learner. By applying the RUSBoost, the resampling method (i.e., undersampling) handled the imbalanced dataset problem by altering the minority and majority class size to provide a balanced distribution in a training dataset. Boosting leverages the random samples of the data to create each tree where each sample is balanced because the algorithm undersamples the majority class to match the size of the minority class. Due to the minimal number of epochs representing poor outcomes (18.52%), the number of weak learner, that is, 30 trees, were utilized in the final models with a number of 20 splits, and a learning rate of 0.1 was merged into a high-quality ensemble predictor using the base function fitensemble to build a RUSBoost prediction algorithm in prediction recovery of moderate TBI.
+        Algorithm 1 RUSBoost Algorithm.Input: Given a set R of examples  (x1,y1), (x2,y2), ⋯, (xm,ym) with minority class yr∈Y,|Y|=2.   Weak learner (decision trees), WeakLearn.   Number of iteration, T   Desired percentage of total examples to be represented by the minority class, N1:Initialize W1(i)=1/n, for all i.2:fort=1,2,⋯,Tdo3:   Create temporary training dataset Rt′ with distribution Wt′ using random under-sampling4:   Train a WeakLearn, providing it with samples Rt′ and their weight Wt′5:   As a result, get a hypothesis ht:X×Y→[0,1].6:   Compute the pseudo-loss for R and Wt′:
+                    ϵt=∑(i,y):yi≠yWt(i)1−ht(xi,yi)+ht(xi,y)7:   Compute the weight update parameter:
+                    ζt=ϵt/(1−ϵt)8:   Update Wt(i) for each sample:
+                    Wt+1(i)=Wt(i)ζt12(1+ht(xi,yi)−ht(xi,y:y≠yi))9:   Normalize the weights Wt+1:         Wt+1(i)←Wt+1(i)/∑i=1HWt+1(i)10:end forOutput: The final hypothesis:      Sfinal(x)=arg maxy∈Y∑t=1Tht(x,y)log10(1/ζt)Input: Given a set R of examples  (x1,y1), (x2,y2), ⋯, (xm,ym) with minority class yr∈Y,|Y|=2.   Weak learner (decision trees), WeakLearn.   Number of iteration, T   Desired percentage of total examples to be represented by the minority class, NInitialize W1(i)=1/n, for all i.fort=1,2,⋯,Tdo   Create temporary training dataset Rt′ with distribution Wt′ using random under-sampling   Train a WeakLearn, providing it with samples Rt′ and their weight Wt′   As a result, get a hypothesis ht:X×Y→[0,1].   Compute the pseudo-loss for R and Wt′:
+                    ϵt=∑(i,y):yi≠yWt(i)1−ht(xi,yi)+ht(xi,y)   Compute the weight update parameter:
+                    ζt=ϵt/(1−ϵt)   Update Wt(i) for each sample:
+                    Wt+1(i)=Wt(i)ζt12(1+ht(xi,yi)−ht(xi,y:y≠yi))   Normalize the weights Wt+1:         Wt+1(i)←Wt+1(i)/∑i=1HWt+1(i)end forOutput: The final hypothesis:      Sfinal(x)=arg maxy∈Y∑t=1Tht(x,y)log10(1/ζt)In order to evaluate the performance of the proposed recovery prediction system, we used a k-fold cross-validation process. To do this, the dataset was divided randomly into k equal-sized subdivisions. At each fold, the k–1 subdivisions were used for training, and the remaining partitioning was used for testing. This process is repeated for k-times (i.e., k = 5). The results of five partitions were averaged and reported as the system performance.Instances in a binary classification task can be labeled as either positive or negative. The minority class is usually regarded as a positive class in binary poorly balanced datasets, while the majority class is usually considered negative. This research classified the poor and good outcomes as positive and negative cases, respectively, resulting in the classification matrix shown in Table 3.True positive (TP) and true negative (TN) are valid predictions, but false negative (FN) and false positive(FP) are wrong predictions. Analyzing the four entries in the confusion matrix does not suffice to assess the classifier’s performance. The confusion matrix provided four types of performance measurements that were used in this study:The sensitivity, often referred to as the True Positive Rate, TPrate, is expressed in terms of:
+            (4)TPrate=TP/(TP+FN)TPrate indicates the ability of a classifier to identify a positive class correctly. It ranges from 0 to 1, with 1 being the perfect score.The specificity, alternatively referred to as the True Negative Rate, TNrate, is determined as;
+            (5)TNrate=TN/(TN+FP)TNrate denotes the ability of a classifier to identify a negative class correctly. The perfect score is 1, and 0 is the worst measure.G-mean (geometric mean), is denoted as;
+            (6)G-Mean=TPrate*TNrateG-Mean introduced by [53] quantifies the ability of a classifier to balance classification accuracy between positive and negative classes. By combining the G-Means of TPrate and TNrate, a low G-Mean score indicates a highly discriminative classifier toward one class and vice versa.F1 score describes the trade-off between precision (TP/(TP+FP) and recall(TP/(TP+FN) in the positive class. This well-known metric is perfectly suitable for skewed dataset problem that is determined as;
+            (7)F1score=2*(precision*recall)/(precision+recall)It is a numeric value between 0 and 1, with 1 representing the perfect value.Area Under Curve (AUC) is a popular overall model performance evaluation, especially for rating binary classifiers in the presence of class imbalance. Receiver operating characteristic (ROC) equals to AUC. To generate the ROC curve, we plotted the TPrate against the false positive rate FPrate, which is calculated as follows:
+            (8)TPrate=TP/(TP+FN)
+            (9)FPrate=FP/(FP+TN)It should be noted that higher AUC values imply a better ROC curve and, thus resulting in better performance.The sensitivity, often referred to as the True Positive Rate, TPrate, is expressed in terms of:
+            (4)TPrate=TP/(TP+FN)TPrate indicates the ability of a classifier to identify a positive class correctly. It ranges from 0 to 1, with 1 being the perfect score.The specificity, alternatively referred to as the True Negative Rate, TNrate, is determined as;
+            (5)TNrate=TN/(TN+FP)TNrate denotes the ability of a classifier to identify a negative class correctly. The perfect score is 1, and 0 is the worst measure.G-mean (geometric mean), is denoted as;
+            (6)G-Mean=TPrate*TNrateG-Mean introduced by [53] quantifies the ability of a classifier to balance classification accuracy between positive and negative classes. By combining the G-Means of TPrate and TNrate, a low G-Mean score indicates a highly discriminative classifier toward one class and vice versa.F1 score describes the trade-off between precision (TP/(TP+FP) and recall(TP/(TP+FN) in the positive class. This well-known metric is perfectly suitable for skewed dataset problem that is determined as;
+            (7)F1score=2*(precision*recall)/(precision+recall)It is a numeric value between 0 and 1, with 1 representing the perfect value.Area Under Curve (AUC) is a popular overall model performance evaluation, especially for rating binary classifiers in the presence of class imbalance. Receiver operating characteristic (ROC) equals to AUC. To generate the ROC curve, we plotted the TPrate against the false positive rate FPrate, which is calculated as follows:
+            (8)TPrate=TP/(TP+FN)
+            (9)FPrate=FP/(FP+TN)It should be noted that higher AUC values imply a better ROC curve and, thus resulting in better performance.The ROC curve and its AUC, TPrate, and TNrate of the evaluated RUSBoost prediction model on moderate TBI data were shown in Table 4. The ROC curve for each frequency band computed from the absolute PSD is shown in Figure 3 and Figure 4. In most cases, the AUC value ranges from 0.5 to 1, where 0.5 means the algorithm performs the same as the chances of flipping a coin. The AUC values were low in absolute PSD of β (i.e., AUCβ = 0.51) (i.e., see Figure 4b) and γ (i.e., AUCγ = 0.54) (i.e., see Figure 4c) bands; indicating that the TPrate and TNrate of these frequencies bands were low (i.e., TPrateβ = 40.0%, TNrateβ = 54.5% (i.e., see Figure 4e); TPrateγ = 80.0%, TNrateγ = 54.5% (i.e., see Figure 4f). The absolute PSD in δ solely achieved the most significant prediction performance with AUCδ values was 0.75 (i.e., see Figure 3a), demonstrating that our proposed RUSBoost prediction model is perfect for dealing with our imbalanced dataset distributions. The AUC values of absolute PSD in θ had AUCθ = 0.71 (i.e., see Figure 3b), and α had AUCα = 0.73 (i.e., see Figure 4a), respectively. The TPrate and TNrate of absolute PSD in δ (i.e., TPrateδ = 80.0%, TNrateδ = 63.6%) and α (i.e., TPrateα = 80.0%, TNrateα= 59.1%) bands were high ((i.e., see confusion matrix in Figure 3c and Figure 4d); indicating their good prediction performance at discriminating the TBI outcomes.The results suggested that the prediction of recovery based on the RUSBoost algorithms had efficiently distinguished between patients with poor and good outcomes with higher AUC values in absolute PSD of δ, α, and θ bands. The absolute PSD in δ, α, and θ bands provided the most significant predictive value and were the best predictors for predicting the recovery outcomes of moderate TBI.The G-Mean is the most acceptable metric to replace the accuracy rate due to the uneven distribution [54,55]. The accuracy rate is a traditional performance indicator for a predictive model with a perfectly balanced class distribution [56]. Based on the results, we found the G-Mean most suited for balancing poor and good outcome classes in terms of total classification accuracy (see Table 4). More specifically, the RUSBoost prediction model contributed the maximum G-Mean (%) and F1 scores in δ, θ (G-Mean (%) = 71.33), meanwhile the G-Mean (%) and F1 scores in α (G-Mean (%) = 68.76) were at the above-average level. The best G-Mean reflects a balanced prediction performance on both positive (i.e. bad outcomes) and negative classes (i.e., favorable outcomes).The F1 scores for the δ and θ bands were good, indicating that the F1 scores are insensitive to FN, and therefore, it accurately measures the quality of an algorithm for predicting the TP. The final PSD in the β (G-Mean (%) = 46.69) band indicates poor performance in predicting the positive and negative classes. The absolute PSD resulted in higher AUC values, and G-Mean above 68.7% suggested the suitability of RUSBoost to predict the moderate TBI outcomes.In addition, for imbalanced dataset problems, the present study provides a performance comparison of three different ML classifiers (i.e., support vector machine (SVM), DT, and k-nearest neighbor (k-NN)) to predict the outcomes of moderate TBI. The results confirmed a superiority in AUC value and a balanced classification performance (i.e., G-Mean (%)) for the RUSBoost over other algorithms. The standard rule-based classifiers (i.e., SVM, DT, and k-NN) had demonstrated algorithm discrimination between a single class of positive and negative (see Table 5 (DT), Table 6 (SVM) and Table 7 (k-NN) for prediction performance comparison).In summary, the results showed that the RUSBoost prediction model was the most suitable for predicting recovery of moderate TBI patients. The AUC values were low in absolute PSD of the following models: k-NN, DT, and SVM; representing that the TPrate, TNrate and G-Mean (%) and F1 scores of these models are low for all frequency bands. The AUC values of absolute PSDβ of the DT algorithm were higher than the AUC of the RUSBoost, SVM, and k-NN. However, it could not indicate its superior prediction performance because the classifier was biased towards a negative class (i.e., good outcomes).In the previous section, we classified the outcomes of moderate TBI using DT, SVM, k-NN, and the RUSBoost. The recovery models were computed based on the absolute PSD in five sub-bands to evaluate the most successful qEEG features in predicting TBI outcomes. We discovered that the RUSBoost prediction model generally outperforms the the DT, SVM, and k-NN, considering only the optimal total accuracy rather than the distribution across different classes. However, this condition may be explained because a classification algorithm that tries to maximize accuracy to meet its objective rule will produce an accuracy of 99% just by correctly classifying all samples from the larger class but misclassifying one sample of the smaller class. As illustrated in Figure 5 and the AUC values in Table 5, Table 6 and Table 7, the ensemble decision trees (i.e., RUSBoost) outperformed the individual classifier.The RUSBoost adopted a hybrid approach from AdaBoost [57] (i.e., adaptive boosting) algorithms that use the combination of sampling and boosting, aiming to achieve higher performance for the dataset with the class imbalance problem [51,58]. As seen from the dataset we used in this study, it is unbalanced, which is why the RUSBoost classifier achieved the highest prediction performance. The learning strategy of the RUSBoost algorithms offered advantages in improving the prediction of the poor outcomes with a slight decrease in the good outcomes class. The undersampling strategy, which balances the class distribution in the dataset, is highly beneficial in learning from skewed training data [40,55]. In sleep spindles detection [59], the RUSBoost algorithms enable an automatic sleep spindles detection with an F-measure of 0.70 and sensitivity of 76.9% without requiring threshold calibration. RUSBoost used majority voting of weak classifiers for discrimination spindles from the extracted EEG features (i.e., synchrosqueezeed wavelet transform (SST)).In comparison to non-sampling techniques [12], recent findings imply that the hybrid strategy with resampling (i.e., undersampling) and boosting can significantly improve model performance [34,60,61,62]. The ensemble DT technique is more flexible and less prone to overfitting (i.e., has a high bias but low variance), demonstrating the generalization power of RUSBoost in predicting outcomes. The present results support the previously reported development of a predictive model using the ensemble DT and resampling is better suited in predicting TBI outcomes than using an individual algorithm (i.e., DT, SVM and k-NN) [40].In this work, the automated artifacts rejection method, which is the MARA, was performed on our continuous EEG of moderate TBI data to separate the contributing sources to the scalp EEG [47]. Artifacts in EEG signals might make interpretation difficult and lead to incorrect analytical judgments. Numerous algorithms and preprocessing pipelines have been developed to address the problem of artifact rejection in electrophysiological data [32,49,63,64,65,66,67]. Each of these algorithms has its own set of strengths and focuses on a different area of artifact rejection than the others. In recent studies, Pedroni et al. [68] suggested that applying a preprocessing pipeline of algorithms to detect defective channels in combination with MARA, which is an ICA-based artifact rejection method, effectively removes a large extent of artifacts.The MARA was initially designed to distinguish the ICA components that originated from the brain and non-brain sources and reject artifactual ICA components in EEG data [46]. The development of the prediction outcomes algorithm is based on an automated EEG artifact rejection method (i.e., MARA) shown to be efficient in identifying artifacts with intelligence ICs selection and provided a good generalization performance of the RUSBoost prediction model. The model obtained the highest prediction performance in δ band (i.e., AUC = 0.75, G-Mean (%) = 71.33) compared to other classifiers. Our results support the claims made in the literature [33,48,68,69] that the automated ICA-artifact preprocessing pipeline offered substantial benefits by increasing consistency and efficiency for classifying artifacts or non-artifacts ICs. On the contrary, Alam et al. [49] have found that the selection of artifact removal had distinct effects on the PSD calculation. However, the recent findings in Noor et al. [40] showed that the artifacts rejection by the automatic continuous rejection and experts confirmation provided promising results in predicting TBI outcomes of specific frequency bands. Although an automated artifacts rejection can help isolate between the neural or non-neural source components from ICA decomposition, subjective method (i.e., visual classification by experts) is still typically advisable [33,70].Designing a clinically useful predictive model is difficult due to the high complexity of EEG measurements. Preprocessing and feature extraction must be done carefully to ensure that high-quality discriminative features can be retrieved to attain a high model performance. Therefore, feature selection and suitable ML algorithms are required to deduce the significant qEEG predictors. Poor data extraction thereby directly affects the accuracy of classification. The great majority of research included in the review [23,28] found that identifying the most informative qEEG that characterizes the recovery outcome level is crucial to ensure early targeted prediction after TBI post-injury. More recently, Noor et al. [40] have demonstrated that qEEG features of PSD in δ, θ, α, and γ showed promising results in the prediction of recovery outcome of moderate TBI patients. A key strength is that the frequency distribution of neuronal activity provides information on the patient level of arousal, restful alertness, and general capacity for focus mental activity [8]. In support, the specific qEEG features are confirmed as invaluable predictors of recovery in TBI, which can complement demographic and clinical information [4,9,71,72].Several studies have suggested that spectrum EEG features may predict the level of consciousness in patients suffering from the disorder of consciousness (DOC) following severe TBI. In comparative studies, a significant reduction in the amplitude oscillations of the α and β bands among patients with DOC but there was a concurrent improvement in θ and δ amplitude for fully conscious participants [73,74,75]. The systematic review by Pauli et al. [22] examined a number of studies exploring continuous EEG as a prognostic measure in DOC following TBI. The resting-state EEG (i.e., continuous EEG) is particularly promising within the 12-month post-injury. Several studies have shown that the α power and variability are significant for modeling the functional outcomes during periods [72,76,77].The α and δ power were extracted from the EEG to be utilized in the random forest (RF) [9], generalized linear model (GLM) [11], and linear regression (LR) [10,72] training features in the previous studies to predict outcomes in severe TBI patients. They have a lower classification performance than our proposed model method (i.e., RUSBoost) [40]. However, we believe this cannot be a fair indicator for a method comparison because their proposed algorithms were mainly suitable for balanced TBI data distribution. The similarities in our findings suggest that modeling prediction models based on computational EEG approaches (i.e., ML and qEEG features) allowed the researchers to identify the most explanatory predictors for a reliable TBI outcome prediction. In support of this, ref. [74] found that spectral density in a specific frequency band provides a strong connection between severe TBI outcomes. The results highlighted that spectral density at different frequency bands has the utmost predictive value, especially two to three months after injury [9,10,11,72]. Overall, this work demonstrates that the association between PSD features in a specific frequency and the clinical outcomes (i.e., GOS scores) is robust enough to develop a reliable TBI prediction outcomes model with multiple combinations of qEEG features and different ML approaches.In conclusion, this study presented a RUSBoost prediction outcomes model that integrates MARA ICA-based into the resting-state eyes-closed EEG preprocessing of moderate TBI data to eliminate artifacts automatically. The prediction performance obtained and reported in this paper is higher than previous studies [8,9,10,66] in predicting TBI outcomes; however, the model’s performance is lower than our prior work [40]. A robust model performance requires rigorous EEG preprocessing and feature extraction techniques to ensure the retrieval of high-quality discriminative features. In addition to that, distortions in EEG signals can significantly diminish the trustworthiness of clinical decisions based on the signals. We believe that the development of a moderate TBI outcomes prediction model based on MARA for automatic tracking and eliminating artifactual ICA components has been demonstrated to be effective in identifying artifacts with intelligence ICs selection, thus providing a good generalization performance of the RUSBoost prediction model. The robustness of RUSBoost algorithms (i.e., ensembles DT) compensated for the inadequacies of single classifiers (i.e., DT, SVM, and k-NN) in classifying the outcomes even with small samples and a minimal set of qEEG features (i.e., PSD). Future research could involve predictive modeling with various parameters (e.g., coherence, connectivity, relative power, spectrum asymmetry) to classify the unique qEEG properties to moderate TBI outcomes.Conceptualization, designed methodology and implemented the algorithm, writing—original draft preparation and revision, N.S.E.M.N.; proposed the idea, editing and supervision, H.I.; writing—review and revision, editing, M.H.C.L.; project administration and funding acquisition, J.M.A. All authors have read and agreed to the published version of the manuscript.This research was funded by the Ministry of Higher Education (MoHE), Malaysia, via the Trans-disciplinary Research Grant Scheme (TRGS) with grant number TRGS/1/2015/USM/01/6/2 and in part by MoHE through Skim Latihan Akademik Bumiputra (SLAB).The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Human Research Ethics Committee, Universiti Sains Malaysia (USM), with an approval number USM/JEPeM/1511045.Informed consent was obtained from all subjects involved in the study. Written informed consent has been obtained from the patient(s) to publish this paper.The authors would like to thank Hazim Omar, Muhammad Ridda for collecting the experimental data, Diana Norma Fitzrol for providing the GOS scores for clinical TBI outcomes assessments and Lai Chi Qin for discussion on algorithm development.The authors declare no conflict of interest.Block diagram of the proposed method.Visualization independent components (ICs) with MARA for automatic artifact rejection. Figure (a) shown the IC17 scalp map for the the eye blink artifact. Figure (b) shown the IC32 scalp map of the neuronal component. The steep power spectrum shown in figure (c); the alpha peak around 10 Hz power spectrum shown in figure (d). Figure (e,f) shown each component of six features that the MARA classification.ROC curve of (a) δ and (b) θ bands and confusion matrix of (c) δ and (d) θ bands shown the ability of RUSBoost prediction model to discriminate between poor and good outcomes on TBI data based on absolute PSD, as measured by AUC.ROC curve of (a) α, (b) β and (c) γ bands and confusion matrix of (d) α, (e) β and (f) γ bands show the ability of RUSBoost prediction model to discriminate between good and poor outcomes on TBI data based on absolute PSD, as measured by AUC.An evaluation of the ROC curves of proposed models in predicting recovery of moderate TBI patients using the RUSBoost, SVM, k-NN and DT in δ band. The ROC curve shows the most significant absolute PSD in δ band of RUSBoost model.Clinical Description of GOS Score.Details in Jennett and Bond [41,42].Moderate TBI Dataset (adapted from Noor et al. [40]).Abbreviation: B (TBI Patient ID); CT (Computed Tomography); F/M (Female/Male); F-T (Frontal-Temporal); GCS (Glasgow Coma Scale); GOS (Glasgow Outcome Scale); L-T (Left-Temporal); L-P (Left-Parietal); L-FTP (Left-Frontal-Temporal-Parietal); P (Parietal); R-F (Right-Frontal).Confusion matrix of binary classification problem.RUSBoost prediction model based on absolute PSD in five EEG frequency bands.Comparison of prediction outcomes results for DT.Comparison of prediction outcomes results for SVM.Comparison of prediction outcomes results for k-NN.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00008.txt

@@ -0,0 +1,6 @@
+These authors contributed equally to this work.In clinical practice, every decision should be reliable and explained to the stakeholders. The high accuracy of deep learning (DL) models pose a great advantage, but the fact that they function as black-boxes hinders their clinical applications. Hence, explainability methods became important as they provide explanation to DL models. In this study, two datasets with electrocardiogram (ECG) image representations of six heartbeats were built, one given the label of the last heartbeat and the other given the label of the first heartbeat. Each dataset was used to train one neural network. Finally, we applied well-known explainability methods to the resulting networks to explain their classifications. Explainability methods produced attribution maps where pixels intensities are proportional to their importance to the classification task. Then, we developed a metric to quantify the focus of the models in the heartbeat of interest. The classification models achieved testing accuracy scores of around 93.66% and 91.72%. The models focused around the heartbeat of interest, with values of the focus metric ranging between 8.8% and 32.4%. Future work will investigate the importance of regions outside the region of interest, besides the contribution of specific ECG waves to the classification.Deep learning (DL) models have had an increasing impact on today’s scientific research. DL models have achieved state of the art results in many fields, such as image classification or natural language processing. However, they lack transparency, i.e., it is not possible to explain their results. For that reason they are often referred to as black-box models. In addition, there are questions about the fairness of the models. For example, lack of fairness in the medical domain may be introduced by bias towards some aspect of the person, such as, gender, ethnicity, sexuality, or disability [1]. The inability to explain DL models is a major handicap, since one cannot understand possible bias towards specific attributes. It prevents these models to be widely used in sensitive tasks, such as, autonomous driving or medical diagnoses. Government organisations and institutions have shown their concern about this issue and a series of guidelines about methods to improve the transparency of artificial intelligence models have been created [2]. For example, the European Commission published a technical report, titled, Robustness and Explainability of AI [3] where the authors emphasised the importance of standardisation and certification tools for AI in order to create AI applications that are more robust and understandable [3].There are two different approaches to explain DL models: To transform the model into a self explainable model (intrinsic approach), which consists of the reduction of the complexity of the model, or to apply explainability methods (post-hoc explainability), consisting of the application of specific methods that extract explanations from complex models [2,4,5]. Furthermore, explainability methods can be categorised as model agnostic or model specific and as local or global [6]. Model agnostic methods are post-hoc and they can be applied to any type of model [5]. Such methods have no access to the model’s internal components, e.g., weights or structural information [5]. On the other hand, global interpretation tools focus on the overall understanding of the DL model features and each of the learned components, while local interpretation tools checking individual predictions of the model. Local methods are less complex to implement compared to global interpretation tools, which are normally applied to simpler DL models [6].The transparency and explainability of DL models is key for medical applications, as discussed in [7]. The same work enumerates numerous flaws that black-box models present in this specific domain, namely in the ethical and disputability sense from the patients perspective. As mentioned in chapter 3 of the General Data Protection and Regulation (GDPR) document, patients should have the right to know the origin of diagnostics, recommended therapeutics, or any other medical intervention that may be supported by artificial intelligence models, such as DL models [8,9]. Thus, the creation of tools capable of explaining DL models is essential for their application in the clinical context.In our study we performed a classification task to detect arrhythmia events in electrocardiographic images using a convolution neural network (CNN). Then we applied three different methods to explain the classification. Those methods created three different attribution maps: Saliency maps, gradient-weighted class activation (grad-CAM) maps, and guided backpropagation (GB) grad-CAM maps. Attribution maps highlight the pixels of an input image that contribute the most to the classification [2,6]. The methods that underlie these maps are local and post-hoc. A detailed description of the three methods will be presented in Section 3.Machine learning and, specifically, deep learning techniques have been applied for ECG signals classification in the case of arrhythmia detection [10].Traditional machine learning algorithms, despite being transparent regarding the classification process, can also depend on tedious and costly tasks, such as feature engineering. For example, in [11], the authors used optimum-path forest and demonstrated its dependence on the feature representation that was given as input. In the same work, the authors presented results using other commonly used classifiers, namely, support vector machines (SVM), multi-layered perceptron (MLP), and Bayesian expert system classifiers, while maintaining the same sets of features. It was concluded that all classifiers depended on the input features for the classification outcomes.In [12], the authors applied an intermediate approach, in which they used the radial basis function (RBF) network to model the ECG heartbeats from the different classes and then, using a deterministic learning algorithm, performed classification with an accuracy score of around 98%. Thus, the application of the RBF allowed to automatically extract dynamic features from every heartbeat.Notwithstanding the issues around DL models in practice, they have been applied in academic works on ECG signals for the detection and classification of arrhythmia events [13,14,15,16]. To overcome the feature engineering process that is required for traditional machine learning algorithms, in [13] the authors applied a combination of a 1-D CNN for feature extraction and three fully connected network (FCN) layers for classification, achieving an accuracy of 86%. In [17], the authors used a denoising autoencoder (DAE) for unsupervised features extraction and stacked its hidden representation layers with a regression layer, which is capable of assigning scores based on the examples in the training set. The innovation is the usage of active learning, in which experts can provide input to the most informative heartbeats given certain criteria, resulting in improved results compared to other works.Most of the studies reported in [13,14,15,16] apply CNN models to the 1D ECG signal either for feature extraction or to prior extracted features to detect or classify arrhythmia in single or multi-lead ECG signals. However, it is not common to use the ECG signal as input images to a CNN, such as performed in our work. Nonetheless, in [18] the authors converted each heartbeat of each ECG signal into a 64 × 64 greyscale image. Those images were used as input of a custom CNN model with to classify 5 different arrhythmia types. The overall accuracy was 99.7% with a F1-score of 99.24%. The authors of [19] performed feature extraction of 32 × 32 binary ECG images (128 ECG images were created from each patient) using 3 different pre-trained CNN models. Those features were merged and used as input of several machine learning models to perform binary classification. The best model achieved an accuracy of 97.6% and a F1-score of 97.9%. Finally, in [20] a multi-label classification pipeline was created by stacking a CNN and a LSTM model, which used ECG images with 10 s of signal as input. The overall accuracy and F1-score was of 99.33% and 96.06%, respectively. Despite the high performance reported in these studies, there was no mention about patient division across the train, validation, and test sets.Works addressing the problem of interpretability in an ECG classification problem are scarce. The work in [21] used a hierarchical attention network combined with bidirectional recurrent neural networks (BiRNN) for the classification task. Explainability was introduced by the hierarchical structure and no specific aforementioned algorithm was used. However, the authors were able to explain the decision process based on the specific results for each hierarchy, which corresponded to windows (set of heartbeats), heartbeat, and, finally, waves (P, QRS complex, or T).The authors in [22] developed an explainability framework specifically addressed to the problem of ECG classification, which included three modules that evaluated the features extracted from a 1-D CNN used for the classification of ECG data. In this case, the authors used segments that contained 5 heartbeats and signal synthesis methods as data augmentation for improved results. The internal states of the CNN were not taken into account for the proposed explainability method, which relied only on the features extracted from the input signals.In [23], the authors reported the creation of an explainable deep learning model (XDM) that performs multi-label classification. The XDM receives a 12-lead ECG time series signal (Sejong ECG datase) with 8 s of data. The model consists in 6 deep learning modules. Each module analyses the presence of a specific feature in the signal. This model requires an increased labelling effort because each cardiologist needs to label each ECG signal not only for the arrhythmia type but also for 6 signal characteristics. By doing this, the authors assured that each arrhythmia label is associated with a specific set of characteristics. Additionally, an attribution map was created to understand which time intervals of the ECG signals had significant impact on the model’s decision for each feature.In [24], a DL model (xECGNet) was created to perform multi-label classification and to provide a visual explanation using a fine-tuned attention map. The used dataset (CPSC 2018) has 12 lead ECG time series signals. With this dataset, the authors created samples of a fixed length with all leads as input of an Attention Branch Network (ABN) that uses attention maps to improve the model’s classification performance and to provide the explanation. This is possible due to the addition of a regularisation term between the attention map and the reference map to the objective function. The reference map is created using the average of the attribution maps of all the ground truth labels. The final attention map, after the model training, will have information about the most important time intervals of the samples to the classification result for each lead.We propose to expand the knowledge of this specific problem by using computer vision allied with specific explainability algorithms to increase interpretability of the results of DL models in ECG images containing six heartbeats. Contrary to other works, we chose to use ECG images to simulate the work of cardiologists when analysing ECG in real life. We trained two different models to predict the label of specific heartbeats. One model classifies the first heartbeat while the other classifies the last heartbeat. We aim to verify if the model focus is on the labelled heartbeat. To the best of our knowledge, this is the first work where explainability methods are applied to a computer vision task using ECG images as an input. By doing this, we aim to know exactly where the model focuses when performing arrhythmia detection, in opposition to the aforementioned works that only identify the most significant time intervals. Additionally, we also performed focus quantitative analysis, in which we computed the proportion of attribution between the area with heartbeat of interest and the rest of the image.Given this, our objectives were to (1) develop two different datasets consisting of ECG figures containing various heartbeats and labelled given a specific heartbeat; (2) train a CNN model for each dataset to detect arrhythmia in ECG figures; (3) apply the aforementioned explainability methods to visually interpret the classification of the CNN with the aim of understanding if the models are able to focus on the heartbeat of interest—the heartbeat that provides the label to each figure; and (4) develop a metric to quantify the amount of focus of the models to the heartbeat of interest. Moreover, with this metric we investigated the effect that different labels have on the focus of the classification models, if accurate classifications are reflected on the focus of the model, and if the nature of the heartbeat, i.e., if it is normal or arrhythmic, is reflected on the focused region.We used the MIT BIH arrhythmia database for the exploration of explainability algorithms applied to a classification model of ECG images. This dataset is comprised of 48 half-hour ECG recordings of 47 different subjects, where 23 were chosen randomly from a set of 4000 recordings, 25 of which were chosen to include unusual arrhythmia events [25]. The sampling frequency of the recording is 360 Hz, with 11-bit resolution over a range of 10 mV. Each heartbeat and arrhythmia event was labelled by two different cardiologists. The total number of labels and, thus, of heartbeats is approximately 110,000, including noisy and virtually intractable parts of signals, that were discarded in this work. Each record is composed of 2 leads of the ECG, however we only considered one for each record. We used the modified limb lead II (MLII) in all cases except two, which did not include that lead. In those cases, we used the V5 lead.Since the data corresponds to 1-D signals, the first step for the application of computer vision techniques is to convert it to images. To mimic real-world applications, these images will comprise sets of 6 heartbeats and the samples will be constructed based on a sliding-window approach. Using those same images, 2 different datasets were created, generating 2 different models with different training data. Those datasets can be described as follows:1.Dataset 1—the binary label of each image corresponds to the label of the last heartbeat;2.Dataset 2—the binary label of each image corresponds to the label of the first heartbeat.Dataset 1—the binary label of each image corresponds to the label of the last heartbeat;Dataset 2—the binary label of each image corresponds to the label of the first heartbeat.Although there are 15 types of arrhythmia, we considered only the normal label and the remaining are comprised in the abnormal label.Some transformations were applied to the created sample images to feed our DL models. The images were cropped to minimise the amount of image without relevant information. They were also resized (224 × 224) and normalised. Finally, following the recommendations in [10], the dataset was divided in two. Each subset had a different group of patients. Then, we created a validation set from the train set. By doing this we obtained: 37,867, 11,204, and 49,617 images, in the training, validation, and test sets, respectively.We used a ResNet50 to perform our binary classification task. This model was created to surpass the difficulty of training very large neural networks [26]. ResNet creators introduced residual blocks that ease the training process. To import, train, and evaluate the model and to perform all data transformations we used Pytorch (https://pytorch.org/ (accessed on 7 January 2022)). The developed code is available in GitHub (https://github.com/ruivarandas/XAI_ECG (accessed on 7 January 2022)).The imported model was already trained for a natural image classification problem, on the ImageNet dataset [27]. However, due to the significant difference between the pretraining task and actual task of this project, the model was trained from scratch with a small learning rate. This was only possible due to the large number of samples in the datasets created for this project. In order to adapt the imported model to the task at hand, the last fully connected layer of the model was replaced by another fully connected layer but with an output size equal to the number of classes, i.e., output size of two.We used the Adam algorithm with weight decay to optimise our models. Additionally, we applied learning rate decay with a decay step of 4 epochs and gamma of 0.1. The initial learning rate was defined as 1 × 10−5 and 1 × 10−4, for datasets 1 and 2, respectively. Reducing the learning rate of the model as the training progresses allows the model to become more stable in advanced epochs.Weighted cross entropy algorithm was used as the loss algorithm. The usage of a weighted loss was important due to the class imbalance of the dataset (very common in any medical dataset). For that reason, a higher weight was given to the less represented class in the loss computation.Accuracy, precision, and F1 score were used as the classification evaluation metrics. We developed an early stop mechanism using the evolution of the F1 score metric in the validation dataset. If the F1 score of a certain epoch is less or equal than the mean of the same parameter of the last 4 epochs, then the training stops. This mechanism reduces the training time without compromising the performance of the model. Equations (1)–(3) show the mathematical formulas of the evaluation metrics.
+        (1)Accuracy(%)=NumbercorrectpredictionsTotalnumberofpredictions×100
+        (2)Precision(%)=TPTP+FP×100
+        (3)F1score(%)=TPTP+12TP(FP+FN)×100
+        wherein TP, FP, and FN are the true positives, false positives, and false negatives, respectively. Positive samples correspond to the abnormal label (arrhythmia) and negative samples correspond to the normal label (healthy).Besides these, the receiver operating characteristic (ROC) curve and the confusion matrix were used to assess the quality of the classification. The ROC curve plots the true positive rate over the false positive rate and shows how well a model can distinguish between two classes. The confusion matrix is illustrated in Figure 1, and illustrates the number of TP, FP, TN, and true negative (TN) graphically.Each model was trained in a computer with 2 NVIDIA Geforce RTX 2080 8GB GPUs with 64 GB of RAM and a i7-9700K 3.6 GHz CPU.We used three different explainability methods: gradients (generates saliency maps), grad-CAM, and guided backpropagation grad-CAM. Besides being local and post hoc methods, they are also backpropagation-based methods. According to [28], backpropagation-based methods compute attribution for all input features with a single forward and backward pass through the network. Some methods inside this category can only provide positive contributions to the final prediction result in the attribution map, while others show the positive and negative contributions, which may degrade the results by increasing the noise in the map.The gradients method originates saliency maps and is the earliest and probably one of the most used methods to explain the predictions of CNNs. The saliency map of the input of a CNN highlights the parts of the input that most contributes to the outcome and, so, the method attributes importance to each pixel of an input image regarding the prediction of the network.Based on the work that introduced this method, the pixel importance is obtained by applying the somewhat inverse operation relative to the training of neural networks (NN) [29]. Neural networks are usually trained by the application of backpropagation regarding the expected labels to optimise the loss function. The backpropagation method is applied from the input to the output of the networks. However, to obtain the saliency maps, the same backpropagation algorithm is applied, but in this case the derivative is applied regarding the input image (Equation (4)):(4)w=∂yc∂I|I0
+          where yc is the class score, I is the image, and I0 is the input image, specific for the task at hand, and w is the attribution map—analogous to the weights of NN.Gradient-weighted class activation mapping (grad-CAM) was first introduced in [30] as a generalisation of class activation mapping (CAM). Unlike CAM, grad-CAM can be used to visualise any type of CNNs. Grad-CAM uses gradient information that flows to the last convolution layer to compute the importance, for the prediction, of each neuron. The last convolutional layers of a CNN retain spatial information and its neurons are focused on semantic class-specific information in the input image. For this reason, grad-CAM is a class discriminative method.Equation (5) shows how to compute the weight αkc. This weight captures the importance of a feature map k for a target class c. This value is the global average pooling of the gradient of the score (before softmax) for class, c, w.r.t the feature maps, Ak: ∂yc∂Aijk:(5)αkc=1Z∑i∑j∂yc∂Aijk.After this step, a weighted combination of forward activation maps and a ReLU are necessary to compute the final map, as shown in Equation (6). The application of the ReLU guarantees that the attribution map only depicts the positive contributions to the classification result:(6)LGrad−CAM c=ReLU ∑kαkcAk.These operations create a coarse heat map of the same size as the convolutional feature maps. Although we can apply grad-CAM to any convolutional layer, as we are trying to explain the decisions of our classifier, we applied the method to the last convolutional layer of our ResNet.The guided backpropagation grad-CAM (GB grad-CAM) was also introduced in [30]. This method was developed to tackle the lack of finer details in the attribution maps created using grad-CAM. GB grad-CAM consists of a pixel-wise multiplication between a grad-CAM map and guided backpropagation (GB) map.GB maps were first described in [31]. These maps are an improved version of the saliency maps. Instead of using a normal backpropagation approach, they use a guided backpropagation. GB combines two methods: ‘Deconvnet’ [32] and backpropagation. These methods differ only in the way they handle backpropagation through the ReLU nonlinearity. The ‘deconvnet’ method, considers only the top gradient signal to compute the gradient in the nonlinearity and ignores the bottom input. GB combines this with backpropagation and masks out the values for which the top gradient or bottom data are negative. This prevents the backward flow of negative gradients.We computed a proportion of attribution between the heartbeat of interest and the rest of the image to study the magnitude of focus in the different regions of the ECG images. Firstly, we computed a rectangular region of interest (ROI) that contains only the heartbeat of interest. This heartbeat is the labelled heartbeat that varies according to the dataset. Figure 2 illustrates an example of a computed ROI. In that case, the last beat was the heartbeat of interest. Then, we determine the proportion between the total sum of the pixel attribution map and the sum of the map inside the region of interest. Using this proportion we are able to measure the percentage of focus inside the ROI.This work can be divided in two distinct parts: One regarding the classification of ECG images and the other focused on the analysis of the created attribution maps using our custom metric.Table 1 presents the number of epochs and training time, the values of accuracy and F1-score computed in the validation and test set, and the precision score computed only at the testing stage. Figure 3 and Figure 4 are related with the testing stage of our models. Figure 3 presents the confusion matrix of each model. Figure 4 presents the receiver operating characteristic (ROC) curve and the area under the curve (AUC) of the ROC. In a brief analysis, we can conclude that model 1 (label corresponding to the last heartbeat of the figure) performs better in terms of global accuracy than model 2 (label corresponding to the first heartbeat of the ECG figure) in the testing set, with an accuracy of 93.66%. However both the confusion matrices and the AUC values are very similar for each model Table 1.Regarding the explored explainability metric, we present three different scenarios.The generic scenario, presented in Table 2 and the histograms shown in Figure 5, is the overall comparison of the three explainability methods relative to the classification of the two models. In this case, we see that the method that better focuses the heartbeat of interest in both classification models is the gradients method, with a focus of around 30% in model 1 and 25% in model 2, while the Grad-CAM is the method with the worst focus with a focus of around 23% in model 1 and 9% in model 2. The comparison between the two models revealed a value p < 0.001 for the null hypothesis that the mean values for each explainability method are equal. A t-test for the means of two independent samples distribution was used.The comparison between correct classification and incorrect classification, presented in Table 3, shows that in the case of the Grad-CAM and GB grad-CAM in model 1, there is a positive relation between correct classification and the focus of the attribution maps. Notwithstanding, in model 2 there is not a clear relation between the classification results and the focus of the attribution maps. The comparison between the two models revealed a value p < 0.001 for the null hypothesis and that the mean values for each correct vs. incorrect comparison of each explainability method are equal. Moreover, the value p for the comparison between sets is lower than 0.001 between all cases, aside from the wrong vs. wrong case using the saliency maps and the GB grad-CAM maps. A t-test for the means of two independent samples distribution was used.Finally, the comparison between the two labels of the images, presented in Table 4 and exemplified in Figure 6, shows that there is not a consistent relation between the label of the images and the focus of the attribution maps. For example, the Grad-CAM shows better results in model 1 for the normal label, while for model 2 the opposite happens. The comparison between the two models revealed a value p < 0.001 for the null hypothesis that the mean values for each normal vs. arrhythmia comparison of each explainability method are equal. Moreover, the value p for the comparison between sets is lower than 0.001 between all cases. A t-test for the means of two independent samples distribution was used.The first step in our project was to detect arrhythmia in the last or first heartbeat within ECG images containing six heartbeats. Thus, we created two different models: One to classify the last heartbeat (model 1) and another to classify the first heartbeat (model 2). Table 1 shows that model 1 performs slightly better than model 2 at the testing set, consisting of patients that are not in the other sets (interpatient classification). We hypothesise that, similarly to how humans classify ECG, the beats prior to the heartbeat of interest help the model to produce the correct prediction.During training, the low number of epochs and the difference of the number of epochs between the two models is also noticeable. This difference is related to how the stopping criteria of the early stop mechanism was implemented. Training stopped after 4 epochs of decreasing F1-score and, thus, if the criteria was different, the number of epochs might have been higher or lower. However, the fact that the training epochs of each model is different does not influence their results, since the stopping criteria was the same. Thus, even if both models continued training, the F1-score would not increase.The confusion matrices show that our classification models struggle to correctly identify the positive label as reflected by the low precision scores (74.10 % and 63.57%) and proximity between the values of false and true positives. However, the AUC-ROC value shows a 85% probability of distinguishing between the positive and negative label. Our classification accuracy scores of 91–94% were slightly below the state of the art results presented in [13,14,15,16]. When comparing with studies [18,19,20] that used ECG as an image we also concluded that our classification results were inferior. However we cannot perform a direct comparison because those studies were not clear regarding if they performed patient division across the data sets. The inferior results can be explained by multiple factors. We use the ECG signal of only one lead of the MIT BIH database without any pre-processing. This can be problematic because some ECG signals in that database have a low signal-to-noise ratio, even after removing the intractable parts of signals. In addition, some arrhythmia events are not visible in all leads. Finally, we used the dataset division proposed in [10], but samples from the training set were used to build the validation set, resulting in a testing set that was larger than the training set. By doing this, we are certain to use the same testing set as most of the reported studies. Since our main objective was the transparency of DL models, we did not focus on the optimisation of the classification results.After the classification task, the next step was the application of the already enumerated explainability methods to create attribution maps. From those maps we computed our metric to measure the amount of focus of the model on the heartbeat of interest. The maps created using the gradients method do not distinguish between positive or negative contributions to the model prediction [28]. Attribution maps created using grad-CAM and GB grad-CAM methods only consider positive contributions. For this reason, the values of our metric are generally higher for attributions maps created using the gradients method. Nevertheless, we expect the model to focus on the region of the labelled heartbeat, even if certain pixels of that region give negative contributions to the prediction.Supposing that the model is not focused in any particular area of the input images, the average value of our metric would be 9.7±4.0%. We called this value the random focus value. This value was estimated by computing the average proportion between the area of the region of interest and the area of the image across all test samples. The values shown in Table 2, Table 3 and Table 4 are higher than the random focus value for all attribution maps except for specific cases in the grad-CAM attribution maps. This fact implies that both our models considered the region of the labelled heartbeat an important region to the prediction process. For further analysis we created three different scenarios for comparison: The generic case; the correct vs. incorrect classification case; and the abnormal vs. normal label case.From the generic case, we can conclude that the obtained values of the attribution metric corroborate the better classification performance of model 1. Model 1 has higher values for all maps. From Table 2, we can also highlight the fact that the attribution metric using grad-CAM in model 2 was the only one below the random focus value. In fact, generally, grad-CAM had the lowest values for all cases. Grad-CAM maps are the most coarse attribution maps of the three different maps that were created [33]. Therefore the higher pixel values are more scattered across the map, when comparing with the other maps, resulting in lower values of our custom attribution metric. The histograms presented in Figure 5 show that despite having the higher average values, model 1 also has higher deviation values for all maps, except for the ones created using gradients method—(a) and (d). The frequency of null values can also be seen as not negligible for the grad-CAM and GB grad-CAM methods—(b), (e) and (c), (f). Those null values represent the cases where there is no pixel attribution of positive contributions inside the region of interest. Gradients maps have a negligible frequency of null values because they consider positive and negative contributions to the prediction process.With the second scenario, Table 3, we extrapolate a relation between the value of our custom attribution metric and the coherence between the classification result and the actual label of the samples. We expected that correct classifications to be related with a more focused model (higher values of the metric) however this is not the case when applying the gradients method. This is again related to the nature of the method, which does not distinguish between positive and negative contributions.Finally, our third scenario, Table 4, focuses on the relation between the actual label of the test sample and its attribution map. Here we cannot find a general tendency for both models. For model 1, both GB grad-CAM and grad-CAM have higher values when classifying normal samples. On the contrary for model 2, higher values are obtained when classifying abnormal samples. Gradients maps have higher values for abnormal samples in both models. These tendencies can be seen in Figure 6.Notwithstanding that we are assessing the focus of our classification model in the region that contains the labelled heartbeat, we also hypothesise that high attribution values outside that region in our input images might be relevant for classification. There, the focus can be on the blank parts of the figures or in the remaining heartbeats that do not contribute to the label. In the first case, the model might be searching for heartbeats that should happen (e.g., if the signal is shorter than expected) or for higher amplitude signals than the ones that are present. In the second case, the model might be looking for inter-beat features, such as, the distance between R peaks, to classify the most important beat (important in some arrhythmia cases, such as tachycardia or bradycardia). In fact, the low values of our metric support this hypothesis, but further research is required for validation.In its present condition our work cannot be applied in clinical settings. First of all, the accuracy and precision scores are not high enough. The fact that there are numerous false negatives could lead to serious misdiagnosis. On the other hand, the explainability still needs improvement, as previously mentioned. Namely, the focused region should be contained on the sites of the abnormality and, in this case, the focus is still somewhat sparse across the figures. However, our explainability metric might allow future works to quantitatively assess the quality of their explainability methods. Ideally, all focus should be on the ROI and the ROI should be clearly identified.In this study, we built two computer vision classification models and applied three different backpropagation-based explainability methods to each to create attribution maps. From the attribution maps, we then created a custom metric that measures the degree of importance of each pixel of the input image given the classification result. Then, we compared the obtained values of our metric across different cases: The generic case, accurate vs. inaccurate classification, and abnormal vs. normal samples.Our classification results were below the state of the art results. Both models achieved testing accuracy scores between 91–94%. The values of our metric ranged between 8 and 38% with high standard deviation values. Those values show that the focus of the classification models is sparse across the ECG image, even though there is a concentration of focus in the heartbeat of interest.For future work, we will improve the classification results by pre-processing the ECG signals. Moreover, we will improve the computation of the heartbeat region of interest to be more precise. We will extend the knowledge on this subject by analysing the importance of other regions of the ECG images. We will explore specific waves inside of the heartbeat of interest to assess their importance to classification (e.g., P and T waves, and QRS complex). Furthermore, it would be interesting to explore if DL models capture pseudo-time dependencies by computing the distant between R peaks and other commonly extracted features, which might explain the importance of regions of interest besides the labelled heartbeat.Conceptualisation, R.V. and B.G.; methodology, R.V. and B.G.; software, R.V. and B.G.; validation, H.G. and P.V.; investigation, R.V. and B.G.; data curation, R.V. and B.G.; writing—original draft preparation, R.V. and B.G.; writing—review and editing, R.V., B.G., H.G. and P.V.; visualisation, R.V. and B.G.; supervision, H.G. and P.V. All authors have read and agreed to the published version of the manuscript.This work was funded by FCT—Portuguese Foundation for Science and Technology and Bee2Fire under the PhD grant with reference PD/BDE/150624/2020. Moreover, it was funded by FCT—Portuguese Foundation for Science and Technology and PLUX Wireless Biosignals S.A. under the PhD grant with reference PD/BDE/150304/2019.Not applicable.Not applicable.The raw data used in this work is publicly available at https://physionet.org/content/mitdb/1.0.0/ (accessed on 7 January 2022). The transformed data can be found in [34].We would like to acknowledge André Martins from the Instituto Superior Técnico, Universidade de Lisboa, for his input in the early stages of this work.The authors declare no conflict of interest.Confusion matrix representation.Normal electrocardiographic signal with a rectangular region of interest (ROI) that contains the last heartbeat.Confusion matrices of the models at testing stage. Arrhythmia—positive label. Normal—negative label. (a) Model 1. (b) Model 2.Receiving Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values of the models at testing stage. (a) Model 1. (b) Model 2.Histograms of the distribution of the custom attribution metric values for the generic case. (a–c) correspond to the metrics estimated using model 1 and the methods: Gradients, grad-CAM, and GB grad-CAM, respectively. (d–f) correspond to the metrics estimated using model 2 and the methods: Gradients, grad-CAM, and GB grad-CAM, respectively.Examples from the obtained attribution maps. The first row of figures corresponds to maps created using model 1 (the heartbeat of interest is the last) and to an abnormal sample: (a) Saliency map; (b) grad-CAM map; and (c) GB grad-CAM map. The second row corresponds to maps created using model 1 and to a normal sample: (d) Saliency map; (e) grad-CAM map; and (f) GB grad-CAM map. The third row corresponds to maps created using model 2 (the heartbeat of interest is the first) and to an abnormal sample: (g) Saliency map; (h) grad-CAM map; and (i) GB grad-CAM map. The fourth row corresponds to maps created using model 2 and a normal sample: (j) Saliency map; (k) grad-CAM map; and (l) GB grad-CAM map. In all the presented cases, the models gave the correct predictions. The scale presented at the top implies that the brighter pixels correspond to a higher focus.Classification metrics. The models 1, 2, were trained with dataset 1, 2, respectively. The presented values are percentages, except for the train time and number of epochs.Attribution metric—generic. Mean value ± standard deviation value of the metric of all test samples. Each line for a different dataset: 1—last heartbeat labelled; 2—first heartbeat labelled. All presented values are percentages.Attribution metric—correct classification vs. incorrect classification. Mean value ± standard deviation value of the metric of the test samples according to their accuracy of classification. Each line for a different dataset: 1—label corresponding to the last heartbeat; 2—label corresponding to the first heartbeat. All presented values are percentages.Attribution metric—abnormal label vs. normal label. Mean value ± standard deviation value of the metric of the test samples that are abnormal and those which are normal. Each line for a different dataset: 1—last heartbeat labelled; 2—first heartbeat labelled. All presented values are percentages.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00009.txt

@@ -0,0 +1,2 @@
+These authors contributed equally to this work.This study aims to reflect on a list of libraries providing decision support to AI models. The goal is to assist in finding suitable libraries that support visual explainability and interpretability of the output of their AI model. Especially in sensitive application areas, such as medicine, this is crucial for understanding the decision-making process and for a safe application. Therefore, we use a glioma classification model’s reasoning as an underlying case. We present a comparison of 11 identified Python libraries that provide an addition to the better known SHAP and LIME libraries for visualizing explainability. The libraries are selected based on certain attributes, such as being implemented in Python, supporting visual analysis, thorough documentation, and active maintenance. We showcase and compare four libraries for global interpretations (ELI5, Dalex, InterpretML, and SHAP) and three libraries for local interpretations (Lime, Dalex, and InterpretML). As use case, we process a combination of openly available data sets on glioma for the task of studying feature importance when classifying the grade II, III, and IV brain tumor subtypes glioblastoma multiforme (GBM), anaplastic astrocytoma (AASTR), and oligodendroglioma (ODG), out of 1276 samples and 252 attributes. The exemplified model confirms known variations and studying local explainability contributes to revealing less known variations as putative biomarkers. The full comparison spreadsheet and implementation examples can be found in the appendix.In recent years, extensive benefits to different application areas have been offered due to successfully applying machine learning (ML) algorithms. In particular, the success of deep learning (DL) approaches are transforming the way we approach real-world tasks performed by humans. ML and DL establish artificial intelligence (AI) models which can be applied in many different fields of research such as healthcare [1], cancer classification [2,3,4], autonomous robots and vehicles [5], image processing [6], manufacturing, and many more [7,8,9,10], thus enhancing and providing various benefits in the corresponding fields. Moreover, these models resulting from ML are suitable for performing different tasks, such as recommendation, ranking, forecasting, classification, or clustering. The variety and the nature of these approaches make them complex to understand and interpret. In the literature, AI models are generally known as black-box, particularly if they result from ML or DL [11]. The opaqueness of such models has negative effects on user acceptance [12]. It also limits the application in sensitive cases such as medicine, finance, or law, where explanations are crucial for users to understand and interpret results in order to effectively manage and use the underlying algorithms [11,13]. From the legal perspective, most applications of AI in medicine are defined as high-risk use cases of AI according to the legal framework for regulating the use of AI proposed by the European Commission (European Commission, 2021). In case of a high-risk application it is required to provide transparency and clear and comprehensible information about the system and its decisions to the user. Such explanations are also dictated by the European General Data Protection Regulation, but also by Californian law (Title 1.81.8. Automated Decision Systems Accountability Act of 2020). Traditionally, software validation or IT auditing is applied in order to fulfill the legal and, in many cases, compliance requirements. However, due to the black-box characteristic resulting from ML and DL, traditional approaches are no longer sufficient, and new guidelines and approaches are needed [14]. In this regard, Explainable Artificial Intelligence (xAI) is proposed as a technical solution, and the first successful validations are already performed in sensitive areas such as pharmaceutical production [15]. In addition, xAI can also increase user acceptance, and the application rate of these models [12]. Thus, xAI approaches seem promising to handle this challenge from a technical perspective.xAI approaches aim to extract knowledge of what the AI algorithm learned during training and how the decision for particular or new instances are generated during the prediction process. xAI mainly focuses on two methods to provide an explanation at a different level of detail: local and global explainability. Local explainability aims to explain particular prediction output, e.g., prediction of single instances. We find many different techniques focusing on local explainability in the xAI literature [16,17]. On the other hand, the goal of global explainability is to explain the overall model behavior, rather than a particular instance. Global methods are extensively applied in different domains, such as health care [18,19], manufacturing [20], administration of justice [21], or biomedical science [22]. These methods mainly rely on dimension reduction and visualization techniques to provide an intuitive explanation to humans. Visualizing a process helps us understand ML models and decision-making processes in a more intuitive way [23]. Moreover, visual inspection is considered as an easy and fast way to recognize new knowledge while analyzing complex processes [24]. As a result, visualization in the context of xAI is widely applied, thus facilitating the interpretation process of black-box models [11,25,26]. Users benefit from visual analytic (VA) systems for xAI [27]. Many of these methods are implemented in Python or R and are openly available [17,28,29]. This helps researchers and, in general, the data-driven community to use and enhance further state-of-the-art solutions. Some existing methods have already been summarized [30,31,32]. However, a comparison of ease of use regarding implementation, as well as details on visualization features, is missing.In this paper, we report on a structured review to investigate the state of the art of mature xAI libraries incorporating VA features. We analyzed the characteristics of xAI libraries with respect to ease of installation and documentation. The comparison is use-case driven: we compare and rank selected libraries regarding their VA capabilities for global and local explainability in general. In particular, we explore different implementations of lime and SHAP approaches and apply selected libraries for the use case of investigating glioma classification based on several clinical and genetic variables. We thereby showcase the applicability of xAI on and supporting the biomedical knowledge creation process.Classification of glioma subtypes is important for therapy decisions and is based on gene variations [33]. This list of central nervous system tumors has been introduced by the World Health Organization and has been updated recently [34]. The community-driven cancer classification platform Oncotree has been developed as clinical decision support system for oncology research and precision medicine and allows for dynamic granularity [35]. For example, grading of diffuse gliomas (DIFG) is still an ongoing discussion and momentarily defined by tumor nomenclature [36]. The process involves molecular and histological features in order to revise risk stratification. Common molecular biomarkers used for clinical classification of glioma include α-thalassemia/mental retardation syndrome X-linked (ATRX), isocitrate dehydrogenase 1 (IDH1), tumor protein p53 (TP53), telomerase reverse transcriptase (TERT), and phosphatase and tensin homolog (PTEN) or the epidermal growth factor receptor (EGFR) among others [34,37]. We have recently highlighted age-based differences in brain tumor diseases using an explainable classification approach [22]. We now extend our studies to include several xAI methods for classifying DIFGs.xAI is defined for the first time in 2004 by Can Lent et al. [38] as a research field that explains the behavior AI models in a more understandable way. However, focus on the topic of xAI has been recently increasing [32] due to increased attention and improvements around the topic of AI/ML across different fields. However, along with the high accuracy results, a more human-centric explanation of the decision-making process of these models is required. This leads the focus toward xAI in the current age. Furthermore, the increase in complexity of ML models has lead to the requirement for developing algorithmic decision-making such as fairness, accountability, and transparency (FAT) principles [39] which are especially evident in highly regulated and mission-critical scenarios.There are several perspectives on the explainability of an AI model (e.g., scope, stage, problem type, etc.). The scope perspective regards the global and local view on model explanations. AI models can be explained either at the global level or local level. Global level interpretation is known as global interpretability in the literature [32], where the entire model behavior is analyzed e.g., feature importance. Global level interpretability summarizes the impact of input features on the model, as well as the model as a whole, while the local interpretation is defined as local interpretability, and it aims to understand the behavior of single predictions and decisions made by the model.Another perspective on the explainability of an AI model is associated with the type of AI model itself. Overall, two types of models exist, white-box and black-box models. White-box models are made to be explainable by design, resulting in no requirement of additional xAI methods for the model to be explainable. Contrarily, black-box models are not explainable by design, so other techniques have to be applied to extract reasoning for certain decisions and predictions.In regard to xAI methods, a recent study [32] reviewed more than 200 scientific articles that aimed to develop new methods for explainability. However, discussing these methods and other xAI concepts falls outside of the scope of this paper. We encourage the reader to consult the work discussed in [30,31,32] for more details about these concepts.Data on glioma samples were downloaded from cbioportal [40,41] with filtering the 6 studies gbm_mayo_pdx_sarkaria_2019, gbm_tcga_pub2013, glioma_mskcc_2019, lgg_tcga, lgg_ucsf_2014, and odg_msk_2017. Only data with the 7 attributes “Oncotree Code”, “Mutation Count”, “Overall Survival (Months)”, “Overall Survival Status”, “Sex”, “Somatic Status”, and “Diagnosis Age” were used. Sample rows without complete data have been removed. Data were extended with gene mutation data of the top 246 mutated genes within selected studies.The top three diffuse glioma (DIFG) subtypes (Glioblastoma multiforme (GBM), Anaplastic Astrocytoma (AASTR), and Oligodendroglioma (ODG)) were further selected and analyzed within this work. We filtered and further processed data for model building comprising of 1276 sample rows with 253 columns out of the 5 studies gbm_mayo_pdx_sarkaria_2019, gbm_tcga_pub2013, glioma_mskcc_2019, lgg_tcga, and lgg_ucsf_2014. The Oncotree Code was selected as the target and the other 252 data columns were selected as features, with 872 GBM sample rows, 234 AASTR sample rows, and 170 ODG sample rows. The data preprocessing and model building can be found on https://github.com/mathabaws/SOTA_xAI_Visual_analytics/blob/main/notebooks/diffuseglioma-dataset-processing.ipynb (accessed on 12 January 2022).We conducted a structured review with the goal of investigating current developments and the state of the art xAI libraries focusing on model interpretation and visualization techniques. State of the art means most up to date, publicly available, implemented consistently with the requirement of current software technology, and following common Python patterns. Moreover, this review aims to investigate various relevant aspects of xAI libraries such as maturity level, documentation, supported programming languages, models and different machine learning tasks, support for data types, etc. The structured review closely follows the methodology for Structured Literature Review (SLR) from Webster and Watson [42]. Additionally, we take necessary attributes for a software selection process into account.The initial set of available libraries was acquired through a search in GitHub. Keywords and the type of the results are the two key limiting factors to guide the initial set of results. For the first limiting factor, the keywords “explainable AI” and “interpretability” were used. The second limiting factor was the type of results and this was set to “repository” which excluded all the results with these keywords in, e.g., the code itself or discussions, issues, commits, etc. Applying these limiting factors resulted in 57 results. To further narrow down the results, three rules were developed for the initial scan of the libraries as shown below:1.Result has to be a repository of a Python library or a software package;2.Result has to implement at least one xAI method;3.Result has to be an overview repository (repository that provides an overview of xAI libaries).Result has to be a repository of a Python library or a software package;Result has to implement at least one xAI method;Result has to be an overview repository (repository that provides an overview of xAI libaries).Supplementary source code together with the overview of library versions and descriptions to recreate an exact development environment used for these experiments can be found on GitHub at the following URL: https://github.com/mathabaws/SOTA_xAI_Visual_analytics (accessed on 12 January 2022).By using the processed data from the combined studies described in the materials section, we trained a model to classify cancer subtypes by distinguishing between the Oncotree codes GBM, AASTR, and ODG. These are the top three most frequent diffuse glioma subtypes samples.In general, 1020 training instances were used for training, and 256 for testing. Testing data remained unbalanced representing a realistic scenario. Ten-fold cross-validation scored a mean accuracy of 0.87 with a standard deviation of 0.02. The results of the trained model are shown in Table 1.In the next subsections, the Python libraries suitable for xAI and VA selected for in-depth analysis are presented, including results from tests with the above described model.Applying the method described in the previous section, 52 relevant repositories were identified. Moreover, several overview repositories in the topic of xAI have been identified. These overview repositories provided information on the libraries other than ones identified through initial scan and were further used for backward and forward search. Next, a process resembling abstract and conclusion scan was conducted to filter out the libraries not focused on xAI and/or VA. In other words, documentation from repositories and implementation of the libraries were scrutinized to identify their focus and scope. As a result, 48 libraries were selected as relevant. These libraries were analyzed, interpreted, and summarized in a concept-centric way [42]. Through an in-depth analysis, metadata was collected, and core libraries and frameworks were identified for further exploration. Figure 1 provides an overview of the process.As a first step, we drill down initial results described in the previous section to the most important libraries aiming for xAI using visualization tools. The complete comparison table can be found in Appendix A.1. We then defined structured rules that help us to identify relevant libraries, which will be further analyzed and experimented. Firstly, we select only those libraries that are implemented in Python and integrate visualization features to communicate xAI results. Furthermore, chosen libraries are able to explain classification models. Last but not least, these libraries are open source, provide good documentation, and support tabular data.After filtering, we identified 11 relevant libraries. Selected libraries based on the aforementioned rules are listed in Table 2. We excluded 6 of the 11 identified libraries as missing criteria were revealed during the in-depth inspection. The remaining relevant libraries were grouped into three different groups: libraries aiming for global explainability in general, libraries focusing on local explanation, and, in particular, libraries which support Lime and SHAP approaches. In the first group, the following libraries are selected: ELI5 [43], Dalex [29], InterpretML [28], and SHAP [17]. In the second group, i.e., local explainability, Lime and SHAP approaches are explored in more detail. Three different libraries focusing on Lime are analyzed: Lime [16], Dalex [29], and InterpretML. Finally, three different libraries focusing on SHAP approaches are analyzed in detail: InterpretML, Dalex, and SHAP. The selected libraries are analyzed and compared within the groups and the results are shown in the sections below. The complete overview table can be found on the GitHub repository (Appendix A.2). All experiments concerning the analyzed libraries in depth are conducted using a notebook with the following characteristics: Lenovo ThinkPad L470, Intel(R) Core(TM) 2.70GHz - 2.90GHz, 16 GB RAM, Windows 10.Several libraries were identified with implementation of different feature importance methods. These are methods that rely on assigning a score to input features based on the predictive performance they add to the model. We are starting this overview with the focus on (1) methods for global explainability of the model and (2) methods that use visualization to communicate the explainability results. During the in-depth analysis, four libraries were identified to contain feature importance visualizations, namely ELI5, Dalex, InterpretML, and SHAP.ELI5 focuses on feature selection with the implementation of permutation importance. It enables extraction and visualization of feature weights and their contribution from the model as a form of global explanations. Visualizations are based on the list view of the features and their weights in a tabular form. The gradient of green and red color indicates the positive or negative impact on the model decisions, and there are no interactive options. Figure 2 depicts feature importance visualization implemented in the ELI5 library. Furthermore, model inspection on the prediction level is supported, which uses similar visualization with weights adding up to either probability of a class in classification models or predicted value in case of regression models.Dalex implements a method called variable importance which provides global explanations of a model based on Permutational Variable Importance [44]. Each variable is randomly shuffled in this method, and the model is inspected for its predictive performance. Intuitively, more important features impact the model performance more than the less important features. Finally, after 10 permutation rounds for each feature, visualization is created, showing the impact of each feature on the model. Such visualization provided by the Dalex library is depicted in Figure 3. Furthermore, the Dalex library provides a simple interactive overview during the mouse hovering over the visualization. This interactive window quantifies their influence on the model and provides additional information. The Dalex library also provides the option to tune the hyperparameters, such as a number of permutation rounds and various thresholds, and enables grouping of the features.The SHAP library provides the opportunity to analyze the model at the global level. This method helps to interpret the model by estimating feature importance altogether with feature effects on prediction with respect to raw data (as shown in Figure 4). The importance of features is shown along the x-axis, with important features listed at the top. For each feature, the contribution towards the specific classes is shown using the corresponding color, as shown in Figure 4a. Furthermore, SHAP provides the opportunity to conduct global interpretation for specific classes as shown in Figure 4b. In this case, the contribution of specific features is shown along x-Axis, where the contribution can be either positive (contributed toward prediction of this class) or negative. Each data point stacked vertically within this visualization represents the contribution for a specific instance. The color gradient encodes the raw values, blue representing the lowest and red the highest value.As mentioned in Section 3.1, InterpretML is focused on navigation through different views and interactive application of different methods. One of the methods that is provided by the library is the overall importance. Overall importance presents the global feature importance of the model. InterpretML makes the distinction of algorithms that are applied in two different model types. These are glassbox models and black-box explainers. To be able to apply and extract global feature importance, a glassbox model needs to be trained. These models are structured for direct interpretability, contrary to the black-box models that provide approximations of explanations. This introduces additional overhead in utilizing InterpretML for model explainability, as an additional model had to be trained to extract important features of the model. An example of such feature importance visualization provided by InterpretML is depicted in Figure 5. Based on the popular visualization library Plotly [45], InterpretML allows simple interaction with the visualization (e.g., zoom-in, selection, export to image format, etc.).Summarizing libraries for global explanation analysis, in terms of computational load, ELI5 provides the most lightweight solution for feature inspection. A simple and unified application programming interface enables a virtually instant overview of the features. On the contrary, all other remaining libraries require some degree of further processing to provide global explainability information. In the context of tabular data, the only supported visualization in ELI5 is a table overview with a gradient of green and red color encoding to indicate the importance of a feature in model predictions. The SHAP library provides more variety in terms of visualization with the implementation of bar chart and summary plot, which combines feature importance with feature effects. In regard to interactivity, visualizations provided by SHAP in the context of global importance are static and do not provide any further interactive features. Furthermore, in comparison to ELI5, SHAP requires an additional computational load that comes with the calculation of shap values. The Dalex library implements additional interactivity features in the model-level variable importance calculation. Visualization implemented in Dalex contains a list of features and their impact on predictions, with additional information provided upon the selection of a feature, which proved particularly useful when inspecting models with large numbers of features. However, this interactivity comes with additional computational load, which was significant in comparison with other libraries. Calculation of the feature importance for the previously developed model took from 1.5 to 5 min, depending on the number of permutation rounds for each feature. Finally, InterpretML provided the most interactivity out of all previously described libraries. Invoking global explanation functions provided a menu system alongside visualizations to investigate feature importance and their interaction. Each visualization enabled extensive inspection through zoom, select, lasso, and export functionality. Despite this interactivity, limitations of InterpretML library are due to the requirement of using built-in GlassBox models such as ExplainableBoostingClassifier. Although showing comparable performance, this restriction to built-in models is quite significant. Furthermore, the additional computation overhead of training an additional model should not be overlooked. Overall, from the perspective of global explainability, all identified libraries provide useful insight into the model behavior, and each comes with its merits and limits from the perspective of visualization options, interactivity, and computational overhead.Models that produce accurate predictions and, at the same time, can explain such predictions are crucial. Researchers often generate global explanations, which try to explain predictions of black-box learning algorithms. However, such a global explanation cannot clarify the prediction of every single instance in the model. Local explainability focuses on gaining the user’s trust for individual predictions and then trusting the model as a whole. Interpretation should make sense from the point of view of individual prediction. Globally important features may not be important locally and vice versa. In this case, the aim is to understand model decisions with respect to local context rather than the global behavior of the model.There are several solutions mentioned in this paper and in this section; we will focus on the local explanations and two most relevant Python libraries, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) [16], identified by the selection rule mentioned in the previous section.We identified three different libraries that fit to the selection rule of the most relevant libraries which are implementations of the SHAP approach: InterpretML [28], Dalex [29], and SHAP [17]. Consequently, we compared and analyzed these libraries showing the state-of-the-art in the topic of SHAP values aiming for the interpretation of black-box models.Dalex (shown in Figure 6) offers basic interaction such as hovering over the visualization. This provides an opportunity to navigate through the results easily. Moreover, it provides the option to download the chart directly from generated visualization.SHAP offers various visualization such as waterfall graphs for global analysis and force plots for local analysis. We specifically compared local interpretation based on the force plot shown in Figure 7 and Figure 8. SHAP provides many alternatives to interpret black-box behaviors, such as the force plot of a single prediction shown in Figure 7, which is a static visualization. Additionally, in Figure 8 a grouped analysis of all predicted instances is shown, where the single instances are stacked over the x-axis. This interactive visualization provides the opportunity to select a method (e.g., ordered by similarity) to order the instances over the x-axis group the results using the drop-down menu on the top of the chart over the x-axis. Moreover, on the y-axis, the drop-down menu offers the option to select the feature which the user wants to analyze. Moreover, hovering over the chart highlights different details, thus increasing the level of information provided from this approach.In contrast, InterpretML provides an opportunity to navigate through different instances using a drop-down menu, presented in Figure 9. The estimated SHAP results for the specific instance are shown automatically by selecting a particular instance. This provides an opportunity to navigate through different instances, having a better overview of the results and the possibility to compare the output of different instances faster. In particular, information such as the predicted class, actual class, and residual error for each instance is shown in the drop-down menu, as well as in the main window. This provides an opportunity to compare similar instances based on predicted class, actual class, or the residual error, thus showing an opportunity to understand a model’s class prediction more comprehensively. Moreover, interactions such as zoom in, zoom out, pan, select, and download are supported. However, InterpretML supports only KernelSHAP methods.Although InterpretML provides multiple interaction possibilities to explore the black-box model, it still presents the highest computation overload. InterpretML takes approximately 95.4 s modeling time per instance and 0.73 s visualization time per instance. SHAP requires 21.2 s modeling time and approximately 0.15 s visualization time for single instances charts and 1.03 s for grouped instances plots. Finally, Dalex needs fewer computation resources with around 0.143 s modeling time and 1 m and 49 s visualization time.LIME (Local Interpretable Model-Agnostic Explanations) is a popular technique that tries to explain the predictions of any classifier by learning an interpretable model locally around the prediction. The key idea behind LIME is that it is easier to approximate a black-box model by a simple model locally. The Lime library can explain any black-box classifier with two or more classes. The visualization output of the LIME library is a list of explanations, reflecting the contribution of each feature to the instance prediction (Figure 10a). Visualization provides local explainability and helps to investigate which feature changes will have the most impact on the instance prediction.Figure 10a,b show instance explanations of LIME. These figures provide explanations for an instance prediction on the class of GBM or ODG, respectively. There are three parts of LIME visualization: a class description with an accurate prediction for each class, a plot showing the impact of features, and a table with actual values in the instance. The left-most section displays prediction probabilities. For the multi-class classification task, we have three colors, blue (GBM), orange (AASTR), and green (ODG). The middle section returns the most important features. The impact of features helps the user to understand which features values are supporting class prediction positively (right side) and which features values are not supporting prediction (left side). If we take Figure 10a as an example, features are represented in two colors: blue and light sea-green. The blue bars indicate supporting (positive) scores towards an instance being predicted as GBM, while the light sea-green bar indicates contradicting (negative) scores towards its prediction. Float point numbers on the horizontal bars represent the relative importance of these features. We can see in Figure 10a that the highest positive influence have genes CIC, BCL6, PKD1L1, and ATRX.Similar to the SHAP approach, besides LIME, InterpretML and Dalex are the most relevant libraries that implement the LIME approach, based on our selection rule. The libraries Dalex and InterpretML were already mentioned and explained in previous sections. The resulting plot for Dalex is shown in Figure 11. The Figure shows an explanation for instance predicted as class GBM. The length of the bar indicates the magnitude, while the color indicates the sign (red for negative, green for positive) of the estimated coefficient. In the previous examples, Dalex offered basic interaction such as hovering over the visualization, as well as the ability to navigate through the results easily. Unfortunately, the resulting plots for the LIME method do not provide any of these features.InterpretML using the LIME approach is shown in Figure 12. As in previous examples (see Figure 9), InterpretML provides an opportunity to navigate through different instances using a drop-down menu. By selecting a specific instance, we can navigate through different instances having a better overview of the results.Regarding computation time, as can be seen in Table 3, InterpretML presents the highest computation overload. InterpretML takes approximately 7.28 s modeling time per instance and 0.72 s visualization time. LIME requires 3.63 s modeling time and 0.4 s visualization time. Finally, Dalex needs little bit more computation resources, with around 3.97 s modeling time and 0.78 s visualization time.The evaluation of features affecting the classification between the diffuse glioma (DIFG) of Glioblastoma multiforme (GBM), Anaplastic Astrocytoma (AASTR), and Oligodendroglioma (ODG) highlights various mutated genes and clinical variables depending on the underlying xAI method. Diagnosis age and survival are among the most important predictors all of the methods, followed by varying gene mutations. Capicua (CIC) depicts an important feature in all approaches and is the most valuable gene feature in Dalex, second in SHAP and InterpretML, and fourth in ELI5. Mutated IDH1 is among the top features and, from a clinical point of view, commonly used for survival prognosis in patients suffering from glioma [34]. Further important variables highlighted by the different xAI methods in different order also include other biomarkers used for clinical classification of glioma, such as ATRX, TP53, TERT, PTEN, or EGFR. Local explanations show a partly different picture and detail decisions of the algorithms on local examples. We present Figures on local instances on the class of GBM (Figure 7, Figure 9, Figure 10 and Figure 11) and (b) ODG (Figure 10). Variables changed place in the hierarchy of importance, while there is additional information on a particular variable’s prediction impact shown as negative or positive factor towards the particular class of the local view.The comparison overview and ranking is shown in Table 3. As a result, the table shows the overview concerning the global and local explainability comparison results of SHAP and LIME.In the context of global explainability, similar criteria can be used for the selection of libraries, i.e., computational overhead, implemented visualizations, and interactivity. From the perspective of the computational overhead, ELI5 provides the most lightweight solution both in terms of computational overhead and implemented visualizations and interactivity. The simple interface provides a good basis for a quick inspection of the existing model and overall model debugging. Feature importance alongside other implemented functionality (e.g., feature selection) of ELI5 can be convenient during the model development process. Increased interactivity and visualization options come with the additional computational overhead in SHAP, Dalex, and InterpretML libraries. From the perspective of interactivity in global explainability, InterpretML provides the most interactive solution. The addition of menu components to select different model components makes it easy to switch between analysis perspectives and extensive visualization features (zoom, lasso, select, and others). This provides excellent analytical insights. However, these functionalities come with limitations in terms of the limited scope of built-in Glassbox models that can be used and additional computation overhead caused by model retraining. In terms of visualization, SHAP and Dalex are in between ELI5 and InterpretML. Compared to ELI5, Dalex requires more computational overhead but provides additional interactivity and visualizations. On the other hand, SHAP requires even more computational overhead but provides excellent visualization options that enable a complex analysis of the interplay between feature importance and feature effect. From the perspective of the stage of the development of the predictive model, ELI5 and Dalex seem to be focused on the model analysis, while SHAP and InterpretML put focus on the underlying data and how this data impacts the model decisions.Regarding local explainability using the SHAP approach, we identified different outcomes. In general, to explain a black-box in the big data context, it is important to find the trade-off between computation resources and explainable results. In the context of local explainability, SHAP outperformed other libraries in terms of computational resources and providing an interactive way to explore the different model predictions. In terms of interactivity, both SHAP and InterpretML outperform Dalex and provide many options to analyze explainable results of multiple instances interactively. However, if the goal is to find a trade-off between computational overhead and interactivity, then Dalex seems as the optimal solution in this context. Finally, if the focus is on exploring the features, the SHAP force plot grouping methods provide many advantages. However, InterpretML offers the option to compare different instances in terms of feature contribution, predicted class, actual class, and residual error. This provides a huge advantage over other methods for analyzing the behavior of block box models in terms of predicted/actual class. Compared to SHAP, LIME has advantages in terms of speed as it builds the model around individual predictions. In the case of large datasets, using SHAP might not be feasible due to the large computational overhead caused by the calculation of all global permutations. Despite the performance overhead, SHAP provides a unified solution, which, once computed, offers more refined explainability and analytical experience.LIME provides an intuitive instance explanation. The LIME library builds the model around individual predictions (neighborhood), thus it does not take additional time to compute the model for all instances. On the other hand, the resulting plots do not provide any interactivity. Using Dalex for the LIME approach does not offer any interaction as for the other libraries. InterpretML is the only library providing interactivity while using the LIME approach. In comparison with the LIME plot, InterpretML’s resulting plot does not offer an extensive summary of features.The main advantage of SHAP for local explanation is that it is the only xAI method based on solid theory (Shapely value) [46]. Moreover, SHAP guarantees that the prediction is fairly distributed among all feature values. On the other hand, LIME for local explanation is faster than SHAP concerning computation time. In particular, if the aim is to analyze huge data sets, then LIME will provide a suitable alternative to the time-consuming computation of Shapely values. The SHAP approach considers this challenge by using approximation and optimization; however, not all model types are supported yet. In particular, LIME supports tabular data, text, and images. In other xAI methods, it is rare that all these types of data are supported.The output of any ML model should be comparable and interpretable. This is of particular interest to researchers in the medical domain as for cancer, where model performance may be compared with the one of clinicians [47]. Some experts from the medical domain argue that transparency for black boxes is not of primary interest to AI applications in their domain, as doctors make diagnoses based on their experience, and complete information on the causality of medical issues are rare [48,49]. However, xAI methods can help to gain new insights and forward biomedical knowledge to better understand interrelated characteristics and signaling components in pathologies.As a modeling approach, classifying glioma sub-types is exemplified: As the chosen dataset combining data from different brain tumor studies comprises sample data primarily from the glioma subtypes GBM, AASTR, and ODG, these three disease types were chosen to be classified to apply VA methods for interpreting global as well as local feature importance. The dataset provides Oncotreecode as identifier. GBM, AASTR, and ODG are all DIFG subtypes. Even combining data from six different studies resulted in a lack of samples for specific subtypes, therefore only the top three were chosen. Open data resources are still set to develop further and to be extended [50]. The chosen dataset is unbalanced and fits this use case insofar as it represents an often-found challenge in molecular sciences. This study aims to describe xAI tools rather than to provide a highly performing classifier solution; still, classifying glioma subtypes is a challenging task, which makes it an ideal example for comparing VA features in xAI. Cross-validation of xAI is not applicable to date, as a matter of ongoing research.From a biomedical point of view, many of the important variables highlighted by the various xAI methods are already known to be involved in cancer signaling and represent common biomarkers in glioma. Generally, such insights into the model can be used for validation. The transcriptional repressor CIC is part of the tyrosine kinase signaling pathway which is known to be involved in tumorigenesis, especially in GBM [51]. Other gene features impacting the classification include mutated IDH1, ATRX, TP53, PTEN, TERT, NF1, and EGFR, all of which are known to be involved in DIFG [22,52]. Among important variables are also the mucin protein family (MUC16 and MUC17) which are involved in epithelial barrier formation and potential biomarkers for favorable prognosis in DIFG, or lysine methyl transferase (KMT2B) also shown to be a player in gliomagenesis [53,54]. One example given, the type I transmembrane protein Notch 1 receptor (Notch1), is involved in the NF-κB signaling pathway effecting cancer development and progression, especially in GBM [55]. Notch 1 is listed in the global top 20 variables listed by SHAP, but not by Dalex. Still, in SHAP it distinguishes primarily between ODG and GBM. Some gene mutations are not primarily common for one class of sub disease, but can increase or mitigate cancer malignancy as given by the example of IDH1. Mutated IDH1 will lead to a favorable outcome, but a complete genetic profile could tell more of cases not concordant with standard prognoses [56]. In the case of local explanations as given in Figure 10b, IDH1 is selected in favor of the ODG class. Local explanations can thereby support further insight on individual cases instead of presenting the big picture of global classes.The local explanation in Figure 12 shows that the low mutation count has been used to select for the class of GBM for this instance. A high mutational burden is indicative for an unfavorable prognosis as given by GBM, which would contradict the observation in this local view. This could be seen as a limitation of model accuracy or be used for future investigations on individual cases and underlying experimental constraints. In Figure 10, we can see another local explanation for GBM classification which is supported by low numbers of mutation count. This could be due to the fact that a high number of samples originate from GBM biopsies, so that samples with low mutation count can also be frequently found. This unbalanced data source can be seen as a certain limitation to the represented model; however, combining local explanations in Figure 10 with global explanations in Figure 4, we can see that even if the mutation count is among the top rated features, there are also other important features that should be taken into account for further analysis. Diagnosis age and overall survival are preferably incorporated by the different algorithms on a global basis. Further local instances by InterpretML and Dalex are presented in Figure 9 and Figure 11. For example, gene mutations With-No-Lysine Kinase 1 (WNK1) are ranked among the top important features, highlighting a possible role of WNK1 in glioma, which has yet to be shown for WNK3 [57]. One local instance presented by Lime in Figure 10a ranks AT-Rich Interaction Domain 1B (ARID1B), shown as putative driver gene in glioma [58], among the most important variables for classifying GBM. The feature is followed by others such as Protein Kinase DNA-Activated Catalytic Subunit (PRKDC), a component of the autophagy-regulating signaling cascades to be alterated also in glioma [59], and the Anaplastic Lymphoma Receptor Tyrosine Kinase (ALK), whose variation has been implicated with pediatric glioma [60]. Another local instance by InterpretML, shown in Figure 9 includes Polycystic Kidney And Hepatic Disease 1 Protein (PKHD1), shown as variant in GBM [61], in the top feature list, followed by Insulin Receptor Substrate 2 (IRS2) [62] and Dynein Axonemal Heavy Chain 11 (DNAH11), which has been recently linked to immune cell infiltration in glioma [63].Applying xAI methods further facilitates the refinement process of the model’s underlying data and thereby helps to understand and enhance a model. By studying the results of local explainability methods, we found an error in the algorithm for computing the different gene’s mutations. The value “NA” had been counted as 1 rather than 0, due to the fact that different gene mutations from the processed data are handled as strings, separated by empty spaces. After evaluating and comparing the results, we corrected the model and revisited the comparison, leading to better results, both in reproducibility of already known markers and better quality, as well as model performance.The comparison of xAI libraries can be used for gaining biomedical insights, but also to detail advantages and challenges using these tools appropriate for certain application scenarios. Figure 8 and Figure 11 show two diverging examples in VA feature range such as interactivity or details on demand regarding xAI quality and quantity. After all, which library and approach to choose depends on the use case, such as finding novel biomarkers in analyzing classification feature importance or investigating survival prediction. Therefore, we compared libraries regarding their global xAI features separately from those with local ones. By making use of the detailed descriptions above, we try to support the decision-making process of choosing a suitable library. F.i. ELI5 is optimal regarding computational load, while InterpretML offers most interactivity at the expense of computation time.We present a comparison of the ease of use of current xAI libraries and exemplify how to support understanding of a black-box model’s results in glioma classification to find novel biomarkers. Thereby, we describe possibilities how to integrate VA features for xAI. We only scratch the surface when it comes to going beyond xAI. The process of understanding can be supported by interactivity and other features to assess the quality of explanations [64]. Future work may also include taking the type of mutation into account by incorporating various types of mutations as different features—for now, the model differentiates between wild-type/mutated and number of mutation if there is more than one mutation for the same gene. Additionally, data could be integrated from miscellaneous sources and cover further subclasses or clinical features, while adding use cases of survival prediction or clustering approaches for signaling insights. Performance experiments for further information on requirements and recommendations could be also part of future work. Finally, we believe that the presented approach, using open data, providing open source implementation, and focusing on ease of use, as well as showcasing the application of xAI to real scientific problems, can contribute to the research fields of cancer science and beyond.Conceptualization, C.J.-Q. and F.J.; methodology, all authors; software, M.V., N.J. and M.G.; validation, F.J.; formal analysis, C.J.-Q.; investigation, all authors; resources, all authors; data curation, C.J.-Q., F.J. and M.V.; writing—original draft preparation, all authors; writing—review and editing, F.J., C.J.-Q., A.H., S.T. and M.G.; visualization, M.V., N.J. and M.G.; and supervision, C.J.-Q. and F.J.; All authors have read and agreed to the published version of the manuscript.Parts of this work have been funded by the Austrian Science Fund (FWF), Project: P-32554 “A reference model of explainable Artificial Intelligence for the Medical Domain”. Additionally, two authors have been partially supported by the FFG, Contract No. 854184: “Pro²Future is funded within the Austrian COMET Program Competence Centers for Excellent Technologies under the auspices of the Austrian Federal Ministry for Climate Action, Environment, Energy, Mobility, Innovation and Technology, the Austrian Federal Ministry for Digital and Economic Affairs and of the Provinces of Upper Austria and Styria. COMET is managed by the Austrian Research Promotion Agency FFG”.Not applicable.Not applicable.Preprocessed data and implementations such as notebooks can be found on https://github.com/mathabaws/SOTA_xAI_Visual_analytics/ (accessed on 12 January 2022).We thank the cBioPortal maintainers and collaborators for providing data on cancer and all the other data providers to make open science possible. We dedicate our work in memoriam to our family members and friends we have lost.The authors declare no conflict of interest.The following abbreviations are used in this manuscript:
+      AASTRAnaplastic AstrocytomaDIFGDiffuse GliomaCICCapicua geneGBMGlioblastoma multiformeLIMELocal Interpretable Model-Agnostic ExplanationsODGOligodendrogliomaSHAPSHapley Additive exPlanationsVAVisual AnalyticsxAIexplainable Artificial IntelligenceThe full table listing all search results and filter criteria for comparing explainable libraries can be found via https://github.com/mathabaws/SOTA_xAI_Visual_analytics/tree/main/data (accessed on 12 January 2022).The repository containing code and experiments can be found via https://github.com/mathabaws/SOTA_xAI_Visual_analytics (accessed on 12 January 2022).Overview of the review process.Visualization of feature weights and their impact to the model in the ELI5 library.Visualization of permutational variable importance in the Dalex library.(a) Visualization of the impact of different variables in the global model performance in the SHAP library (b) Visualization (summary ploy) that combines feature importance with feature effects for a specific class (class “GBM” in this case).Visualization of overview of feature importance provided by InterpretML.Visual explanation of black-box prediction results using the Dalex library. In this case, an exemplary local view of class GBM is detailed.Local visual explanation of black-box prediction results (exemplary instance of class GBM) using the SHAP library.Grouped based analysis of instance prediction. Local visual explanation of black-box prediction results using the SHAP library. In the x-axis, the local explanation results of every instance are stacked. The y-axis shows the contribution to prediction and the option to select the feature that will be explored for every instance in terms of SHAP contribution.Black-box model interpretation using InterpretML. In this case, an exemplary local view of class GBM is explored in detail.Local visual explanation of black-box prediction results using LIME library—exemplary instance prediction of class (a) GBM and (b) ODG.Visual explanation of black-box prediction results using the Dalex library (class GBM)—LIME approach.Visual interpretability of black-box model using the InterpretML library with the LIME approach. In this case, the local view of class GBM is explored in detail. Colors are encoded as follows: blue = negative contribution, orange = positive contribution, and gray = intercept.Predictive results using RF classifier.Summary containing library names and analyzed properties.Library comparison with respect to global and local explainability.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00010.txt

@@ -0,0 +1 @@
+Deep learning techniques, such as convolutional neural networks (CNNs), generative adversarial networks (GANs), and graph neural networks (GNNs) have, over the past decade, changed the accuracy of prediction in many diverse fields. In recent years, the application of deep learning techniques in computer vision tasks in pathology has demonstrated extraordinary potential in assisting clinicians, automating diagnoses, and reducing costs for patients. Formerly unknown pathological evidence, such as morphological features related to specific biomarkers, copy number variations, and other molecular features, could also be captured by deep learning models. In this paper, we review popular deep learning methods and some recent publications about their applications in pathology.With the development of artificial intelligence and machine learning techniques in the past decade, many deep-learning-based computer vision models are playing important roles in daily life, and have revolutionized various industries through their superior performance and efficiency in prediction tasks, such as autopilot, machine translation, electronic sports, and biometry [1,2,3,4,5,6]. Recently, these technologies have also shown their extraordinary potential and capabilities in solving many complicated questions in the biomedical field by analyzing massive amounts of biomedical data, such as protein structural predictions with Alphafold which outperforms experimental results [7,8], and tumor segmentation in MRI scans [1]. In particular, computational pathology, a discipline that involves the effort of both pathologists and informaticians, has especially benefitted from the advancement of deep learning in recent years [9,10,11,12]. Several models have also been demonstrated to be useful in clinical diagnoses based on histopathology images [13,14,15]. In addition, some model-extracted morphological features show correlations with features at a molecular level, including single mutations and subtypes, most of which are previously unknown to human pathologists and clinicians [14]. Here, we discuss these deep learning methods from a technical perspective and summarize their successful applications in pathology from recent publications. Deep learning is a type of machine learning method that is using a multi-layer perceptron called artificial neural networks (ANN) [1,2,16]. Training a deep learning model involves designing and selecting a neural network architecture, loss functions, and evaluation metrics, as well as tuning the hyperparameters of batch size, step size, and regularization methods [1,2,17,18]. Convolutional neural networks (CNNs), variants of ANNs, have proved their power in tackling various computer vision tasks, such as image classification, segmentation, and object detection [19,20,21,22,23]. The first modern CNN architecture, LeNet5, was introduced by Yann LeCun et al. in 1998 [24]. This gradient-based six-layer convolutional neural network shows its power in recognizing hand-written digits and characters [24]. However, the development of CNNs was restricted by limited computational compacities and resources for over a decade. The advancement of computational hardware in recent years, especially graphical processing units (GPUs) and tensor processing units (TPUs), empowers the development of deep neural networks. Many CNN architectures, such as AlexNet [25], VGG [26], InceptionNet [20], and ResNet [27], can be trained into models that even outperform human beings in a computer vision classification challenge called ImageNet, which contains 1.2 million high-resolution images of more than 1000 classes [19,27,28,29,30]. AlexNet was introduced in 2012. This architecture is much larger than the previous LeNet5, with 650,000 neurons and 60 million trainable parameters packed into this design, with 5 convolutional and 3 fully connected layers [25]. Overlapping max pooling, ReLU nonlinearity, and dropout regularization are also incorporated. Due to the development of hardware, AlexNet at that time could only be trained and run on two GPUs [25]. It achieved a top-five test error rate of 15.3%, which made it the winner of the ILSVRC-2012 competition and outperformed the second-best model by more than 10% [25]. The overwhelming success of AlexNet drew people’s attention back to CNNs, and numerous other new architectures, including the VGG, InceptionNet and ResNet, were developed in the following years.VGG architecture was introduced in 2014, when it won the ImageNet challenge [26]. Compared with AlexNet, VGG increases the depth of the model by adding more convolutional layers with smaller convolutional filters [26]. However, with the introduction of newer architectures in the following years, VGG architecture has lost its popularity due to the gigantic size, high complexity to train, and less accurate performance.InceptionV1 architecture was announced in 2015, with the name GoogLeNet, which is a 22-layer deep CNN (Figure 1) [20]. The two key innovations that make Inception architectures outstanding are the inception module and auxiliary classifier. The inception module consists of multiple convolutional kernels with different sizes on the same layer [20]. This design allows the model to capture similar features of various sizes. Deep CNNs are prone to overfitting and passing gradient updates through the entire network is hard, which is often referred to as the vanishing gradient problem. By adding auxiliary classifiers in the middle of the network, the auxiliary loss from the middle of the model is taken in the final loss calculation, so that the gradients also represent the middle part of the network [20]. InceptionV2 and InceptionV3 were introduced in 2016, which modified the inception module by factorizing the larger kernels into a stack of smaller kernels to make the architecture more computationally efficient [28]. In addition, InceptionV3 uses the RMSProp optimizer and adds batch normalization into the auxiliary classifiers, significantly improving the performance, with a top-five error of 3.57% and top-one error of 17.2% on ImageNet, much better than a human [28]. InceptionV4 further refined the architecture by adding reduction blocks and unifying the inception modules [29]. A major competitor of InceptionNet is Resnet, which applies the idea of residual connection (Figure 2) [27]. In this architecture, each layer learns the residuals from the previous layer with reference to the layer inputs [27]. The top-five error on ImageNet is 3.57%, which is similar to the performance of InceptionV3 [27]. Interestingly, this residual connection idea was later adapted by the InceptionNet team to develop InceptionResNetV1, a modified version of InceptionV3, and InceptionResNetV2, a modified version of InceptionV4 [29]. Using the residual connection, InceptionResNetV2 achieved a markedly improved 3.1% top-five error on ImageNet [29]. Ever since the introduction of deep neural networks, people have been eager to know what their models have learned [31]. For image-based tasks with CNN models, visualizing the captured features is the most straightforward way. Class activation mapping (CAM) and saliency maps are two simple ways to visualize the learned features by projecting the weights and gradients of the output layer back to the input image [32,33,34]. However, these visualization methods are image-specific and will only roughly imply where the models are focusing. In addition, many saliency methods have been criticized recently for giving misleading visualization interpretations, and researchers are advised to use them with caution [35]. To unveil the CNN models further, direct deconvolution and indirect optimization are the two major approaches [36]. Deconvolution starts with finding an image from the dataset that triggers high activity to the neuron of interest and the gradient of neuron activity is calculated [36]. In general, a deconvolutional network is a reversed convolutional network, which maps features back to pixels [37]. However, deconvolution visualization can be noisy and may contain features that are not easy to interpret [38]. The indirect optimization approach can provide more accurate visualization than that from deconvolution [39]. The algorithms optimize the colors of the pixels of an image to maximize activation of the neuron of interest [38,39,40]. Once a set of optimized images for many neurons has been obtained, a dimensional reduction visualization method, such as UMAP and tSNE, can create an atlas that systematically displays the correlations of features captured by different neurons at the same layer [38,41,42,43]. Graph neural networks (GNNs) are a type of neural network which deal with data consisting of relational information [44]. Data with a non-Euclidean structure of information, such as particle interactions, molecular structures, and object relationships in images, could be modeled by GNNs [45]. In general, GNNs can be further classified into four categories: recurrent GNNs, convolutional GNNs, graph autoencoders, and spatial–temporal GNNs [45]. Generative adversarial networks (GANs) are a type of neural network consisting of two networks that are trained at the same time [46]. The generator part is trained to create fake images which tries to fool the discriminator, while the discriminator, trained with both real and generated fake images, is able to distinguish them [46]. Many variants of GANs have been applied to different tasks, such as style transfer, the visualization of neural networks, and object segmentation [46,47,48,49,50,51,52,53]. Cycle-GAN, a GAN variant using cycle-consistence loss to train two pairs of generators and discriminators simultaneously, has become increasingly popular for image-to-image translation tasks [50]. Unlike conditional GANs, which require two styles of paired data, cycle-GANs only need two sets of images of two styles, which significantly lowers the data requirements while preserving the quality of style transfer [50]. With the success of CNN models in various real-world computer-vision classification tasks, researchers and scientists have also trained and tested these models in case scenarios in biomedical fields, including pathology. These studies may involve training an existing CNN architecture from scratch. However, it requires more data, and the data augmentation techniques may not always be suitable for biomedical images. Alternatively, transfer learning techniques, which freeze most of the parameters from a model often pre-trained on ImageNet, have more advantages in terms of the data size requirements. For example, an InceptionV3-based ImageNet pre-trained CNN model can achieve a high level of accuracy in determining skin lesion malignancy and the possibility of melanoma  [13,54]. In clinical settings, pathologists typically examine histopathology slides under microscopes to provide diagnosis or other clinical information. Due to the development of digital pathology equipment, digitizing histopathology slides is cheaper and more accessible. As a result, more and more deidentified digital histopathology slide images have become available in many databases. These images, often with extremely large dimensions, are saved in special image file formats (e.g., .svs or .scn), which is a tuple of the same image with different resolutions [55]. Thus, in order to fit these digital histopathology images into CNN architectures, people usually develop their own customized pipelines with commonly used techniques, such as tiling the whole slide images (WSI) or sampling regions of interest (ROIs) (Figure 3) [56]. In the past few years, classification CNN models trained on histopathology images have shown phenomenally high performance and promising clinical potential in predicting both morphological features and molecular features. The visualization techniques also reveal results that often match pathologists’ expectations and many models are generalizable to independent real-world clinical images. For example, Inception and InceptionResNet architectural models demonstrate high accuracy and other statistical metrics in predicting subtypes and key biomarker mutations, such as STK11 and EGFR, in non-small-cell lung cancer histopathology slides [14,57,58]. With the integration of other critical clinical variables and images, immune response, G-CIMP, and telomere length can be predicted in glioblastoma patients [59]. BRAF mutation, a well-known biomarker in malignant melanoma, can also be accurately predicted with a CNN-based model [60]. Other molecular and genomic features, such as microsatellite instability (MSI), can be predicted from histopathology slides with a reasonable accuracy as well [15]. The critical gene expression level could also be inferred by applying these CNN classification models to WSI [61]. Some contemporary models also show successful classification results in the histopathology images of multiple tissue types [62,63]. These successful cases indicate that CNNs represent a suitable approach to study the correlation between molecular features and morphological features in histopathology slides, some of which may be undetectable or often ignored by human pathologists. However, histopathology images are quite different from the images in the ImageNet because of their extremely large sizes, higher resolution, and sparser useful feature distributions [11,55,64]. Deep learning architectures that could take advantage of these characteristics are very likely to achieve better results, unveiling more interesting hidden features in histopathology image classification tasks. For instance, a multi-resolution CNN model, which takes advantage of the data structure of .svs and .scn image files, achieves higher performance in classifying endometrial cancer molecular features than its single resolution counterparts [65]. Weakly supervised techniques, such as multiple instance learning, also demonstrate decent performance in classification tasks of histopathology images, and have gained popularity in recent years [64,66,67]. The innovative idea of bringing GNN models into solving histopathology classification problems develops greater capacity in understanding the subtle relationships between features of different tissue structures and at different locations on giant digital histopathology slides [68,69]. In addition to classification tasks, CNN models are also capable of segmenting cells or tissue in histopathology slides [9,55]. The segmented cells or tissue could then be used to train classification models for different prediction tasks, including the recurrence of non-small-cell lung cancer [70] and endometrial tissue types [71]. A popular segmentation CNN architecture used in the biomedical field is U-net, which has a similar structure to an autoencoder [72]. A 3D version of U-net, which has 3D convolutional layers instead of 2D convolutional layers, is capable of segmenting volumetric images [22]. Modified U-net architectures, such as USE-net [73], Focus U-net [74], and U-net with attention gate [75], have achieved even better performance in various biomedical image segmentation tasks than vanilla U-net. Other autoencoder-based methods have also achieved promising results in segmentation tasks of histopathology images, such as highlighting tumor regions in liver cancer WSI [76]. Well-trained style transfer models are also viable options for segmentation tasks [48]. With the introduction of GAN, using conditional-GAN or cycle-GAN models and in combination with CNN models for segmentation problems is also shown to be viable, with less stringent training data requirements [46,53,77]. Unlike most classification models, the segmentation models can be more adaptive to different types of tissues due to the similarities of the stained features and textures of the histopathology slides [78]. Additionally, the evaluation metrics of these classification models can be drastically different from those of the classification models. The segmentation labels are also usually images; therefore, it is not easy to determine a binary prediction or even a prediction score at the per-image level. Hence, typical statistical metrics, such as AUROC or precision and recall, are often not capable of fairly evaluating segmentation tasks. Pixel-level metrics, such as intersection over the union (IoU), also pose weaknesses because it cannot objectively give relative importance to pixels of different regions. Object-level metrics can be an optimal alternative, but the requirement of identifying all objects on the label images prohibits its adoption in real-world model evaluation. Therefore, researchers often use customized evaluation metrics with a combination of customized pixel weights, dice loss, and IoU with specific thresholds [79,80]. In this review paper, we have introduced popular deep learning algorithms, CNN, GNN, and GAN, and also highlighted mechanisms of how they work and how they can be applied to solve clinical and scientific questions in pathology. We have also discussed recent publications which show that these deep learning techniques have the potential to be useful in classifying or segmenting histopathology imaging data. With the continuing advancement of machine learning and deep learning techniques and the development of hardware and software, it is realistic to believe that the integration of artificial intelligence and pathology will become an even more attractive field to explore. Compared with conventional computational methods, deep learning techniques generally run faster and have much better performance in pathology tasks. Although one has to be rigorous and ethical about translating these AI-based technologies into clinical settings, we still hold an optimistic view that they will eventually revolutionize medical diagnosis processes and really push the development of precision medicine forward. Above all, the ultimate goal of introducing AI into pathology and biomedicine in general is to make healthcare more accessible, affordable, and agreeable. Nevertheless, there are still a number of limitations of the current studies and potential obstacles which prevent these models from implementation in contemporary real-world clinical settings. For example, only patterns with prior understanding from pathologists can be used as reliable evidence for prediction, which significantly limits the tasks for which deep learning models can be applied. In addition, the patient samples that can be used as training, validation, and test sets are also very limited for each of the specific tasks of interests. Moreover, detailed labels of medical images are often not available, and the labeling standards among clinicians also vary significantly in different countries. Additionally, the interpretability of deep learning models applied to histopathology images remains debatable, especially among clinicians. More advanced self-supervised or semi-supervised methods may solve some of these problems from a technical perspective in the future. Conceptualization, R.H. and D.F.; methodology, R.H. and D.F.; resources, R.H.; data curation, R.H.; writing—original draft preparation, R.H.; writing—review and editing, D.F.; visualization, R.H.; supervision, D.F.; project administration, D.F.; funding acquisition, D.F. All authors have read and agreed to the published version of the manuscript.This work was supported by NIH/NCI U24CA210972.Not applicable.Not applicable.Not applicable.We would like to thank all members of the Fenyö laboratory and the administration team of ISG at NYU. The authors declare no conflict of interest.Diagram of the inception module [19], containing 4 branches with convolutional kernels of different sizes.The concept of residual connection in ResNet [27]. Here, the middle 2 convolutional layers are skipped by the residual connection.Typical classification model pipeline for histopathology images.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00011.txt

@@ -0,0 +1 @@
+The RNA polymerase III–specific TFIIIB complex is targeted by oncogenes and tumor suppressors, specifically the TFIIIB subunits BRF1, BRF2, and TBP. Currently, it is unclear if the TFIIIB subunit BDP1 is universally deregulated in human cancers. We performed a meta-analysis of patient data in the Oncomine database to analyze BDP1 alterations in human cancers. Herein, we report a possible role for BDP1 in non-Hodgkin’s lymphoma (NHL) for the first time. To the best of our knowledge, this is the first study to report a statistically significant decrease in BDP1 expression in patients with anaplastic lymphoma kinase–positive (ALK+) anaplastic large-cell lymphoma (ALCL) (p = 1.67 × 10−6) and Burkitt’s lymphoma (BL) (p = 1.54 × 10−11). Analysis of the BDP1 promoter identified putative binding sites for MYC, BCL6, E2F4, and KLF4 transcription factors, which were previously demonstrated to be deregulated in lymphomas. MYC and BDP1 expression were inversely correlated in ALK+ ALCL, suggesting a possible mechanism for the significant and specific decrease in BDP1 expression. In activated B-cell (ABC) diffuse large B-cell lymphoma (DLBCL), decreased BDP1 expression correlated with clinical outcomes, including recurrence at 1 year (p = 0.021) and 3 years (p = 0.005). Mortality at 1 (p = 0.030) and 3 (p = 0.012) years correlated with decreased BDP1 expression in ABC DLBCL. Together, these data suggest that BDP1 alterations may be of clinical significance in specific NHL subtypes and warrant further investigation.Lymphoma is characterized by the deregulated growth of lymphocytes, including natural killer, B-, and T-cells. It is anticipated that, in 2021, 8830 individuals will be diagnosed with Hodgkin’s lymphoma (HL), and 81,560 individuals will be diagnosed with non-Hodgkin’s lymphoma (NHL) in the United States (US) [1]. Thus, NHL accounts for approximately 80% of all lymphoma diagnoses in the US. Subtypes of aggressive (fast-growing) NHL include: diffuse large B-cell lymphoma (DLBCL), anaplastic large-cell lymphoma (ALCL), Burkitt’s lymphoma (BL), lymphoblastic lymphoma (LBL), mantle cell lymphoma (MCL), and peripheral T-cell lymphoma (PTCL) [2].Uncontrolled cell proliferation is a common characteristic of many human cancers, including aggressive forms of lymphoma [2]. The regulation of eukaryotic cell proliferation is controlled by three distinct RNA polymerases (pol) [3], including RNA pol III, which controls transcription of untranslated small RNA molecules involved in processing and translation. Together, these regulate the biosynthetic capacity of a cell. Accurate transcription by RNA pol III requires general and gene-specific transcription factors [3], including the RNA pol III–specific TFIIIB complex (3, 4). To date, two forms of TFIIIB have been well characterized in humans [4,5], with both forms of human TFIIIB requiring BDP1 [4,6]. In humans, multiple BDP1 isoforms have been identified [4]. The identified eukaryotic BDP1 isoforms contain a conserved SANT domain (Swi3, Ada2, N-Cor, and TFIIIB) involved in chromatin remodeling and transcription regulation [3,4]. The isolated BDP1 isoforms vary in the length of c-terminal extensions characterized by a 55-residue repetitive motif [3,4,5,6,7].TFIIIB activity, via the TBP, BRF1, and BRF2 subunit(s), is targeted both directly and indirectly by various oncogenes and tumor suppressors [7,8]. For example, the oncogenes MAP kinase ERK and MYC [9,10] stimulate TFIIIB activity in vitro. The tumor suppressors p53 [10,11], PTEN [12,13], BRCA1 [14], the retinoblastoma protein (RB) [10], and the Rb family members p107 and p130 [15] inhibit TFIIIB activity. The TFIIIB subunit BRF2 is deregulated in various human cancers and is an oncogene in lung squamous cell carcinoma [8,16,17,18]. To date, alterations in BDP1 have been demonstrated in nonsyndromic hereditary hearing loss [17] and have been shown to promote tumorigenicity in TP53-mutated prostate cancers [18]. Additionally, BDP1 is overexpressed in cells transformed by papovaviruses [19]. However, specific BDP1 alterations in human cancers have not been investigated.In this study, we queried patient data from the Oncomine microarray database and an integrated data-mining platform to analyze BDP1 alterations in human cancers using publicly available datasets. Herein, we report a possible role for decreased BDP1 expression in lymphoma for the first time. To the best of our knowledge, this is the first study to report a statistically significant decrease in BDP1 expression in patients with anaplastic lymphoma kinase–positive (ALK+) anaplastic large-cell lymphoma (ALCL) (p = 1.67 × 10−6) and Burkitt’s lymphoma (BL) (p = 1.54 × 10−11). An analysis of the BDP1 promoter identified putative binding sites for myelocytomatosis oncogene (MYC), B-cell lymphoma 6 protein (BCL6), E2 factor transcription factor 4 (E2F4), and Krüppel-like factor 4 (KLF4) transcription factors. MYC, BCL-6, E2, and E2F4 are demonstrated to be deregulated in lymphomas. Specifically, MYC and BDP1 expression are inversely correlated in ALK+ ALCL, suggesting a possible mechanism for the significant and specific decrease in BDP1 expression. In activated B-cell (ABC) diffuse large B-cell lymphoma (DLBCL), decreased BDP1 expression correlated with clinical outcomes, including recurrence at 1 year (p = 0.021) and 3 years (p = 0.005). Mortality at 1 (p = 0.030) and 3 (p = 0.012) years correlated with decreased BDP1 expression in ABC DLBCL. Together, these data suggest that BDP1 alterations may be of clinical significance in lymphoma and warrant further investigation.From September 2019 through December 2021, we performed comprehensive queries of the Oncomine Research Premium Edition platform [20,21]. The Oncomine Research Premium Edition platform is a cancer microarray database and web-based data-mining platform [20] containing 729 datasets (91,866 samples) to determine the frequency of BDP1 alterations in human cancers. The Oncomine Research Premium Edition platform uses statistical tests conducted both as two-sided for differential expression analysis and as one-sided for specific over- and underexpression analysis [20,21]. For the overall study analysis, p-values were corrected for multiple comparisons by the false discovery rate method [20,21]. For BDP1 expression analyses in specific datasets, cutoff values, sample numbers, and p-values are indicated in the figure legends. The Oncomine™ Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization. The public datasets used are noted in Table 1, with study descriptions and hyperlinks to the available datasets, and are cited in figure legends.The Eukaryotic Promoter Database [25] (https://epd.epfl.ch//index.php, accessed on 5 January 2022 was queried to identify putative transcription factor binding sites within the BDP1 promoter, specifically targeting transcription factors known to be deregulated in NHL. A threshold p-value of 0.001 was used while querying the Eukaryotic Promoter Database [25].Recently, it was observed that p53-deficient prostate cancer cells display high levels of BDP1 [18], suggesting that deregulation of BDP1 may be of functional significance in human cancers. Deregulation of the human TFIIIB subunits BRF1 [26,27,28,29,30] and BRF2 [7,16,27,31,32,33,34,35,36,37] has been well documented in human cancers. The primary aim of this study was to determine if the TFIIIB subunit BDP1 is specifically altered in human cancers and if the observed alterations correlate with clinical outcomes. Oncomine 4.5 was queried for BDP1 expression in 729 datasets (91,866 samples) based on cancer type, cancer versus normal, and cancer versus cancer, including histology and multicancer analysis and outlier analyses. The disease summary analysis for BDP1 is presented in Figure 1. Red shading of boxes denotes gene overexpression; blue shading represents decreased gene expression. This disease summary was performed using the following criteria: a minimum 2-fold change in BDP1 gene expression, a p-value of 1 × 10−4, and a gene rank percentile of 10%. BDP1 was overexpressed in breast and colorectal cancer versus normal datasets but underexpressed in breast and lymphoma cancer versus normal datasets (Figure 1A). In cancer versus cancer datasets, BDP1 was over- and underexpressed in kidney cancer (cancer histology dataset) (Figure 1A). In a cancer subtype analysis, BDP1 was overexpressed in castrate-resistant metastatic prostate cancer [38] (n = 122, p = 2.60 × 10−11), suggesting that alterations in BDP1 may be of clinical significance, as previously reported [18]. Interestingly, BDP1 expression was decreased in the pathway and drug analysis in lung cancers (Figure 1A). Specifically, BDP1 expression decreased in the HCC 1299 lung cancer cell line transfected with the epidermal growth factor receptor (EGFR) and treated with the EGFR signal transduction inhibitor gefitinib (p = 2.40 × 10−5) [39]. The disease summary analysis for BDP1 includes an Oncomine outlier analysis reporting the number of unique datasets in which BDP1 had the highest-ranking cancer outlier profile analysis (COPA) score [20]. The outlier analysis demonstrated that BDP1 was both over- and underexpressed across the analyzed cancer datasets.Interestingly, BDP1 was significantly overexpressed in colorectal cancer (p = 2.07 × 10−5, 105 patients, median gene rank of 318) across five datasets (10%) (Figure 1B). However, further analysis did not identify any correlation with clinical outcomes (unpublished data). In contrast, BDP1 was significantly underexpressed in lymphoma (p = 8.37 × 10−7, 131 patients, median gene rank of 107) across two datasets (28.5%) (Figure 1C). To the best of our knowledge, BDP1 alterations in lymphoma have not been investigated previously. Thus, the observed statistically significant BDP1 underexpression in lymphoma datasets (Figure 1A,C) warranted further in-depth analysis.In Figure 2A, we show BDP1 expression in the Brune lymphoma dataset [38]; BDP1 expression was significantly decreased in BL (normal versus cancer; gene rank of 25, p = 1.54 × 10−11, fold change of −2.148, n = 67). In Figure 2B, we show BDP1 expression in the Eckerle lymphoma dataset [39]; BDP1 expression was significantly decreased in ALK+ ALCL, an ALCL subtype that responds well to standard chemotherapy treatment (gene rank of 190, p = 1.67 × 10−6, fold change of −2.635, n = 64) [23].Subsequently, we determined if the observed statistically significant decrease in BDP1 expression in lymphoma patients was unique to BDP1 or was common to the TFIIIB subunits BRF1 and BRF2. The heat maps depict TFIIIB expression in BL (Figure 2C,D) and ALK+ ALCL (Figure 2E,F). There was no significant decrease in BRF1 or BRF2 expression concurrent with the significant decrease in BDP1 expression (Figure 2C,E). However, BRF1 (p = 5.75 × 10−4) and BRF2 (p = 8.50 × 10−4) were significantly and specifically overexpressed exclusively in the Brune lymphoma dataset (Figure 2D) [22]. Only BRF2 (p = 0.005) was overexpressed in the Eckerle lymphoma dataset (Figure 2F) [23].Gene expression profiling identifying molecular heterogeneity in various lymphomas has provided additional genetic information with the potential to develop targeted therapies [40]. As such, we re-examined the Brune [22] and Eckerle [23] lymphoma datasets to identify the top over- and underexpressed genes in BL and ALK+ ALCL to determine the potential significance of BDP1 underexpression in the context of these NHL subtypes (Figure 3). Using the Steidl lymphoma [41] concept cluster (Oncomine cluster-ID n9239), we queried the Brune [22] and Eckerle [23] lymphoma datasets to identify genes with the top median gene rank that were significantly over- and underexpressed. The top genes identified as significantly under- or overexpressed are labeled with median gene rank, p-value, and fold change in gene expression. Analysis of the Brune dataset (Figure 3A,B) identified BDP1 (p = 1.54 × 10−11) as significantly underexpressed in BL vs. normal (log2 median-centered intensity), with a median gene rank of 25 and −2.15-fold change (Figure 3A). The significantly overexpressed genes in BL are presented in Figure 3B using the same parameters utilized in Figure 3A. Analysis of the Eckerle dataset (Figure 3C,D) identified BDP1 (p =1.67 × 10−6) as significantly underexpressed in ALK+ ALCL vs. normal (log2 median-centered intensity), with a median gene rank of 190 and −2.63-fold change (Figure 3C). The identification of significantly overexpressed genes in ALK+ ALCL is presented in Figure 3D using the same criteria as in Figure 3C.The results in Figure 3 prompted an in silico analysis to determine if the transcription factors with significantly altered expression have the potential to deregulate BDP1 activity and expression. The BDP1 protein contains a highly conserved SANT domain, originally identified in Swi3, Ada2, N-Cor, and TFIIIB (yeast BDP1) [3,4]. The SANT domain has been implicated in DNA binding and chromatin remodeling [42]. Of the transcription factors with significantly altered expression in BL and ALK+ ALCL identified in Figure 3, we identified two transcription factors with the potential to regulate BDP1 activity via the SANT domain (Table 2).Further analysis of deregulated transcription factors in BL and ALK+ ALCL identified histone deacetylase 4 (HDAC4) as specifically and significantly overexpressed in both BL and ALK+ ALCL (Figure 3). Previously, SANT domain proteins were identified in chromatin remodeling complexes [45]. In addition, corepressor of nuclear receptors (n-CoR) is a SANT domain protein known to interact with and activate histone deacetylase 3 (HDAC3) [46]. In esophageal carcinoma, HDAC4 overexpression has been associated with poor survival and promotes tumor progression [47]. It was recently demonstrated that the miRNA miR-155 targets HDAC4 and indirectly regulates B-cell lymphoma 6 (BCL6) expression, a key event in B-cell leukemia development [48]. A meta-analysis of the diffuse large B-cell lymphoma patient microarray data demonstrated that miR-155 expression inversely correlates with HDAC4 and BCL6 [48]. More experiments are required to determine if BDP1, potentially via the SANT domain, interacts with HDAC4, playing a regulatory role in NHL. In addition, we noted that the REST corepressor 3 (RCOR3) was both significantly over- and underexpressed in both BL and ALK+ ALCL.There are three REST corepressor family members (1–3), each with two SANT domains [44]. Deletions in RCOR1 are associated with unfavorable survival outcomes in patients with DLBCL [49]. RCOR1 and RCOR2 facilitate nucleosome demethylation during blood cell maturation, whereas RCOR3 inhibits this process [50]. It is unclear if RCOR3 plays a role in NHL.We could not determine whether the observed significant decrease in BDP1 mRNA expression in BL and ALK+ ALCL is the result of decreased transcription from the BDP1 promoter or decreased BDP1 mRNA stability. Thus, we examined the BDP1 promoter for putative transcription factor binding sites known to play a role in lymphoma. Thus, we performed a query of the Eukaryotic Promoter Database [25] for lymphoma-associated putative transcription factor binding sites in the BDP1 promoter, located −1000 to +100 relative to the transcriptional start site (TSS), using a cutoff p-value of 0.001 [25]. Table 3 summarizes the location of putative binding sites in the BDP1 promoter for lymphoma-associated transcription factors.In NHL, Krüppel-like factor 4 (KLF4) has been characterized as a tumor suppressor, and overexpression inhibits cell proliferation in BL cell lines [51]. In ALK+ ALCL, KLF4 overexpression is significant (p = 0.002, gene rank 2491) and BDP1 expression is significantly decreased (p = 1.67 × 10−6, gene rank 190) (Figure 4). Thus, the KLF4 binding sites identified at −734, −593, −553, −459, and −291 may partially explain the significant decrease in BDP1 observed in subtypes of NHL [51].In NHL, MYC amplification is associated with poor prognosis [52]. It is plausible that MYC binding to the BDP1 promoter, located at −582 and −581, may permit the recruitment of MYC-associated proteins to silence BDP1 expression. MYC is significantly overexpressed (p = 1.4 × 10−7, gene rank 218) and BDP1 expression is significantly decreased (p = 1.67 × 10−6, gene rank 190) in ALK+ ALCL (Figure 4). BCL6 overexpression has been implicated in lymphoma [53]. We identified BCL6 binding sites in the BDP1 promoter at −985, −936, −384, −362, −287, −276, and −173, relative to the transcription start site (TSS). Overexpression of BCL6 has been implicated in lymphoma [53]. However, both BCL (p = 7.31 × 10−4, gene rank 1145) and BDP1 (p = 1.67 × 10−6, gene rank 190) expression are significantly decreased in ALK+ ALCL (Figure 4A) [53]. We did not observe the same correlation in Burkitt’s lymphoma using the Brune dataset (data not shown) [22].We identified several putative Forkhead box protein P1 (FOXP1) and E2 factor transcription factor 4 (E2F4) binding sites in the BDP1 promoter (Table 3). FOXP1 is overexpressed in a subset of DLBCL patients [54]. In ALK+ ALCL, both FOXP1 (p = 2.62 × 10−5, gene rank 414) and BDP1 (p = 1.67 × 10−6, gene rank 190) expression is significantly decreased (Figure 4). Decreased E2F4 protein expression in BL tumor samples has been reported [55]. However, our analysis of E2F4 expression shows that it remains relatively unchanged in ALK+ ALCL (Figure 4). Analyses of these putative binding sites suggest that the MYC binding sites at −582 and −581 in the BDP1 promoter may play a key role in regulating BDP1 mRNA expression in ALK+ ALCL.Together, these data suggest a role for BDP1 alterations in NHL patients. To determine if the decreased BDP1 expression in lymphoma subtypes is clinically relevant, we examined whether BDP1 expression changes correlated with clinical outcomes. Clinical outcomes analyses, depicted in Figure 5, were performed using the Shaknovich lymphoma dataset [24] (n = 69). In activated B-cell (ABC) DLBCL [40], decreased BDP1 expression correlated with clinical outcomes, including recurrence at 1 year (p = 0.021) and 3 years (p = 0.005) (Figure 5A,B).DLBCL patients can be divided into two groups based on expression profiling: ABC DLBCL and germinal center B-cell (GCB) DLBCL subtypes [56]. Patients with ABC DLBCL have poorer clinical outcomes than GCB patients [56]. Consequently, we carried out an analysis to determine if BDP1 expression and mortality were correlated in ABC DLBCL. Mortality at 1 (p = 0.030) and 3 (p = 0.012) years correlated with a decrease in BDP1 expression in ABC DLBCL (Figure 5C,D). Using the Shaknovich lymphoma dataset [24], recurrence (p = 0.614) and mortality (p = 0.858) outcomes in Figure 5 did not correlate with BDP1 expression in patients with GCB DLBCL (data not shown). However, BDP1 underexpression correlated with mortality at 1 (p = 0.023) and 3 (p = 0.009) years in the Lenz DLBCL dataset [56] (data not shown).A variety of predictive and diagnostic biomarkers have been defined in ABC DLBCL [57], including the following common loss-of-function ABC DLBCL molecular biomarkers: beta-2-microglobulin (B2M), CD58 molecule (CD58), cyclin-dependent kinase inhibitor 2A (CDKN2A), CREB-binding protein (CREBBP), E1A-binding protein p300 (EP300), myeloid/lymphoid or mixed-lineage leukemia 2 (MLL2), and the myeloid differentiation primary response gene (88) (MYD88) [57]. FOXP1 [57] and the melanoma-associated antigen (mutated) 1 (MUM1) are immunohistochemical biomarkers in ABC DLBCL [57]. We queried the Shaknovich [24] lymphoma dataset in Oncomine to determine if BDP1 and established ABC DLBCL biomarkers correlate with clinical outcomes, and the results are shown in Figure 6. The expression heat maps represent recurrence at one year (Figure 6A) and three years (Figure 6B) and ABC DLBCL patients who died at one year (Figure 6C) and three years (Figure 6D); p-value and fold-change are indicated. BDP1 expression was significantly decreased, as were many of the established ABC DLBCL biomarkers. This significant correlation of BDP1 expression with both clinical outcomes and identified biomarkers in lymphoma suggests that more extensive analyses are warranted to determine if decreased BDP1 expression is a global feature of DLBCL or specific to DLBCL subtypes.To the best of our knowledge, this is the first study to identify BDP1 alterations in NHL. In this study, we performed a meta-analysis of cancer patient data from the Oncomine web-based data-mining platform to analyze BDP1 alterations in human cancers. Interestingly, there is a statistically significant decrease in BDP1 expression in patients with ALK+ ALCL (p = 1.67 × 10−6) and BL (p = 1.54 × 10−11). To potentially identify mechanisms that drive the decrease in BDP1 mRNA expression, we analyzed the BDP1 promoter for transcription factor binding sites with relevance in NHL. Analysis of the BDP1 promoter identified putative binding sites for MYC, BCL6, E2F4, and KLF4 transcription factors, which were previously demonstrated to be deregulated in lymphomas. MYC and BDP1 expression are inversely correlated in ALK+ ALCL, suggesting a possible mechanism for the significant and specific decrease in BDP1 expression. In ABC DLBCL, decreased BDP1 expression correlated with clinical outcomes, including recurrence at 1 year (p = 0.021) and 3 years (p = 0.005). Mortality at 1 (p = 0.030) and 3 (p = 0.012) years correlated with decreased BDP1 expression in ABC DLBCL. Lastly, BDP1 underexpression correlates with previously identified biomarkers in ABC DLBCL patient clinical data. DCBCL is the most common lymphoma diagnosed in adults, with ABC DCBCL having a poor prognosis [58].All microarray dataset analyses have limitations and should be interpreted with caution. In this study, we examined BDP1 alterations in NHL and clinical outcomes. We exclusively used publicly available microarray datasets from the NCBI Gene Expression Omnibus repository containing clinical outcome data identified using the Oncomine Research Platform in Figure 1. We believe that larger studies of NHL patients using RNA-seq analysis would provide an unbiased approach to analyzing all transcripts in a genome.Additional clinical studies are required to determine if the observed correlation between BDP1 expression and clinical outcomes is specific to ABC DCBCL, potentially identifying BDP1 as a predictive biomarker in ABC DCBCL, or a general observation in NHL. Together, the data presented suggest that BDP1, a unique factor in the RNA pol III machinery, may be a novel target for therapeutic intervention for patients with NHL and warrants further investigation in the clinic.L.S. conceived the study, performed data analysis, prepared figures, and prepared the manuscript. S.C.-P. performed data analysis and revised the manuscript. All authors have read and agreed to the published version of the manuscript.This research received no external funding.Not applicable.Not applicable.The present study used publicly available datasets archived in NCBI Gene Expression Omnibus. Hyperlinks to datasets are provided in the Methods section with study descriptions. Data analysis was performed using Oncomine Research Edition, retired on 17 January 2022.The authors thank St. John’s University for funding this research.The authors declare no conflict of interest.BDP1 expression is significantly altered in a subset of human cancers. (A) Oncomine 4.5 database disease summary for BDP1. Oncomine 4.5 was queried for BDP1 expression in 729 datasets (91,866 samples) based on cancer type, cancer versus normal, and cancer versus cancer, including histology and multicancer analysis types and outlier analyses. Red shading of boxes denotes gene overexpression; blue shading represents decreased gene expression. This disease summary was performed using the following criteria: a 2-fold change for gene expression, a p-value of 1 × 10−4, and a gene rank percentile of 10%. BDP1 was overexpressed in breast and colorectal cancer vs. normal datasets but underexpressed in breast and lymphoma cancer vs. normal datasets. In cancer vs. cancer datasets, BDP1 was over- and underexpressed in kidney cancer (cancer histology dataset). BDP1 was overexpressed in prostate cancer (metastasis vs. primary) in a cancer subtype analysis but decreased in drug and perturbation analysis in lung cancers. The outlier analysis demonstrated that BDP1 was both over- and underexpressed across analyzed cancers. (B) BDP1 was significantly overexpressed in colorectal cancer (p = 2.07 × 10−5, 105 patients) across 5 datasets (10%). (C) BDP1 was significantly underexpressed in lymphoma (p = 8.37 × 10−7, 131 patients) across 2 datasets (28.5%). The Oncomine™ Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization.BDP1 mRNA is significantly and specifically underexpressed in lymphoma. (A) BDP1 expression in Brune lymphoma [22] (Burkitt’s lymphoma vs. normal), gene rank of 25 (top 1%), p-value = 1.54 × 10−11, fold change of −2.148, n = 67. (B) BDP1 expression in Eckerle lymphoma [23] (anaplastic large-cell lymphoma, ALK-positive vs. normal), gene rank of 190 (top 1%), p-value = 1.67 × 10−6, fold change of −2.635, n = 64. Heat maps denoting underexpression of the TFIIIB subunit BDP1 (C,E) in the Brune [22] and Eckerle [23] lymphoma datasets are specific. BRF1 and BRF2 were significantly overexpressed exclusively in the Brune dataset (D). Only BRF2 was significantly overexpressed in the Eckerle dataset (F). The Oncomine™ Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization.Heat map identifies BDP1 expression as significantly underexpressed in BL and ALK+ ALCL. Using the Steidl lymphoma [41] concept cluster (Oncomine cluster-ID n9239), we queried the Brune [22] and Eckerle [23] lymphoma datasets to identify genes with the top median gene rank that are significantly over- and underexpressed. (A) BDP1 (p = 1.54 × 10−11) was significantly underexpressed in BL vs. normal (log2 median-centered intensity), with a median gene rank of 25 and −2.15-fold change. The top genes identified as significantly underexpressed are labeled with median gene rank, p-value, and fold change in gene expression. (B) Identification of significantly overexpressed genes in BL using the same parameters identified in (A). (C) BDP1 (p = 1.67 × 10−6) was significantly underexpressed in ALK+ ALCL vs. normal (log2 median-centered intensity), with a median gene rank of 190 and −2.63-fold change. The top genes identified as significantly underexpressed are labeled with median gene rank, p-value, and fold change in gene expression. (D) Identification of significantly overexpressed genes in ALK+ ALCL using the same criteria applied in (C). The Oncomine™ Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization.Coexpression analysis of BDP1, KLF4, MYC, BCL6, FOXP1, and E3F4 in ALK+ ALCL. Under- (A,B) overexpression of BDP1, KLF4, MYC, BCL6, FOXP1, and E3F4 in ALK+ ALCL. The analysis was performed using the Eckerle lymphoma dataset [23]. Gene rank, fold change in expression, and p-values are indicated. The Oncomine™ Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization.BDP1 expression in activated B-cell (ABC) diffuse large B-cell lymphoma (DLBCL) correlates with clinical outcomes. BDP1 expression was significantly altered in ABC DLBCL recurrence at 1 (A) and 3 (B) years. In addition, in ABC DLBCL, BDP1 expression was significantly altered in patients who died in year 1 (C) and year 3 (D). Clinical outcomes analyses were performed using the Shaknovich lymphoma dataset [24], n = 69. The Oncomine™ Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization.Correlation of ABC DLBCL biomarkers and BDP1 expression with clinical outcomes. BDP1 expression and ABC DLBCL biomarker expression were significantly altered in ABC DLBCL recurrence at 1 (A) and 3 (B) years. ABC DLBCL biomarker and BDP1 expression were significantly altered in patients who died in year 1 (C) and year 3 (D). Clinical outcomes analyses were performed using the Shaknovich lymphoma dataset [24], n = 69. The Oncomine Platform (Thermo Fisher, Ann Arbor, MI, USA) was used for analysis and visualization.Public datasets used in this study. Study descriptions and hyperlinks to datasets are provided.Transcription factors significantly altered in BL and ALK+ ALCL known to interact with the SANT domain of BDP1.Identification of transcription factors deregulated in NHL with putative transcription factor binding sites in the BDP1 promoter.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00012.txt

@@ -0,0 +1 @@
+Previous studies demonstrate the feasibility of predicting obesity using various machine learning techniques; however, these studies do not address the limitations of these methods in real-life settings where available data for children may vary. We investigated the medical history required for machine learning models to accurately predict body mass index (BMI) during early childhood. Within a longitudinal dataset of children ages 0–4 years, we developed predictive models based on long short-term memory (LSTM), a recurrent neural network architecture, using history EHR data from 2 to 8 clinical encounters to estimate child BMI. We developed separate, sex-stratified models using 80% of the data for training and 20% for external validation. We evaluated model performance using K-fold cross-validation, mean average error (MAE), and Pearson’s correlation coefficient (R2). Two history encounters and a 4-month prediction yielded a high prediction error and low correlation between predicted and actual BMI (MAE of 1.60 for girls and 1.49 for boys). Model performance improved with additional history encounters; improvement was not significant beyond five history encounters. The combined model outperformed the sex-stratified models, with a MAE = 0.98 (SD 0.03) and R2 = 0.72. Our models show that five history encounters are sufficient to predict BMI prior to age 4 for both boys and girls. Moreover, starting from an initial dataset with more than 269 exposure variables, we were able to identify a limited set of 24 variables that can facilitate BMI prediction in early childhood. Nine of these final variables are collected once, and the remaining 15 need to be updated during each visit.While previously uncommon in young children, obesity is now a worldwide epidemic affecting over 40 million children under the age of 5 [1,2]. Obesity in childhood is associated with both adverse outcomes like hyperlipidemia, diabetes and hypertension [3,4,5,6], as well as with higher morbidity and mortality in adulthood [7]. The underlying causes of obesity are modifiable risk factors throughout the life course; these risk factors represent major causes of health inequalities [8]. Thus, the prevention of obesity is considered a national and global health priority [9].Unhealthy weight gain during early childhood significantly increases the risk for obesity later in life [10,11], so the ability to identify children at a young age who carry the greatest risk for obesity could significantly improve prevention efforts [12]. Several important and potentially modifiable indicators of obesity have been identified during this timeframe, including rapid infant weight gain, poor infant sleep quality, birth weight, and maternal characteristics (e.g., current and pre-pregnancy weight, depression) [13,14]. Despite this, there has been relatively limited research into predictive modeling of childhood obesity risk, leaving many unanswered questions about how and when to intervene.Existing research to evaluate obesity risk has predominantly employed logistic regression techniques, with limited success. The constraints of traditional regression approaches (e.g., restricting analyses to a relatively small number of predictors and assumptions of independence and linearity) have prompted others to examine non-linear interactions via machine learning [14,15,16]. Machine learning is increasingly recognized as useful for preventive care [17] because of its ability to characterize, adapt, learn, predict and analyze clinical data. However, one of the main challenges in employing machine learning in the clinical domain is that electronic health record (EHR) data are often incomplete and irregularly sampled (e.g., lacking regular time intervals between patient visits). In addition, height and weight, which are necessary to calculate BMI, are collected during pediatric visits in the first 2 years of life [18], but not routinely as pediatric appointments are often missed [19]. These issues hinder the performance of predictive models using EHR data. Recent techniques in deep learning and artificial neural networks address these issues and have the potential to predict health outcomes more accurately by using EHR data.In this study, we used a longitudinal, EHR-derived dataset of children to investigate the medical history needed for a recurrent machine learning model to accurately predict BMI prior to age 4 years. Our secondary aim was to understand whether BMI prediction varies considerably between boys and girls, which would require separate BMI prediction models for each sex.Previous studies have used machine learning techniques to develop obesity prediction models or to determine key determinants of obesity for designing intervention tools [14,20]. However, as discussed by Siddiqui et al. [20], very few of these studies analyze sex-specific prediction models, use large-scale datasets, or examine geographic/neighborhood exposure variables (e.g., access to food and opportunities for physical activity) [21,22,22,23,24] that might be associated with childhood obesity [25,26,27].Existing models of childhood obesity risk also tend to focus on predictive variables that are routinely collected in clinical practice [28], and therefore tend to include only biological predictors and postnatal factors like infant sex and birthweight [29]. It has been suggested that one of the reasons for the intractability of childhood obesity is the failure to take into account the complexity and interconnectedness of contributing factors across the life course, ranging from the social, built, and economic environments to behavior, physiology, and epigenetics [30]. A number of childhood obesity risk factors that operate during the first 1000 days of life have been identified [13] and have special significance for obesity risk prediction. For instance, programming effects occurring during pregnancy increase children’s obesity risk. Adding this information could lead to improvements in a model’s ability to identify children at risk for obesity in early life, but EHR data typically contain information on maternal prenatal risk factors separately from risk factors during infancy and from measures of height and weight across childhood. The models presented in this study leverage data from a population-based, longitudinal database that combines data from multiple stages of the life course and thus add a valuable contribution to our understanding of obesity risk in early life.Finally, the lack of effective interventions to reduce the risk for obesity in early life [31,32] suggests that efforts must be made to identify very young children with a high risk of developing obesity that could be specifically targeted for intervention. The methodology in the present paper employs long short-term memory (LSTM) [33] models to predict children’s BMI prior to age 4 using different lengths of history data, determined by the number of previous clinical encounters. LSTM is a recurrent neural network model that learns from an ordered sequence of events, in this case, prior clinical encounters of the patient. While several machine learning techniques could have been used, an LSTM model was selected because the history encounter constitutes a time series. In particular, the variables height and weight that are used to calculate BMI as well as the age of the child vary from one encounter to the next. LSTM models are particularly well suited for time-series applications and continue to outperform other architectures in various fields. For example, in Wang et al.’s analysis [34], LSTM outperformed RF, SVM, Naive Bayes, and Feed forward neural networks when predicting patient-reported outcomes using history responses from cancer patients. In other applications [35], LSTM models were used to predict post-operative risk for patients suffering from obesity and risk for complications after bariatric surgery.Data were extracted from the Obesity Prediction in Early Life (OPEL) database, a unique longitudinal, epidemiologic data repository that combines birth certificate, contextual-level, and health outcome data for 19,857 children born in Marion County, Indiana. We constructed the OPEL database by linking three independent data sources:The Child Health Improvement through Computer Automation (CHICA) system; a computer-based pediatric primary care clinical decision support system that operated in eight pediatric primary care practices in Indianapolis between 2004–2019 [36]. The CHICA system includes data for over 47,000 patients on factors such as measured height and weight, demographics (e.g., child sex, age, race/ethnicity, Medicaid insurance status), and social determinants of health (e.g., parent health literacy, food and housing insecurity, parental depression, and infant feeding practices);The IN Standard Certificate of Live Birth (i.e., ‘birth certificate’), which consists of 235 variables covering parental sociodemographic information as well as information on prenatal care, labor/delivery, and neonatal conditions and procedures. Birth certificate data were made available from the Marion County Public Health Department (MCPHD); andThe Social Assets and Vulnerabilities Indicators (SAVI) Project, which collects geocodes, organizes, and presents integrated data on communities in the 11-county Indianapolis metropolitan statistical area drawn from more than 30 federal, state, and local providers. All are linked to the lowest available geographic level [37]. SAVI is the nation’s largest community information system, with more than 10,000 time-series variables from 1980 to the present, including welfare, education, health, public safety, housing, demographics, locations of health facilities, health and human services, community facilities, and associated service areas.The Child Health Improvement through Computer Automation (CHICA) system; a computer-based pediatric primary care clinical decision support system that operated in eight pediatric primary care practices in Indianapolis between 2004–2019 [36]. The CHICA system includes data for over 47,000 patients on factors such as measured height and weight, demographics (e.g., child sex, age, race/ethnicity, Medicaid insurance status), and social determinants of health (e.g., parent health literacy, food and housing insecurity, parental depression, and infant feeding practices);The IN Standard Certificate of Live Birth (i.e., ‘birth certificate’), which consists of 235 variables covering parental sociodemographic information as well as information on prenatal care, labor/delivery, and neonatal conditions and procedures. Birth certificate data were made available from the Marion County Public Health Department (MCPHD); andThe Social Assets and Vulnerabilities Indicators (SAVI) Project, which collects geocodes, organizes, and presents integrated data on communities in the 11-county Indianapolis metropolitan statistical area drawn from more than 30 federal, state, and local providers. All are linked to the lowest available geographic level [37]. SAVI is the nation’s largest community information system, with more than 10,000 time-series variables from 1980 to the present, including welfare, education, health, public safety, housing, demographics, locations of health facilities, health and human services, community facilities, and associated service areas.Institutional Review Board approval to construct the OPEL database was obtained from the Indiana University School of Medicine. All data analyses for this study occurred on a restricted-access server provisioned specifically for research purposes.From the OPEL database, we identified 73,957 clinical encounters from 6614 children ages 0 to 4 years. Within this limited dataset, we performed data preprocessing to remove erroneous records, impute missing values, and encode variables into normalized features for use in our predictive model. For example, encounters where height decreased more than 2 inches from the previous encounter or with implausible recorded BMIs were categorized as input error. We also established valid ranges for the mother’s gestational weight gain and the child’s birth weight. Variables that were one-hot encoded (e.g., race of the mother or father) were converted to multi-class nominal variables. Finally, we deleted duplicative variables, administrative variables not directly relevant to the aims of our analysis, and variables without enough data to be useful.This preprocessing yielded a list of 269 variables derived from the OPEL database that we initially considered for modeling (Appendix A). From this list, we performed feature reduction guided by existing peer-reviewed literature on early life obesity risk (e.g., [13]), expert opinion (ERC), and the results of a LASSO regression. Feature reduction also took into account noisy and sparsely populated variables.Our outcome of interest was BMI as defined by the Center for Disease Control and Prevention (CDC) guidelines [38]. We imputed missing and invalid BMIs using linear interpolation and height and weight data from previous encounters.After preprocessing, we randomly selected an equal number of boy and girl patients, then split the dataset by patient such that 80% of our data was used for model training and 20% was used for model testing while maintaining an equal split according to patient sex. We normalized all input variables to values between −1 and 1. In the initial dataset, the girl class was the minority class.We then developed separate long short-term memory (LSTM) [33] models to predict BMI using different lengths of history data, determined by the number of previous clinical encounters. We defined history data as either 2, 3, 5, or 8 prior encounters, and modeled our predictions of patient BMI at each encounter immediately following the set of history encounters. We modeled predictive variables as both fixed (e.g., maternal and paternal race, infant birthweight, mother’s age at birth) and varying (e.g., patient’s age, visit type, sleep quality) between encounters.The model architecture consisted of an LSTM layer followed by a single Feed forward linear layer. The number of hidden nodes in the LSTM layer was set to half the number of input features. The Adam optimizer was used to update the weights in the model. Each model was trained using an input-output sequence with a varying number of history encounters. For example, when using five history encounters the model was trained to predict BMI at the sixth encounter.Based on prior research demonstrating different obesity determinants for boys and girls [39], we developed three models: one for boys, one for girls, and a combined model for both. K-fold cross-validation [40] with k = 5 was used to evaluate each model and to estimate variabilities induced by the data selection. The accuracy of the models was measured using MAE and Pearson’s correlation coefficient (R2). We report the standard deviation of these metrics from the K-fold cross-validation.The feature reduction process resulted in a set of 24 exposure variables: 15 were derived from the CHICA dataset, 7 from the birth certificate, 1 from CHICA/birth certificate, and 1 from SAVI (Table 1).Table 2 and Figure 1 show the distribution of the patients in the training and testing cohorts. As designed, there were approximately the same number of boys and girls included in both training and testing cohorts. There were no clinically meaningful differences across the cohorts in terms of mean BMI and age at the clinical encounter. The mean age at the encounter, defined as the average age across all encounters, was approximately 68 weeks (17 months), with no difference between the training and testing cohorts. There were also no significant differences between the cohorts with respect to the average number of encounters during the study period, although the average number of encounters for boys showed a higher standard deviation than for girls.Data in Table 2 were used to develop the three types of models discussed above. The boy BMI model used a total of 2694 patients during training and was tested on 657 patients. Similarly, the girl model was trained on 2614 patients and tested on 649 patients. The combined model was trained using both training cohorts (i.e., 5308 boy and girl patients) and was tested on the combined testing cohorts (i.e., 1306 boy and girl patients).Table 3 and Figure 2 show the results of the LTSM models. Models with five or eight history encounters were determined to more accurately predict the patient’s BMI than models using two or three history encounters. These models fit the observed data well, as shown by the mean average error and correlation between actual BMI and predicted BMI. Models were not trained with more than eight encounters due to concerns of reduced data quantity. Mean average error and correlation estimates were less optimal when using two or three history encounters, with the highest mean average error (1.49 for boys and 1.60 for girls) and the lowest correlation between actual and predicted BMI observed using two history encounters (R2 = 0.55 in the boy only model and R2 = 0.49 in the girl only model). Moreover, the K-fold standard deviation was low for both the mean average error and the R2 in models with five and eight history encounters, indicating that these models were not susceptible to the selection of the training data and were more likely to generalize to new data. We observed higher K-fold standard deviations in models with two or three history encounters, suggesting less optimal performance in predicting BMI.The above-mentioned advantages of the five and eight history encounter models were achieved despite having longer prediction horizons compared to the two or three history encounters models. For instance, the five history encounters boy model had an average prediction horizon of more than 20 weeks. That is, the model predicted BMI, on average, 20 weeks into the future. Conversely, the two history encounters model had an average prediction horizon of less than 18 weeks.We did not observe significant model differences between boys and girls. The combined model showed optimal performance with the lowest mean average error (0.98, SD = 0.03) and the highest correlation (R2 = 0.72), likely owing to the greater number of patients included.Within the entire cohort, the mean age at which children reached five clinical encounters was 10.1 months with a standard deviation of 6.5 months.The purpose of this study was to understand the importance of historical health data in developing machine learning models to identify pediatric patients with increased risk of future overweight and obesity. Our LSTM models suggest that clinical data from at least five clinical encounters are needed to accurately predict child BMI prior to age four years with prediction horizons approximately 20 weeks in the future. In contrast to prior research [39], our combined model performed better than the models separated by sex, negating the need to develop and employ separate models for boys and girls.Although previous studies have successfully applied machine learning to predict childhood obesity [14], few have investigated the application of these models in clinical care [28]. Our model could be employed in a pediatric clinical setting to dynamically track and predict children’s BMI progression, facilitating obesity prevention through anticipatory guidance during each wellness visit. The results also suggest that having height and weight data from at least five clinical encounters may be necessary to accurately predict future BMI values. Encouragingly, the majority of patients in our sample achieved this threshold within the first 17 months of life, with 10 months being the average age at which children reached five clinical encounters. This suggests that employing our model to identify children at risk for suboptimal weight outcomes is feasible in very early childhood.The input variables used by our model are consistent with previous findings in the literature [13]. For instance, characteristics of children’s sleep such as duration, timing, and quality have been associated with obesity [41,42]. In this study, we conducted an ablation test on the two sleep quality variables (i.e., frequency of nighttime waking and parental perception of sleep quality) for the combined boys and girls model with five history encounters. The result of the ablation test shows a higher mean average error (1.03 vs. 0.98) with a larger standard deviation (0.07 vs. 0.03). The BMI correlation also dropped from 0.72 to 0.70, underscoring the important association of early sleep quality for the prediction of children’s obesity risk.Pediatricians are well-positioned to provide parents with information regarding obesity risk in early life, but many consensus guidelines recommend obesity screening in the pediatric setting only after 2 years of age when the “tipping point” of obesity onset may have already passed [43]. Further, meta-analyses indicate that BMI surveillance and counseling have only marginal effects on reducing children’s BMI [44]. There is evidence that unhealthy weight gain in very early childhood of age tracks into later childhood, adolescence, and adulthood [10,11], which suggests that new approaches to help providers and parents address this problem are needed. Our screener, administered in the clinic setting, could help identify very young children at risk of unhealthy weight gain, enabling preventive counseling focused on healthy feeding, activity, and family lifestyle behaviors. Even though our findings show statistical support for postponing BMI prediction until it is possible to obtain information from five clinical encounters, the proposed models still facilitate early identification and intervention as existing guidelines recommend at least this many pediatric visits by six months of age [18]. The prediction horizon of 20 weeks and the frequency of encounters during children’s first year of life means that there are numerous opportunities for providers to monitor growth, identify weight issues, and take appropriate action.Consistent with prior research [45], the performance of our models diminished as the temporal distance between the acquisition of the exposure variables and the time of BMI prediction in the future increased. While requiring only two history encounters is attractive in practice because it enables the use of the model for a wider population, the high mean average error of the resulting predictive models makes their utility to predict obesity risk limited. The model’s improvement when using five history encounters suggests that more clinical data are needed before one can correctly predict future BMI. However, further research is needed to evaluate the reproducibility and generalizability of our models before they can be applied in clinical practice for similar and related populations. Future work may wish to investigate the relative importance of the variables in our model using an external validation dataset and by conducting ablation experiments as performed in the present study for the subjective sleep quality variables.Machine learning has been widely applied in the field of obesity research, both for the prediction of future weight outcomes and for identifying targets for intervention. Several previous studies proposed classifiers for obesity in both adults and for early childhood. For instance, Thamrin et al. [46] used linear regression and various machine learning approaches (Bayesian networks and CART models) to classify adults 18 and older as having or not having obesity based on survey data on indicators such as age, parental obesity, and activity level. Here, we predict children’s future BMI rather than classify risks for obesity. We stipulate that the transparency of our proposed approach can better support intervention. Another earlier study by Dugan et al. [47] used longitudinal data from CHICA to compare different machine learning techniques (decision trees, random forest, and Bayesian networks) using 167 features from the first 2 years of life. They found that decision trees provided the best accuracy when predicting obesity between ages 2 and 10 years. Our study expands on this work by using historical data to predict children at risk for obesity. Other research focused on machine learning and obesity prediction has provided thresholds for obesity rather than BMI [48,49,50], which may not be as applicable for patients at younger ages. The models proposed in the present paper estimate exact BMI values and are dynamic. They predict future BMI based on the nearest history and can therefore be used for children of varying ages. Moreover, the proposed models leverage routinely-collected EHR data, which is a practical approach compared with previous models that, for example, predict obesity using more costly and less accessible genetic data [48,51]. Importantly, the limited number of features we identify makes our model practical for use in other settings. Although the relatively narrow set of variables we identify are not all typically included in the EHR, they could be easily collected using existing screeners [28]. This data collection approach was successfully used in previous studies to obtain child birthweight and weight change between birth and 6, 9, and 12 months [52]; and to obtain data on paternal weight, maternal smoking, and breastfeeding [53].Our study is subject to some limitations. First, it is possible that our results may be confounded by child age. While the distribution of the data (Table 2) shows that the average at encounter is approximately 68 weeks for all cohorts, patients with five or eight encounters may be older than those with two or three encounters. Their BMI may be more stable and easier to predict. This potential for confounding is the subject of a current investigation. In addition, the EHR data within the OPEL database is derived from a predominately low-income, urban population in Indianapolis, IN. Additional work in other populations is needed to externally validate our findings, as children’s growth patterns may vary by socioeconomic factors [54]. Finally, we were unable to examine other variables that are potentially impactful to children’s early weight gain, like physical activity, as they were not included in the OPEL database. Future research may wish to incorporate such measures for a better understanding of the children’s weight trajectories.The present study shows that five history encounters and a limited number of exposure variables are sufficient to predict BMI for both boys and girls in very early childhood. These findings can inform efforts to identify infants at risk of developing overweight and obesity. We envision using the proposed model in a pediatric clinic to dynamically track the progression of children’s BMI four months into the future during each wellness visit. Our findings have implications for future work aimed at early identification and intervention of obesity, as well as for other chronic diseases that begin in early life.Conceptualization, E.R.C. and Z.B.M.; methodology, R.S. and Z.B.M.; software, Z.B.M.; validation, R.S. and Z.B.M.; formal analysis, R.S. and Z.B.M.; resources, E.R.C.; data curation, E.R.C., R.S., and Z.B.M.; writing–original draft preparation, E.R.C.; writing–review and editing, R.S. and Z.B.M.; supervision E.R.C. and Z.B.M.; funding acquisition, E.R.C. All authors have read and agreed to the published version of the manuscript.This work was supported by NIH Grant K01 DK114383.The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Institutional Review Board of Indiana University School of Medicine (protocol code 2006099750, approved 8 March 2020).Not applicable because this is a retrospective study.The data presented in this study are available on request from the corresponding author. The data are not publicly available due to privacy laws.The authors wish to thank Sami Gharbi for his contribution to the data acquisition and interpretation.The authors have no conflict of interest to disclose.Complete list of starting features before LASSO reduction by data source.Distribution of average child age at the encounter.Results from the long short-term memory (LSTM) models: mean average error (MAE) by number of history encounters, stratified by child sex.Features from the OPEL database used in the analysis.Number of patients, average BMI, age, and number of encounters per patients included in the training and testing datasets.SD, standard deviation; * Represents the average number of encounters during the timeframe of analysis.Results from the long short-term memory (LSTM) models: mean average error, Pearson’s correlation coefficient, and mean prediction horizon in weeks.Each entry is the mean value of all folds in a 5 K-fold evaluation. MAE, mean average error; SD, standard deviation.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomedinformatics/biomedinformatics-02-01-00013.txt

@@ -0,0 +1 @@
+Polysomnography is the gold-standard method for measuring sleep but is inconvenient and limited to a laboratory or a hospital setting. As a result, the vast majority of patients do not receive a proper diagnosis. In an attempt to solve this issue, sleep experts are continually looking for unobtrusive and affordable alternatives that can provide longitudinal sleep tracking. Collecting longitudinal data on sleep can accelerate epidemiological studies exploring the effect of sleep on health and disease. These alternatives can be in the form of wearables (e.g., actigraphs) or nonwearable (e.g., under-mattress sleep trackers). To this end, this paper aims to review the several attempts made by researchers toward unobtrusive sleep monitoring, specifically sleep cycle. We have performed a literature search between 2016 and 2021 and the following databases were used for retrieving related articles to unobtrusive sleep cycle monitoring: IEEE, Google Scholar, Journal of Clinical Sleep Medicine (JCSM), and PubMed Central (PMC). Following our survey, although existing devices showed promising results, most of the studies are restricted to a small sample of healthy individuals. Therefore, a broader scope of participants should be taken into consideration during future proposals and assessments of sleep cycle tracking systems. This is because factors such as gender, age, profession, and social class can largely affect sleep quality. Furthermore, a combination of sensors, e.g., smartwatches and under-mattress sleep trackers, are necessary to achieve reliable results. That is, wearables and nonwearable devices are complementary to each other, and so both are needed to boost the field of at-home sleep monitoring.Sleep is important for the physical and mental well-being of an individual. The quantity and quality of sleep are generally associated with chronic diseases and health risks such as diabetes, cardiovascular diseases, renal failure, anxiety, and depression [1,2]. The fast pace of modern society and the rapid increase in the aging population have contributed to the population of people being affected by sleep disorders. The Centers for Disease Control and Prevention (CDC) reported that a third of the United States population does not get enough sleep [1]. A similar statistic was reported by the Canadian Men’s Health Foundation (2016) stating that 30% of Canadian men are sleep deprived. This is reflective of the global populace as sleep disorders are rapidly becoming a global concern, leading to a range of societal problems [3]. Sleep monitoring is important and could be a lifesaver for people with undiagnosed sleep disorders [4,5]. A major motivation for sleep monitoring is the effect it has on health and well-being [6]. However, the process requires trained sleep technicians to perform polysomnography (PSG).The PSG is the medical gold standard for sleep studies. It uses various intrusive sensors to record multiple physiological signals during sleep, namely electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), electrocardiogram (ECG), body position, oronasal airflow, photoplethysmogram, abdomen, and thorax respiratory efforts, and others (Figure 1).It is a fairly inconvenient, time-consuming, and expensive technique, used only clinically, and cannot be used for longitudinal sleep tracking [7,8]. The PSG protocol can itself affect the quality of sleep because of unfamiliarity with the environment and the multiple attachments used in the study. The results from the study (sleep score) include sleep onset latency, sleep efficiency, and sleep stages. The structure of sleep includes various stages characterized by specific physiological changes. These stages are Wake, Non-Rapid Eye Movements (NREM), and Rapid Eye Movement (REM), i.e., Wake (“wakefulness before sleep”), NREM1 or N1 (“very light sleep”), NREM2 or N2 (“light sleep” defined by EEG recordings), NREM3 or N3 (“deep sleep”), and REM (“dream state”) [9]. The transition from one stage to the next is described as the sleep cycle.Due to the complexity of PSG, other methods have been proposed as alternatives. Actigraphy (ACT) and photoplethysmography (PPG) are two solutions that enable long-term monitoring and produce a valid assessment of sleep/wake behavior. Metrics derived from longitudinal sleep tracking can help detect and manage various diseases, e.g., cardiorespiratory disorders and dementia [10]. That is, the collection of longitudinal sleep data on a large scale can boost epidemiological studies that examine the influence of sleep on health and disease [11]. There are also less cumbersome approaches to sleep monitoring owing to the advancement, adoption, and integration of technology into healthcare in the form of non-contact systems, wearables, and mobile systems [11,12,13,14,15]. These systems capitalize on the strong correlation between bio-vital signs and sleep.Researchers have been focusing on creating non-contact sleep tracking methods (e.g., under-mattress sleep trackers shown in Figure 2) that can achieve closer outcomes to PSG [16]. These systems can potentially be used for sleep-quality monitoring. Examples include systems working on the principle of ballistocardiography (BCG) (Sadek et al. [17]), strain gauge (Lima et al. [18]), seismometer (Li et al., 2018), ultrasonic (Hsu et al., 2017; Tran et al., 2019), ultra-wideband system (Kang et al., 2020), RF signals (Liu et al. [4]), fiber optics (Koyama et al. [19]), and smart textiles (Zhou et al. [20]).To this end, several devices and algorithms have been suggested, presented, and implemented, but fewer for sleep-cycle monitoring which is an important aspect of sleep. As a result, this paper aims to review existing works on sleep-cycle monitoring using unobtrusive sensors. The rest of the paper is organized as follows. Section 2 and Section 3 presents the methodology used in the literature selection. Discussion and opinions are presented in Section 4. Lastly, the paper is concluded in Section 5.For this review, the literature search was performed using a systematic computerized approach: IEEE, Google Scholar, Journal of Clinical Sleep Medicine (JCSM), and PubMed Central (PMC). The keywords used to retrieve publications were chosen based on common terms used in the field, from the topic under review, and database suggestions (Appendix A). Due to the number of results from the search on Google scholar (2810), sleep-cycle monitoring using contactless sensors research cannot be exhaustively reviewed. Therefore, only studies between 2016 and 2021 (1730) were considered. The titles and abstracts of the articles were screened, and references from relevant articles were scanned for other relevant publications. Articles included in the review were read and evaluated. Other conditions for inclusion were implemented tools, presentation of a method to measure sleep stage, studies published in a scientific journal or a scientific conference, studies with participants or clinical population, and studies with validation. A table showing the overview of the 14 reviewed publications is presented in Table 1.Nam et al. [21], proposed a system based on a tri-axial accelerometer and a pressure sensor to quantify sleep quality. The system was able to monitor the sleeping position, non-REM sleep time, movement, heart-rate variability (HRV), and variations in the breathing amplitude (i.e., an estimate of the presence and number of apneic episodes). The sleeping posture was determined using a wearable sensing belt integrating the tri-axial accelerometer. The wake–sleep period was determined via respiratory signals obtained through the bed pressure sensor. Ten volunteers (nine males and one female) participated in the experiments, and the system was validated against the PSG and a digital video camera. The authors managed to determine sleep quality based on three parameters (i.e., non-REM sleep time, the number of apneic episodes, and the total duration of the subject’s dominant sleeping pose). At last, the estimated sleep quality was found to be consistent with reference devices.Nguyen et al. [22] presented a prototype of their Light-weight In-ear BioSensing (LIBS) system that can be used for staging a whole-night sleep study. The system was placed inside the ear canal to continuously record electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) signals representing electrical activities of the human brain, eyes, and muscles. The overlapping features (i.e., temporal, spectral, and non-linear) of the signals were separated by a non-negative matrix factorization (NMF)-based model to match the reference signals obtained from PSG. LIBS was tested on eight participants (three females and five males) and PSG recordings were acquired in parallel. On average, the system achieved 95% accuracy for sleep-stage classification. The classification accuracy was 87% for NREM1, 89% for NREM2, and 78.31% for REM. Besides, the sensitivity was 80% and 83% for NREM1 and REM, respectively. Considering how the system was worn, the users took a survey, and it was concluded that 85.8% agreed that LIBS did not disturb their sleep.A mobile service (i.e., Sleep Hunter) for sleep quality monitoring and smart wake-up calls that detect transitions between sleep phases was introduced by Gu et al. [23]. For sleep-stage detection, Sleep Hunter employed a statistical model denoted as the linear-chain conditional random field (CRF). It used sensors installed in smartphones to monitor body movements, acoustic events, ambient lighting conditions, sleep length, and personal factors. At various stages of sleep, a smartphone placed next to the pillow will detect body movements such as rollover, leg stretching, and leg jerking. Sleep data were collected from 45 volunteers to train the CRF model, and the system was tested on 15 participants (eight males and seven females) for a month. Sleep Hunter was compared to other actigraphy-based technologies, Zeo and Jawbone Up, to validate its performance. The detection accuracy in a three-stage classification was 64.55%. The system was also capable of providing smart wake-up services based on sleep stage detection. The user can set a one-hour time frame when they want to be woken up, and the device will wake them up if it detects a light sleep stage within that time frame; otherwise, it will wake them up at the end of the hour.Tal et al. [24] conducted experiments to validate a contact-free monitoring system (EarlySense) made up of a piezoelectric sensor and a mobile application. The goals of the study were to verify the accuracy of the system in assessing sleep/wake state and sleep parameters as compared to PSG, as well as to see if the system can detect sleep architecture in different sleeping situations, such as when another person is in bed. The pressure-based device was placed under the mattress around the chest region to measure the respiratory rate, heart rate, and sleep stage. The results obtained from 63 (45 males and 18 females) participants produced comparable results with PSG for the total sleep time (TST), wake, REM, and non-REM. EarlySense produced 96.1% and 93.3% accuracy of continuous measurement of heart rate (HR) and respiratory rate (RR). The sleep detection sensitivity, specificity, and accuracy were 92.5%, 80.4%, and 90.5%, respectively.The research by Guettari et al. [25] was aimed at detecting the presence of patients in bed and estimating their sleep quality. A thermopile sensor producing thermal signals was fixed on the wall to achieve this goal. The Symbolic Aggregate Approximation (SAX) method was implemented in thermal signal segmentation processing. The SAX method created each segment by first segmenting the mid-variance and then determining its sleep phase. The system extracted the length of each thermal data segment, the variance of each segment’s thermal segment, and the level of each segment. The Kohonen self-organized map (SOM) was used to classify the signal segments into three sleep phases: deep sleep (R, N3), light sleep (N1, N2), and wake phase (W). The number of phases was fewer than professional systems, but the system was efficient for long-term monitoring. Further work is being carried out to improve the classification. Based on the 13 patients (nine males and four females) who took part in the study, the obtained classification results showed an accuracy of 87% with 95% confidence intervals for the recognition of the three sleep stages.A new approach for sleep analysis was developed and presented by Seba et al. [26]. In this approach, sleep activity was classified into three stages based on thermal signature: waking, relaxed sleep, and restless sleep. This device, which was based on temperature monitoring (both patient and ambient), was integrated into the framework of the Smart-EEG project by the SYEL—SYstèmes ELectroniques team. The system was made up of a thermopile sensor TMP007, a thermal camera, an accelerometer, and an iButton. The thermal camera was placed on the wall, the thermopile sensor was placed on a frame above the subject, and the iButton was worn on the wrist. The sensor was used to determine the temperature of the upper “bed + patient” region. Images from the thermal camera in medical format gave information that could be analyzed by experts to understand postural changes and changes in temperature measurement relating to the upper part of the “bed + patient” and the ambient environment. An inertial unit was used to obtain wrist acceleration in three axes to evaluate the responses of the thermopile sensor. iButtons were used to autonomously test the temperatures of the wrist, distal, and proximal skin. It was observed that there was a relationship between the day/night alternation, wake/sleep alternation, and high and low temperature. The temperature of the body drops during sleep and rises during the day, while skin temperature rises during sleep and falls during waking hours. The classification for calm and restless sleep was carried out using the acceleration module. In sum, this study validated several studies linking body temperature to sleep.De Zambotti et al. [27] carried out a comparison of a multi-sensor sleep tracker (ŌURA ring) with PSG in terms of measuring sleep and sleep phases. ŌURA ring is capable of detecting the pulse rate, variation in inter-beat-intervals (IBIs), and pulse amplitude from the finger optical pulse waveform. It also measures motion and body temperature and with machine learning methods, it can calculate and classify sleep into stages. The study was aimed at validating the accuracy of these functionalities. Sleep data were collected from the 41 participants (28 males and 13 females) recruited for the exercise by both the ring and PSG. The ŌURA ring showed good agreement with the PSG measurements in terms of TST, SOL, WASO, and light sleep (N1 + N2). However, the ring overestimated REM and underestimated “deep sleep” (N3). Epoch-by-epoch (EBE) analysis showed that it had a high sensitivity for detecting sleep (95.5%), 65% for detecting light sleep, 51% for detecting deep sleep, and 61% for detecting REM sleep, but low specificity in wake detection (48%). Furthermore, the accuracies of classifying PSG-defined TST ranges of (<6 h, 6–7 h, >7 h) were 90.9%, 81.3%, and 92.9%, respectively.In another validation study by de Zambotti et al. [28], the authors assessed the performance of a consumer multi-sensory wristband (Fitbit Charge 2) for sleep-stage classification versus PSG. The Fitbit device can monitor time spent awake, light sleep, deep sleep, and REM sleep, in addition to sleep/wake states. Forty-four subjects (18 males and 26 females) participated in the study, during which participants wore a Fitbit on their wrist while undergoing PSG. The data captured from the systems were compared using t-tests, Bland–Altman plots, and epoch-by-epoch (EBE) analysis. The result from Bland–Altman plots showed that Fitbit overestimated TST and “light sleep” (N1 + N2) while it underestimated SOL and deep sleep (N3). There were, however, no significant differences in the recordings for the wake after sleep onset and time spent in REM sleep. Based on the EBE analysis, Fitbit had accuracies of 96% in detecting sleep (sensitivity), 61% in detecting PSG wake (specificity), 81% in detecting “light sleep”, 49% in detecting “deep sleep” and 74% in detecting REM sleep. Overall, Fitbit achieved 82% accuracy in sleep cycle classification.Pallesen et al. [29] conducted a pilot study to validate IR-UWB pulse-doppler radar technology against polysomnography (PSG) for sleep assessment. UWB technology uses short-range radio waves with very low energy levels. This technology is based on the idea that body, limbs, and breathing motions trigger shifts in the frequency (Doppler shift) of radio waves. Twelve volunteers (six males and six females) were assessed overnight by a Novelda XeThru radar and PSG. Comparisons between bedtime and wake-up time were made using the respiratory signal. The result of the study showed the mean differences between the radar parameters and PSG estimates for SOL, WASO, and TST. The mean values obtained for accuracy, sensitivity, specificity, and Cohen kappa were 93.1%, 96.1%, 69.5% and 67%, respectively. The findings indicated that IR-UWB radar could be an alternative objective measure to actigraphy. The ability to assess movements from several parts of the body simultaneously, such as movement from the extremities and respiration movements, which both shift significantly during sleep, is an obvious advantage over actigraphy. It was also indicated that the presence of more than one person in the bed will affect the reading, as the system was unable to differentiate between movements caused by different individuals.A validation study was carried out by Tuominen et al. [30] to assess the accuracy of the BCG Beddit Sleep Tracker (BST) for monitoring sleep. Ten participants (five males and five females) were recruited for the test and data from PSG, BST, and other technologies were collected for two nights. Analysis showed that BST was able to identify SOL. However, the system underestimated wake after sleep onset and overestimated TST and sleep efficiency. There was also a poor outcome for sleep classification as BST failed to differentiate between NREM stages and did not detect the REM stage. It was concluded that further research and development work in sleep tracking devices is still required, as well as more validation studies for other emerging technologies.Kalkbrenner et al. [31] presented the assessment of their novel type-4 sleep monitor. The study was aimed at classifying sleep stages based on tracheal body sound and actigraphy. The tracheal body sound was used to extract cardiorespiratory signals which are commonly used for sleep assessment, while the IMU was used to extract movement features such as sleeping position and movement. The system was made up of a body sound microphone attached to the suprasternal notch (near the trachea) and the IMU and other peripherals (battery and Bluetooth gateway) were attached using an abdominal belt. A linear discriminant classifier was used for the sleep stage automation. Data were obtained from 53 subjects (33 males and 20 females) for validation purposes. Sleep/wake classification yielded 96.9% accuracy and 0.69 Cohen’s Kappa, Wake/REM/NREM classification resulted in 76.3% accuracy and 0.42 Kappa, and Wake/REM/light sleep/deep sleep classification produced 56.5% accuracy and 0.36 Kappa.The study by Lauteslager et al. [32] was aimed at assessing the capability of the radar-based Circadia Contactless Breathing Monitor (model C100) and proprietary Sleep Analysis Algorithm for sleep-stage classification. The system predicts bed occupancy, sleep stages, and derives standardized sleep metrics using its analysis algorithm and pulsed ultra-wideband radar. Sleep stage classification was carried out on the dataset obtained using the C100 device and PSG. For nine participants (six males and three females) in 17 nights, an epoch-by-epoch recall was 75.0%, 59.9%, 74.8%, and 57.1%, for deep sleep, light sleep, REM, and wake, respectively. The overall accuracy was 66.7%. A group from the University of Fribourg in Switzerland performed an independent validation and the recall was 70.7%, 52.5%, 83.0%, and 55.3% for deep sleep, light sleep, REM, and wake, respectively, with an overall accuracy of 62.7% using data obtained from 24 participants. A direct comparison with a Fitbit device and Philips Actiwatch showed that the C100 outperforms them in estimating TST, SOL, WASO, REM Sleep, Deep Sleep, and REM Latency.The purpose of the study by Zhang et al. [33] was to exploit Ambient Radio Signals for recognizing sleep stages and assessing sleep quality. The study presented the model, design, and implementation of SMARS, a system that uses tiny changes in breathing patterns to measure the quality of sleep. The system was built on a single RF link and has a coverage of up to 10 m to monitor breathing. It was designed using off-the-shelf devices. The system consists of a Tx equipped with two antennas that by default transmits standard Wi-Fi packets at a rate of 30 Hz, and an Rx with three antennas that capture Channel State Information (CSI) of every packet it received from the Tx. It combines instantaneous breathing rate estimation and sleep monitoring. CSI is a statistical model on the motion that was developed to take into account both reflection and scattering multipath indoors. For fast estimation, a statistical approach that examines the autocorrelation function of CSI power response was adopted. SMARS was deployed in six homes and sleep data of 32 nights (about 234 h) were collected in total. Tx and Rx were placed on opposite sides of the bed during data collection. For comparison, PSG data of six participants (five males and one female) were collected to establish ground truth. Additionally, an open dataset on four state-of-the-art RF-based respiratory monitoring systems containing 160 h of overnight sleep data was used for validation. In terms of sleep stage recognition accuracy of SMARS compared to commercial products EMFIT and ResMed, SMARS achieved accuracies of 87%, 89%, and 87% for the wake, NREM, and REM detection, respectively. This was better than EMFIT with 77%, 75%, and 46%, and ResMed with 53%, 87%, and 79% accuracies. Additionally, SMARS has a wide coverage as it achieves a detection rate above 90% when the subject is 8 m away and 88.7% and 65% at 9 and 10 m, respectively. SMARS is a promising system for remote sleep monitoring. The system provided a good estimation of sleep stages compared with PSG based on the results. SMARS has served as a benchmark of comparison for other sleep monitoring devices such as Wi-Fi-Sleep by Yu et al. [15].Inspired by recent advancements in Wi-Fi-based sensing, Yu et al. [15] presented a system to monitor and classify sleep. Wi-Fi-Sleep is based on Wi-Fi transceivers and a deep learning method. The system extracts accurate respiration and body movement and can classify sleep into four stages. The Channel State Information ratio was used to eliminate blind spots for improved detection. The effectiveness of the system was evaluated by experimenting with 12 subjects over 19 nights, a process in which it achieved an accuracy of 81.8% for four-stage sleep classification. The ground truth was obtained from PSG and the performance was compared with SMARS and RF-Sleep. The accuracy of the four-stage sleep classification for the other two devices is 79.8% and 69.4%, respectively.In the literature, it can be noted that there is an increasing interest in sleep monitoring (in particular, sleep cycles) using unobtrusive sensors. Due to the unsuitability of PSG for in-home monitoring, researchers have developed and are developing various unobtrusive systems as alternatives, leveraging on recent technological advancements. Generally, proposed sleep monitoring methods are based on one or a combination of the following: respiratory cycle, cardiac cycle, body movement [34]. We also observed that the majority of the proposed systems are limited to three-stage sleep classification [34]. Although the results from the abovementioned studies are encouraging in terms of accuracy, commercial devices cannot produce identical results to PSG. It makes sense because EEG-based systems are the most accurate for detecting all the stages of sleep [35].That said, the ease and relative performance of actigraphy-based devices for sleep and sleep-cycle monitoring has given rise to numerous wearable devices and smartphone-based technologies. Actigraphy devices enable the user to wear dedicated sensors to help track vital signs and movements while sleeping. This review shows that the use of wearable devices for sleep-cycle monitoring is feasible but inaccurate compared to the gold standard PSG [30,36,37]. This is because even in healthy adults accelerometry has high sensitivity but low specificity for sleep detection. These devices often tend to underestimate or overestimate some key parameters such as TST, sleep efficiency, wake, or the transition between the sleep stages [30,36,37]. Patients with sleep disorders, or those who are chronically sleep-deprived, are more likely to suffer from fragmented sleep and reduced ability to understand their functional impairment. Therefore, wearing sleep trackers with incorrect readings could have adverse effects on these patients. This happens because most patients do not realize that the claims of these devices typically outweigh the science to support them as devices to measure and improve sleep. As a result, the importance of precise measurements cannot be overstated [36].A recent study by Chinoy et al. [38] has shown that off-the-shelf sleep trackers (i.e., Fatigue Science Readiband, Fitbit Alta HR, EarlySense Live, ResMed S+, SleepScore Max) provided mixed results for sleep stage classification and the trackers tended to perform worse on nights with poorer/disrupted sleep. Similarly, Roomkham et al. [39] have come to the same conclusion that further studies are needed to assess the longer-term performance of sleep trackers, namely, the Apple Watch in natural conditions, and against PSG in clinical settings. Furthermore, Kholghi et al. [40] concluded that EMFIT QS failed to distinguish sleep stages against PSG and additional development is needed before using EMIFT QS in clinical settings. Moreover, studies have shown that although smartphone-based sensing systems are simpler and less expensive, they correlate poorly with the PSG [41].Frankly, it is impractical to compare or generalize the accuracy across sleep trackers, specifically for under-mattress sleep trackers. This inconsistency occurs because the morphological characteristics of acquired BCG signals are device dependent. Besides, the signals can be different between and within subjects [37,42]. As a result, there is a need for a comprehensive and open dataset of BCG signals that will enable researchers to utilize them in their environments and improve the field into an accepted technique suitable for clinical studies [34]. To date, there is only one publicly available dataset of BCG signals; the purpose of the dataset was to assess the ability of the BCG to monitor changes in cardiovascular function [43].Inventors have proposed, produced, and presented several methods (models and devices) for sleep monitoring by acquiring physiological data unobtrusively. However, the efficacy of a few systems was clinically validated. Experiment-wise, most of the studies are limited to a small sample of healthy individuals [39]. Thus, a broader scope of participants should be taken into consideration during future proposals and assessments of sleep-cycle tracking systems. This is because factors such as gender, age, profession, and social class affect the quality of sleep (Cappuccio et al. [44]).Despite the above criticism, commercial sleep trackers can provide continuous and long-term monitoring of patients’ sleep quality for days and weeks, which is impossible in hospitals. In other words, they can be used as predictive screening methods before performing the sleep studies [34]. For example, Sadek et al. [45] have shown the efficiency of an under-mattress sleep tracker for long-term monitoring of specific sleep parameters, namely, wake-up time, bedtime, and time in bed. These parameters were trended over time, and the authors were able to detect anomalies and notify corresponding caregivers.Typically, under-mattress-based sensors can monitor the sleep quality of patients without interfering with their daily activities. However, this may not always be the case for wearable sensors considering vulnerable populations with behavioral symptoms. To explain, if the sensor is not waterproof, it has to be removed before showering. In addition, if the sensor has a short battery life, it needs to be removed frequently for charging. These situations will undoubtedly distract patients and similarly disrupt the data collection [46,47]. The choice between wearable and non-wearable sensors should be based on the medical conditions of each patient group. Hence, there will always be a trade-off between data continuity and patient comfort [46]. Following our discussion, we conclude the paper in the next section.This review gives an overview of the current state and performance of sleep-cycle monitoring using contactless sensors. The review features the importance of sleep, a discussion of sleep monitoring and polysomnography, a review of existing works in sleep cycle monitoring, a discussion of the takes from the review, and highlights potential concepts that could be explored. With the rising interest in sleep monitoring generally and the clinical need for sleep cycle monitoring, there is an opportunity for researchers and commercial organizations to produce systems that will provide reliable and valid sleep information. Sleep monitoring is a very critical medical issue that could avert negative consequences on the life of individuals. It could potentially reduce the volume of fatigue-related work injuries, health issues, underperformance, road accidents, and aid health workers in managing sleep disorder patients. The performance and features of the systems examined in this review are encouraging. They could be set up for remote sleep-cycle monitoring and long-term studies, and they are easy to use. Unlike the gold standard-PSG, they are unobtrusive and contactless.Conceptualization, B.A. and I.S.; writing—original draft preparation, J.S.; writing—review and editing, I.S.; supervision, B.A. and I.S.; project administration, B.A. All authors have read and agreed to the published version of the manuscript.This research received no external funding.Not applicable.Not applicable.Not applicable.The authors declare no conflict of interest.The keywords and phrases used to retrieve the publications related to contactless monitoring of sleep cycles.Sleep;Sleep cycle;Sleep monitoring;Sleep cycle monitoring;Sleep monitoring system;Sleep cycle monitoring using contactless sensors;Sleep stage monitoring;Automatic sleep stage classification;Automatic sleep stage detection using contactless sensors;Sleep stage monitoring based on heart rate;Contactless sleep monitoring;Sleep stage monitoring based on cardiac cycle;Sleep stage monitoring based on the respiratory cycle;Sleep stage monitoring based on physiological factors;Polysomnography;Ballistocardiography;Non-contact sleep monitoring;Sleep and wearable devices;Bed sensor system;Sleep-cycle monitoring based on physiological factors using contactless sensors.An illustration showing the several sensors attached to a monitored individual during an overnight PSG study.An illustration showing the positing a contactless sensor under the mattress of a monitored individual.Brief description of the literature covered in the review.Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Med-MDPI/biomolecules/biomolecules-01-01-00001.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).I am pleased to introduce Biomolecules, a new journal to report on all aspects of science that focuses on biologically derived substances, from small molecules to complex polymers. Some examples are lipids, carbohydrates, vitamins, hormones, amino acids, nucleotides, peptides, RNA and polysaccharides, but this list is far from exhaustive. Research on biomolecules encompasses multiple fascinating questions. How are biomolecules synthesized and modified? What are their structures and interactions with other biomolecules? How do biomolecules function in biological processes, at the level of organelles, cells, organs, organisms, or even ecosystems? How do biomolecules affect either the organism that produces them or other organisms of the same or different species? How are biomolecules shaped by evolution, and how in turn do they affect cellular phenotypes? What is the systems-level contribution of biomolecules to biological function?The scope of Biomolecules is broad and multidisciplinary, covering biochemical, molecular, cell biological, genetic, physiological, and computational approaches to name a few. I anticipate that this journal will foster fruitful crosstalk between the various disciplines and approaches applied to biomolecule research. We will publish any manuscript of high scientific quality that pertains to diverse aspects relevant to biogenic substances, irrespective of biological question or methodology.To kick-start Biomolecules, we will also publish ambitious series of special issues that cover selected topics of current interest and relevance, including both reviews and original research. Example topics are non-coding RNAs, DNA damage responses, protein folding, sumoylation, and glycoproteins. We welcome suggestions for additional topics. These special issues will be edited by respected leaders in the specific fields to increase the profile and visibility of the papers.The open access format of Biomolecules will provide effective and unrestricted dissemination of the papers to a wide readership. This format will help to realize the ambition of the journal to promote stimulating research for readers with multiple backgrounds and perspectives. We aim at competent, fair peer review and rapid publication to make it attractive for prospective authors to submit to Biomolecules.On behalf of the editorial office and board, I welcome all authors and reviewers and thank them for all their valuable contributions to this exciting new journal. Together we can develop Biomolecules into a respected venue for the fast and cost-effective publication of quality research from diverse scientists across the globe. We look very much forward to working with you all.

Med-MDPI/biomolecules/biomolecules-01-01-00003.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Intregins are heterodimeric α- and β-subunit containing membrane receptor proteins which serve various cell adhesion roles in tissue repair, hemostasis, immune response, embryogenesis and metastasis. At least 18 α- (ITA or ITGA) and 8 β-integrin subunits (ITB or ITGB) are encoded on mammalian genomes. Comparative ITB amino acid sequences and protein structures and ITB gene locations were examined using data from several vertebrate genome projects. Vertebrate ITB genes usually contained 13–16 coding exons and encoded protein subunits with ∼800 amino acids, whereas vertebrate ITB4 genes contained 36-39 coding exons and encoded larger proteins with ∼1800 amino acids. The ITB sequences exhibited several conserved domains including signal peptide, extracellular β-integrin, β-tail domain and integrin β-cytoplasmic domains. Sequence alignments of the integrin β-cytoplasmic domains revealed highly conserved regions possibly for performing essential functions and its maintenance during vertebrate evolution. With the exception of the human ITB8 sequence, the other ITB sequences shared a predicted 19 residue α-helix for this region. Potential sites for regulating human ITB gene expression were identified which included CpG islands, transcription factor binding sites and microRNA binding sites within the 3′-UTR of human ITB genes. Phylogenetic analyses examined the relationships of vertebrate beta-integrin genes which were consistent with four major groups: 1: ITB1, ITB2, ITB7; 2: ITB3, ITB5, ITB6; 3: ITB4; and 4: ITB8 and a common evolutionary origin from an ancestral gene, prior to the appearance of fish during vertebrate evolution. The phylogenetic analyses revealed that ITB4 is the most likely primordial form of the vertebrate β integrin subunit encoding genes, that is the only β subunit expressed as a constituent of the sole integrin receptor ‘α6β4’ in the hemidesmosomes of unicellular organisms.Cell surface integrin receptors regulate cell-cell and cell-extra cellular matrix (ECM) interactions and are involved in mediating all known basic cellular processes (proliferation, migration, differentiation and death) in the body. Precise regulations of these cellular processes by a wide range of integrin receptors are witnessed in cells during development and later in life [1,2,3,4,5,6,7]. Disturbance of integrin function/s lead to suboptimal organogenesis in rodent animal models [5,6,7,8] and disease states in human populations [9,10].An integrin receptor is a heterodimer consisting of an α and a β subunit, each containing extracellular, transmembrane and cytosolic domains. The extracellular domains of receptor subunits bind with the ECM proteins (such as fibronectin, laminin and collagen) and the cytosolic domains of β subunits interact with kinases (focal adhesion kinase and Src kinase), adaptor molecules (such as talin and kindlin) and the cytoskeleton (actin and microtubules) [5,11,12]. These interactions facilitate the ‘outside-in’ and the ‘inside-out’ signaling across the cell membrane by the integrin heterodimers [12,13,14].While the evolutionary path of integrins in development and maintenance of cellular process are intensive areas of investigation [15,16,17], the evolution of different integrin subunits encoded within the vertebrate genomes remains to be fully elucidated. This knowledge is necessary for understanding the integrin receptors and the evolution of cellular functions that are coordinated by these versatile receptors. Evolution of integrin genes dates back to the time of transition of unicellular life forms into multicellular organisms [18,19]. It is known that the integrin-mediated adhesion system existed in the single celled Amastigomonas (Phylum Apusozoa), possibly for the purpose of attachment with the basal lamina, a transition towards sedentary life and multicellularity [19]. In vertebrates, a phylum that includes ∼53,000 species, the genes coding integrins are identified as early as in fishes (Actinopterygians) that evolved about 450 millions of years ago (Mya) [20]. Here we report the gene structures and amino acid sequences for vertebrate β-integrin encoding genes (ITB) and proteins (ITB), respectively, as well as their phylogenetic and evolutionary relationships. Potential regulatory sites for several human ITB genes, predicted secondary structures of signal peptides and cytoplasmic domains and tissue specific expression for mammalian ITB genes are also discussed in terms of their homology and evolution.Table 1 summarizes the locations and predicted structures for vertebrate ITB genes based upon BLAT interrogations of several vertebrate genomes using the reported sequences for human ITB1 [21,22,23,24], ITB2 [25,26,27]; ITB3 [28,29,30,31]; ITB4 [32,33,34]; ITB5 [35,36,37]; ITB6 [38,39,40]; ITB7 [41,42]; and ITB8 [43,44] and the University of California Santa Cruz (UCSC) Genome Browser [45]. The predicted vertebrate ITB genes predominantly contained 13–16 coding exons, with the exception of vertebrate ITB4 genes which exhibited 36 (opossum ITB8) to 39 coding exons and encoded larger ITB protein subunits (∼1,800 amino acids) as compared with other ITB subunits which contained ∼800 amino acids in sequence (Table 1). ITB genes were separately located on vertebrate chromosomes for each of the genomes examined in comparison with other gene families which may be clustered on a single chromosome (e.g., the alcohol dehydrogenase (ADH) gene family) [46] or a small number of chromosomes such as the lactate dehydrogenase (LDH) gene family [47].Application of the SignalP 3.0 server predicted a standard length of signal peptides for ITB1 (20 aa), ITB2 (22 aa), ITB3 (26 aa), ITB4 (27 aa), ITB5 (24 aa), ITB6 (21 aa) and ITB7 (19 aa) subunits except for a long signal sequence (42 aa) for the smallest size integrin isoform ITB8. Although, the server provided a distinct cleavage site for signal peptides in all human ITB forms, a recent study has shown that the signal peptide of the β2 integrin subunit in ruminants containing cleavage inhibition glutamine (Q) was not processed [48]. Not much is known about the signal peptide processing of integrins; however, human ITB1 and ITB2 integrin subunits contain cleavage inhibiting ‘Q’ at the predicted cleavage sites (data not shown). Domain annotation of signal sequences of ITB genes predicted a central helical domain with an anterior and a posterior coiled motifs in ITB1, ITB2, ITB3, ITB5, ITB6 and ITB7 subunits. The signal peptide for ITB8, however, consisted of two central helical motifs separated by a coiled motif and two additional coiled motifs at the N and C terminal ends of the signal peptide. The predicted signal peptide sequences from different ITB subunits showed little evidence of sequence similarity (data not shown) that is not uncommon for signal peptides [49]. The lack of identity amongst the primary structures of signal sequences of different ITB subunits and the similarity amongst the secondary structures (a central hydrophobic core with coiled motifs at the ends) implies that these secondary structures of signal sequences are indispensible conformations for the insertion of the N-terminal ends of beta-subunits into the cell membrane. The reason for the very long signal sequence and two hydrophobic motifs in the ITB8 structure is unclear although it is possible that an additional hydrophobic motif may enhance the processing and translocation of ITB8 into the lipid bilayer [50,51].Figure 1 illustrates the predicted domain structures for ITB2 and ITB4, with the former representing the domain structures for ITB1, ITB3, ITB5, ITB6, ITB7 and ITB8 [52] including the N-signal peptide previously described (residues 1–22 for ITB2); an extracellular integrin beta region (pfam00362) (residues 32–447) including a potential cell attachment site (residues 397–399) and a region of cysteine-rich tandem repeats (residues 414–617); an integrin beta tail domain (pfam07965) (residues 622–700); a transmembrane helical region (residues 701–723) (see Figure 1 for ITB2 TMHMM region), which anchors ITB2 to the cell membrane; and an ITB2 cytoplasmic region (residues 724–768).Vertebrate beta integrin and nematode beta integrin-like genes and proteins. RefSeq: the reference amino acid sequence; 1,3 predicted Ensembl amino acid sequence; 2 not available; 4 Contig refers to a DNA scaffold for sequencing analyses; GenBank IDs are derived NCBI http://www.ncbi.nlm.nih.gov/genbank/; Ensembl ID was derived from Ensembl genome database http://www.ensembl.org/; UNIPROT refers to UniprotKB/Swiss-Prot IDs for individual acid lipases (see http://kr.expasy.org/); bps refers to base pairs of nucleotide sequences; pI refers to theoretical isoelectric points; the number of coding exons are listed.Predicted domains and transmembrane helix for human ITB2 and ITB4. Domains and key regions are identified for ITB2 and ITB4 amino acid sequences using NCBI web tools (http://www.ncbi.nlm.nih.gov/Structure/cdd/wrpsb.cgi) to identify functional domains and ExPasy web tools to identify predicted transmembrane domains (TMHMM) (http://www.cbs.dtu.dk/services/TMHMM-2.0/); INB or integrin_beta; integrin B tail (in pink); TMHMM transmembranes (in red); cytosolic domain (in blue); Calx-beta domain (in green); FN3 (fibronectin 3), cytokine receptor and interdomain contacts (red triangles); Note: lack of FN3 binding domains and the cytokine receptor motifs in β4 subunit that interacts only with laminin-332.In contrast, ITB4 is a much larger protein compared with the other beta-integrins (1822 residues compared with 769–798 residues) and contains several beta-integrin-like domains in the N-terminal half of the protein [12,53,54] including the N-signal peptide (residues 1–27); an N-terminal extracellular beta-integrin domain (residues 37–453); a cysteine-rich tandem-repeat region (residues 456–619); an integrin beta-tail region (residues 626–711); and an ITB4 transmembrane-helix anchor region (residues 711–733) (Figure 1). The cytoplasmic region of ITB4 contains four fibronectin type III (FN3) domains (residues 1128–1215; 1220–1313; 1528–1628; and 1641–1734) and a Calx-beta motif (residues 991–1054) which are responsible for most of the intracellular interactions of the integrin [53,54,55,56]. Other key cytoplasmic ITB4 domains include several interdomain contact sites (residues 1641, 1708 and 1723) and cytokine-receptor motifs (1608–1609 and 1611–1612).The cytoplasmic domains of β subunits interact with several intracellular proteins. Many of these interactions are known to cause conformational change in the extracellular domain changing the affinity of the receptor with the ECM. The interactions of extracellular domains with the ECM may also cause change in the conformation of the cytosolic domain allowing its interaction with the non-receptor tyrosine kinases and the actin cytoskeleton. Therefore, the cytoplasmic domain of integrin β subunits plays crucial roles in both ‘outside-in’ and ‘inside-out’signaling [12,13,14]. Figure 2 examines alignments of vertebrate ITB1 cytosolic domain sequences which are color coded for amino acid residue properties. With the exception of a second duplicated ITB1.1 (designated as ITB1B) gene product observed in zebrafish (Danio rerio), identical sequences were observed for the ITB1 cytosolic domain for all vertebrates examined, which indicates that this is a highly conserved region of ITB1 which undertakes essential functions and is subject to selection and maintenance of this sequence. Comparisons of the cytosolic domain sequences for the other ITB proteins (alignments not shown) revealed lower levels of amino acid sequence identities as compared with the highly conserved ITB1 cytosolic domain sequence: ITB2 (37% identities); ITB3 (77% excluding the gene duplicate product ITB3B from zebrafish); ITB5 (50%); ITB6 (74%); ITB7 (35%); and ITB8 (45%).Amino acid alignments for vertebrate ITB1 cytosolic domain sequences. ITB1 sequences examined included Hu-human; Rh-rhesus; Ma-marmoset; Mo-mouse; Ra-rat; Gp-guinea pig; Ho-horse; Co-cow; Pi-pig; Op-opossum; Ch-chicken; Fr-Xenopus tropicalis; Zf-zebrafish; see Table 1 for details; note that 2 ITB1-like genes were observed in zebrafish (designated as ITB1A and ITB1B); * shows identical residues for ITB subunits; : similar alternate residues;. dissimilar alternate residues; α-helix for vertebrate ITB sequences is in shaded yellow; β-sheet is in shaded grey; colors for amino acids are shown as: basic (R and K); acidic (D and E); neutral hydrophilic (G, Y, Q, S, T, N, Y, C, H); and hydrophobic (M, A, F, I, L, W, P, V); the Cyto-1, Cyto-2 (NPXY) and Cyto-3 (NXXY) domains are shown in dotted lines (see text for details).Figure 3A shows amino acid sequence alignments for the six major human ITB1 isoforms designated as ITB1a-ITB1f [57]. Residues 1–26 were identical for each of the isoforms which contained the 19 residue α-helix region, whereas the C-terminal differed in length and sequence and exhibited 1–2 predicted β-sheet regions. Recent studies [58] have shown that ITB1a is expressed in fetal muscles but is substituted by ITB1d during postnatal development. The C-terminal region is exposed at the cytoplasmic face of the plasma membrane where it is bound to the actin filaments. ITB1d is expressed only in striated muscle tissues and binds to both cytoskeletal and extracellular matrix proteins with an affinity higher than ITB1a which provides a stronger link between the cytoskeleton and extracellular matrix to support mechanical tension during muscle contraction. ITA1a and ITA1b have been shown to be similar as far as the alpha/beta association and fibronectin binding are concerned but differ, however, in their subcellular localization. ITB1a has been localized in focal adhesions whereas ITBb does not and exhibits distinct properties [22]. Human ITB1 isoforms are differentially expressed in tissues and exhibit distinct binding properties. HumanITB1a is widely expressed and usually coexpressed with other isoforms with a more restricted distribution. ITB1b is expressed in skin, liver, skeletal muscle, cardiac muscle, placenta, umbilical vein endothelial cells, neuroblastoma cells, lymphoma cells, hepatoma cells and astrocytoma cells. ITB1c is expressed in muscle, kidney, liver, placenta, cervical epithelium, umbilical vein endothelial cells, fibroblast cells, embryonic kidney cells, platelets and several blood cell lines, whereas ITB1d is expressed specifically in striated muscle (skeletal and cardiac muscle).Amino acid sequence alignments for vertebrate ITB cytosolic domain sequences. (A) Comparison and alignments of human ITB1 major isoforms for cytosolic domain sequences; (B) Consensus sequences of vertebrate ITB cytosolic domains; see Table 1 for sources of beta integrin cytosolic domain sequences: * shows identical residues for ITB subunits; : similar alternate residues;. dissimilar alternate residues; α-helix for vertebrate ITB sequences is in shaded yellow; β-sheet is in shaded grey; colors for amino acids are shown as: basic (R and K); acidic (D and E); neutral hydrophilic (G, Y, Q, S, T, N, Y, C, H); and hydrophobic (M, A, F, I, L, W, P, V); the Cyto-1, Cyto-2 (NPXY) and Cyto-3 (NXXY) domains are shown in dotted lines (see text for details).Amino acid sequence alignments for vertebrate consensus sequences ITB cytosolic domains are shown in Figure 3B. Human ITB sequences were based on previous reports for ITB1 [21,22,23,24]; ITB2 [25,26,27]; ITB3 [28,29,30,31]; ITB5 [35,36,37]; ITB6 [38,40]; ITB7 [39,41,42]; and ITB8 [43,44]. With the exception of the human ITB8 sequence, the other ITB sequences were of similar length (46-58 amino acids) and shared a predicted 19 residue α-helix region (residues 2–20 for human ITB1) whereas the human ITB8 sequence contained 65 amino acids with 6 predicted β-sheets.The cytoplasmic tail of the ITB4 subunit is exceptionally long (1072 residues) compared to other ITB subunits [33] that are much shorter (Figure 3). Point mutation analysis of the cytoplasmic sequences of these β integrin subunits reveal three clusters of amino acids in the β cytoplasmic tail that regulate the interaction of integrins with the cytoskeleton, localization of receptors at the adhesion complex and inside-out signaling [12,13,59,60,61,62]. These three clusters of amino acids (Signalins) are commonly known as cyto 1, cyto-2 and cyto-3: cyto-1 is present in the vicinity of transmembrane domain, whereas cyto-2 (NPXY motif) and cyto-3 (NXXY motif) are located in the proximal and distal regions respectively of a tail (Figure 2) [63]. Alignment results of ITB1 subunits of different species (Figure 2) and spliced versions of ITB1 subunits (Figure 3A) show that the cyto-1 residues remain highly conserved indicating their conservation during vertebrate evolution for their specificity in function. Recent studies have shown that the interaction between the conserved arginine residue in the α-tail and aspartate residue in the β-tail, and by the hydrophobic residues immediately N-terminal to the arginine and aspartate residues play important role in ‘inside-out’ signalling by forming a ‘clasp’ between the α and β subunits [64,65].The cyto-3 sequence in contrary varied amongst the spliced versions of ITB1 and the different ITB subunits (Figure 3). Therefore, the variability in the functions of different spliced versions of ITB1 (Figure 3A) and amongst different ITB isoforms (Figure 3B) may be derived from the differences in the cyto 2 and cyto 3 sequences. Moreover, each β subunit conceals distinct differences in its affinity towards intracellular proteins that is shown to be dictated by the ‘X’ and the neighboring amino acids of these motifs [66]. For instance, the binding of ICAP-1α, a 200amino acid protein, with the cyto-3 is influenced by the proximal Val787 and Val 790 [66,67,68]. The NPXY and NXXY motifs, with the propensity to form β turns, act as canonical recognition sequences of intracellular proteins with phosphotyrosine-binding domains (PTB) [66]. These include the interaction of β1A tail with the PTB domain of talin, EPS8 and Dab1; β2 tail with Dok-1 and talin; β3 tail with Numb, Dab1, EPS8, Tensin, Dok-1 and talin; β5 tail with Numb, Dab1, Dab2, EPS8, Tensin, Dok-1 and talin and the β7 tail with tensin, Dok-1 and Talin [67].Recent studies have reported that cytosolic proteins kindlin-1, 2 and 3 are essential for integrin activation [68,69]. Immuno-precipitation assays with β integrin tails show that isoforms of kindlins bind with membrane proximal NPXY and membrane distal NXXY motifs as well as neighboring residues (NPXY linker region) of the β integrin subunit [69]. Several other cytosolic proteins (including filamin, melusin and myosin) also bind both conserved and non-conserved domains of β cytoplasmic tails [70]. Therefore, differences in the residues within and around these motifs in vertebrate β integrin subunits may change the affinities of these cytosolic proteins with the β integrin tail. Phosphorylation of the tyrosine residue in the distal NXXY motif of the β3 subunit disrupts the recognition by kindlin-2 and co-activation of aIIb.b3 integrin by talin [71,72]. The phosphoryalated or unphosphorylated state of tyrosines may also determine the affinities of proteins with the cytosolic tail. The unphosphorylated state of Y747 in the of β3 integrin tail has a 3 fold preference for the talin over the PTB domain of Dok1, whereas with the phosphorylated state of Y747, this affinity is increased 400 fold for Dok1 and decreased 2 fold for talin [73]. A recent study shows that phosphorylation of tyrosine 759 inhibits binding of kindling-2 with the C-terminal β3 chain [71]. Thus the expression patterns of different β subunits (Figure 5) and interacting proteins in the cytosol as well the phosphorylation state of ‘Y’ may determine the functional output of the integrin receptors.Figure 4 shows the predicted structures of mRNAs for human ITB transcripts for the major isoform in each case [57]. The transcripts were 3.0–9.2 kbs in length and exhibited distinct exonic structures in each case, including extended 3′-untranslated regions (UTR), especially for ITB3a, ITB6a and ITB8a transcripts. The number of ITB introns varied widely among the vertebrate genes examined: the ITB4 gene contained the largest number of introns (39) followed by ITB1 (16), ITB2 and ITB7 (15), ITB3, ITB5 and ITB6 (14) and ITB8 (13).Structures and major splicing isoforms for human beta integrin genes. Derived from the AceView website http://www.ncbi.nlm.nih.gov/IEB/Research/Acembly/ [57]; mature isoform variants (a) are shown with capped 5′- and 3′- ends for the predicted mRNA sequences; NM refers to the NCBI reference sequence; exons are in pink; the directions for transcription are shown as 5′ → 3′; sizes of mRNA sequences are shown in kilobases (kb).The human ITB genome sequences contained several predicted transcription factor binding sites (TFBS), microRNA sites located in the 3′-untranslated region and CpG islands, which included CpG158, CpG92, CpG91, CpG152 and CpG133 located in the 5′-untranslated region of human ITB1, ITB3, ITB4, ITB5 and ITB8, respectively (see Table 2). These CpG islands within the ITB gene promoters may play major contributing roles in maintaining high levels of gene expression (1.4–6.1 times the average for human genes) [57] which are similar to CpG islands within housekeeping gene promoters expressed in most tissues [74]. Large numbers of TFBS sites were observed for most of the human ITB genes examined, including 51, 56 and 105 such sites for ITB4, ITB6 and ITB8, respectively. Of particular significance for the human ITB1 and ITB3 gene promoters is the transcription factor, HoxD3, that binds directly to these promoters and assists in regulating the expression of integrins α5β1 and αVβ3 during angiogenesis [75]; the PPARα (peroxisome proliferator-activated receptor-α) that regulates gene expression in vascular cells and inhibits TGF (transforming growth factor)- β-induced ITB5 transcription [76] and Hox A10, that directs the regulation of the ITB3 gene in human endometrial cells and regulates transcription of ITB3 during myeloid differentiation [77,78]. Moreover, the genes encoding the integrin subunits β7, β3, β6 and β8 map to 12q13.13, 17q21.32, 2q23-q31 and 7p15-p21 positions respectively which are close to HOXC, HOXB, HOXD and HOXA genes suggesting a common divergence of these genes during vertebrate evolution [79].Predicted transcription factor binding sites (TFBS), CpG islands and MiRNA (MiR) regions for human and mouse ITB Genes. The human and mouse genome browsers (http://genome.ucsc.edu) [45] were used to examine the predicted transcription factor binding sites (TFBS), CpG islands and Mi-RNA binding sites for human and mouse ITB genes.Several microRNA (miRNA) binding sites within the 3′-untranslated region (3′-UTR) of human ITB mRNA were also identified (Table 2). These microRNA species are phylogenetically conserved and regulate mRNA and protein expression during embryonic development [80,81]. MiRNA 183, for example, inhibits tumor invasiveness and participates in the development and function of neurosensory organs by targeting the ITB1 (mRNA) gene [82] whereas ITB3 gene expression is apparently regulated by miRNA let-7a in malignant melanoma [83]. Over-expression of mir-124 attenuates endogenous ITB1 expression in oral squamous cell carcinomas [84] while microRNA miR-93 promotes tumor growth and angiogenesis by decreasing ITB8 transcripts [85]. The number of microRNAs that target the 3′ UTR of human ITB transcripts (4 for ITB1, 44 for ITB2, 5 for ITB3, 1 for ITB4, 53 for ITB5, 11 for ITB6, 30 for ITB7 and 2 for ITB8) varies widely among the human ITB genes examined. The absence of redundancy among this wide range of microRNA species regulating the levels of human integrin subunits suggests that the evolution of the C-terminal non-coding regions of these subunits followed a divergent path for the purpose of regulating the levels of expressions of each ITB subunits in different cells. The regulation of ITB4 by a single microRNA further suggests that the expression of this subunit is not intensely regulated at the post-transcriptional stage in comparison with the other human ITB genes.Brendle and coworkers [86] have also examined single nucleotide polymorphisms (SNPs) in predicted miRNA sites for several ITA and ITB genes and the potential association of these SNPs with breast cancer risk (BCR) and reported a potential BCR marker for one of the ITB4 miRNA binding sites. A likely mechanism for mi-RNA translational regulation has been recently reported [87]. MicroRNAs have been shown to be transcribed as long primary-miRNAs (pri-miRNAs) in the nucleus and processed in the cytoplasm into 19-22 bp mature mi-RNAs which anneal to the 3′-UTR of target mRNAs to promote degradation or translational repression [88]. Moreover, considerable flexibility has been reported for mi-RNAs which are capable of targeting hundreds of genes while individual 3′-UTR mi-RNA regions may be a target for several distinct mi-RNAs [89,90]. The miRNA sequences within the 3′-UTR of human ITB genes are therefore likely to play a major role in regulating the translation of these genes within vertebrate tissues.Figure 5 presents ‘heat maps’ showing comparative Itb gene expression for various mouse tissues obtained from GNF Expression Atlas Data using GNF1M chips (http://genome.ucsc.edu; http://biogps.gnf.org) [91]. These data supported a broad and high level tissue expression for mouse Itb7, including during early embryonic development. A very high level of expression for Itb2 and Itb7 in bone marrow, spleen and lymphocytes are consistent with their involvement in forming the integrin receptors in blood cells [92]. The Itb4 expression was highest in epidermal tissues and is consistent with its presence in the hemidesmosomes of these epithelial cells [93,94]. It may be noted that ITB4 pairs only with the α6 subunit forming a laminin-binding receptor providing stable adhesion of epithelial cells with the basement membrane [95,96]. Other comparisons of mouse Itb tissue expression indicated significant differences, including higher levels of Itb3 expression in bone but with lower expression levels for Itb8 in most tissues examined. Overall, mouse Itb tissue expressions levels were up to 3.7 times the average level of gene expression [57] which supported key roles played by these membrane receptor proteins which serve various cell adhesion roles in tissue repair, hemostasis, immune response, embryogenesis and metastasis [92]. Similar tissue distribution profiles for ITB gene expression were observed for human tissues, including an overall high level gene expression ranging from 0.8–6.1 times the average level of human gene expression (Table 2).Comparative tissue expression for mouse beta integrin genes (ITB). Expression ‘heat maps’ (GNF Expression Atlas 2 data) (http://biogps.gnf.org) [91] were examined for comparative gene expression levels among human and mouse tissues for ITB genes showing high (red); intermediate (black); and low (green) expression levels; derived from mouse genome browsers (http://genome.ucsc.edu) [45].A phylogenetic tree (Figure 6) was calculated by the progressive alignment of 61 vertebrate ITB amino acid sequences with vertebrate ITB1-8 sequences which was rooted with the Caenorhabitis elegans (nematode) ITB-like sequence (see Table 1). The phylogram showed clustering of the ITB sequences into groups which were consistent with their evolutionary relatedness, as well as groups for each of vertebrate ITB1–ITB8 which were distinct from the nematode ITB-like sequence. These groups were significantly different from each other (with bootstrap values of >90) and showed closer relatedness for the following ITB gene groupings: group 1: ITB1-ITB2-ITB7; group 2: vertebrate ITB4 with the elegans ITB-like sequence (PAT3); group 3: ITB3-ITB5-ITB6; and group 4: ITB8, which is the most distinct group in terms of its relatedness to other ITB gene families. It is apparent from this study of vertebrate ITB genes and proteins that these are ancient proteins for which a proposed common ancestor for the ITB genes may have predated the appearance of fish >500 millions of years ago [96].Among the ITB integrin genes examined, the ITB4 integrin subunit gene related most closely with the C. elegans (nematode) PAT3 sequence indicating that it may represent the primordial vertebrate beta integrin gene and the first to appear in the vertebrate ancestor. The ITB4 differs from other ITB subunits. It is unusually longer (1778 residues) compared with other integrin β subunits and contains a long amino-terminal (683 aa) and cytosolic (1072 aa) domains [33]. The extracellular domains of β4 subunit showed low identity (∼35%) with other β integrin subunits. Moreover, the transmembrane domain of the ITB4 subunit is poorly conserved and is exceptionally long [92,97,98].Phylogenetic tree of vertebrate beta integrin cytosolic domain amino acid sequences. The tree is labeled with the ITB name and the name of the animal and is ‘rooted’ with the Caenorhabitis elegans (nematode) ITB-like sequence (see Table 1). Note the 7 major clusters corresponding to the ITB1, ITB2, ITB3, ITB4, ITB5, ITB6, ITB7 and ITB8 gene families. A genetic distance scale is shown (% amino acid substitutions). The number of times a clade (sequences common to a node or branch) occurred in the bootstrap replicates are shown. Only replicate values of 90 or more which are highly significant are shown with 100 bootstrap replicates performed in each case.Different integrin beta receptor proteins are known to interact with at least 22 different ligands and matrix proteins [14,92, 99,100,101,102] that are summarized in Table 3. Of these, ITB1 pairs with the largest number of ligands (12 ligands) followed by ITB2 and ITB3 (7 ligands each), ITB5, ITB6 and ITB7 (3 ligands each), ITB8 (2 ligands) and ITB4 (1 ligand). Moreover, ITB1 pairs with the largest number of α subunits (fourteen) followed by ITB2 (four), ITB3 and ITB7 (two), and ITB4, ITB5, ITB6 and ITB8 (single α subunit). The ITB4 subunit pairs only with the ITA6 subunit forming α6β4 as the sole integrin receptor of the hemidesmosomes, a structural component, that is required for the attachment of cells with the basal lamina [103,104,105]. The genes and proteins of hemidesmosomes [107] date back to metazoans/holozoans, suggesting that the attachment of unicellular life forms on the basal lamina via the hemidesmosomes possibly initiated the formation of multicellular organisms with the evolution of other cell-cell junction components (tight, adherent, desmosmal and gap junctions). The tissue specific expression of integrin ITB subunits (Figure 6) showed that mammalian Type I hemidesmosomes are found in the epithelial cells of skin, mouth and esophagus whereas Type II hemidesmosomes are found in the intestinal epithelial cells [93]. This is consistent with the high expression of ITB4 in epithelial cells. While the ITB4 subunit in the α6β4 integrin receptor primarily plays a role in the formation of stable adhesions of epithelial cells with laminin-332, recent studies have suggested an additional role in the migration of keratinocytes and cancer cells [106,108]. Prior to migration, keratinocytes lose their stable adhesion mediated by hemidesmosomes and migrate over collagen and then secrete a provisional matrix of laminin-332 for its motility [109]. The cancer cells also require laminin-332 to migrate [110]. It is now known that the proteolytic cleavage of laminin-332 triggers cell motility of cells via the α6β4 receptor [111]. Other evidence suggests that the migration on laminin-332 is indeed mediated by the α3β1 integrin rather than the α6β4 integrin which actually has transdominating inhibiting effects on migration mediated by the α3β1 integrin [112]. Overall these reports indicate that the ancestral role of integrin in forming stable adhesions of epithelial cells via hemidesmosome might have evolved to support migratory roles of cells by the introduction of additional integrin receptors to perform specialized functions. In this regard, the evolution of the ITB1 subunit from the primordial ITB4 may have played a significant role in influencing cell migration. The transmigration of blood cells across the endothelial layers, a highly specialized function mediated by the integrins, may be associated with the evolution of receptors αvβ3 and those formed by the association of the ITB2 subunit with αL, αM, αX or αD subunits [113,114,115,116,117,118,119].The extracellular domains of both α and β subunits of a receptor interact with wide spectrum of ECM molecules (Table 3) to perform various cellular functions. This suggests that these receptors may have evolved along with the evolution of ECM molecules for performing diverse functions in the context of presence or absence of specific ligands. Consequently, the ITB1 subunit may be the most promiscuous of all of the vertebrate β subunits as it pairs with the largest number of α subunits, and these alpha/beta1 heterodimers also interact with a large number of ligands. This is consistent with the observation that ITB1 like subunits had already diverged in the earlier stages of metazoans (corals and sponges) [120]. Therefore, the clues to the evolution of different vertebrate integrin receptors may lie in their evolution to interact with different ECM molecules. However, in the absence of comprehensive information on the different domains/motifs of the ECM molecules that interact with the specific domains of different integrin receptor, further conclusions may not be derived. Nevertheless, the clues to the evolutionary proximity amongst different β subunits might be found in their ability to pair with common α subunit/s, since these β subunits are likely to preserve domain/s that determine their ability to associate with similar α subunit/s or vice versa [121]. With this notion and based on the overlapping subunit compositions of functional integrin receptors (Figure 7), it is predicted that ITB1 that shares the sole alpha subunit (α6) with ITB4, is the closest to the ancestor ITB4. The ITB1 evolved to pair with the largest number of alpha subunits (Table 3) including α4 that is shared with ITB7, and the αv subunit that is shared with ITB3, ITB5, ITB6 and ITB8. Therefore, the cluster containing ITB3, ITB5, ITB6 and ITB8, and the cluster consisting of ITB7 and ITB2 may have been derived directly from ITB1. The origin of ITB5, ITB6 and ITB8 from ITB3 (rather from the versatile ITB1) is less likely because ITB3 is the most specialized of this cluster and is expressed in both blood cells (platelets) and other cell types such as placental trophoblast and cancer cells [122,123,124]. In contrast, ITB5, ITB6 and ITB8 including ITB1 are not expressed in blood cells. The ITB2 and ITB7 subunits, that constitute solely the integrins of hematopoietic and immune system [125], and specifically ITB2 that does not share an α subunit with other ITB subunits, are likely to be the most specialized ITB subunits. The αLβ2 mediates migration of T-Cells across the endothelium (invasion or transmigration) and the α4β7expressed on memory T cells directs their trafficking to the sites of inflammation.Multiplicity and specificity of ligand binding by ITB subunits; ECM refers to extracellular matrix.An analysis of α subunit sharing by different ITB subunits suggests that evolution of the ITB1 subunit led to the emergence of two groups of ITB subunits, one consisting of ITB3, ITB5, ITB6 and ITB8 subunits and the other consisting of ITB7 and ITB2. This conclusion from the subunit sharing concept (Figure 7) is very similar to our phylogenetic analysis data that suggests that ITB1-ITB7-ITB2 belong to one cluster and the ITB3-ITB5-ITB6 as another cluster. A previous phylogenetic study on ITBs [16] supported ITB1-ITB7-ITB2 as one cluster and the ITB3- ITB5-ITB8 as another cluster. Therefore, two phylogenetic analyses differed by one subunit (ITB6/ITB8) in their second cluster but both found ITB4 either an outlier or an ancestral integrin subunit. The ITB8 is found to be a distinct member of ITBs in our study whereas in the previous study it was found to diverge from the ITB6 earlier in evolution. The subunit pairing concept (Figure 7), however, groups the ITB8 subunit belonging to cluster 2 of both studies together (ITB3-ITB5-ITB6-ITB8) which is consistent with a previous report [15].The blue arrows show the predicted evolutionary paths of β subunits from the ancestral β4 subunit. Pairings of different α and β subunits are shown by thin black lines. This concept shows two lines of evolution diverging from β1, one towards blood cell integrins consisting of β2 or β7 subunits and the other towards a cluster consisting of β5, β6, β8 and β3 subunits that are primarily expressed in tissues other than blood with exception of α IIb.β3 that is expressed also in blood platelets (see text).BLAST (Basic Local Alignment Search Tool) studies were undertaken using web tools from the National Center for Biotechnology Information (NCBI) (http://blast.ncbi.nlm.nih.gov/Blast.cgi) [127]. Protein BLAST analyses used human and mouse ITB amino acid sequences previously described (Table 1). Non-redundant protein sequence databases for several vertebrate genomes were examined using the blastp algorithm, including human (Homo sapiens) [128]; horse (Equus caballus) [129]; mouse (Mus musculus) [130]; opossum (Monodelphis domestica) [131]; chicken (Gallus gallus) [132]; frog (Xenopus tropicalis) (http://genome.jgi-psf.org/Xentr3/Xentr3.home.html); zebrafish (Danio rerio) (http://www.sanger.ac.uk/Projects/D_rerio/); and nematode (Caenorhabditis elegans) (http://genome.ucsc.edu/). This procedure produced multiple BLAST ‘hits’ for each of the protein databases which were individually examined and retained in FASTA format, and a record kept of the sequences for predicted mRNAs and encoded ITB-like proteins. These records were derived from annotated genomic sequences using the gene prediction method: GNOMON and predicted sequences with high similarity scores for human ITB. Predicted ITB-like protein sequences were obtained in each case and subjected to analyses of predicted protein and gene structures.BLAT analyses were subsequently undertaken for each of the predicted ITB amino acid sequences using the UCSC Genome Browser (http://genome.ucsc.edu/cgi-bin/hgBlat) [45] with the default settings to obtain the predicted locations for each of the mammalian ITB genes, including predicted exon boundary locations and gene sizes. Structures for human and mouse isoforms (splicing variants) were obtained using the AceView website to examine predicted gene and protein structures (http://www.ncbi.nlm.nih.gov/IEB/Research/Acembly/index.html?human) [57].FASTA sequence of different human β integrin amino acid sequences were subjected to SignalP 3.0 Server (http://www.cbs.dtu.dk/services/SignalP) [133] to determine the number of amino-acids and the predicted secondary structures in the N-terminal end of the ITGB isoform involved in the formation of the signal peptide. The secondary structures of each signal peptide were determined using a SWISS-MODEL workspace (http://swissmodel.expasy.org) [134].Predicted secondary and tertiary structures for the predicted cytoplasmic domains of vertebrate ITB-like proteins were obtained using the SWISS MODEL web tools [134]. The tertiary structures of the cytoplasmic tails for human ITB1 [residues 1–36], ITB2 [residues 1–47] and ITB7 [residues 1–38] were predicted using a model (PDB: 3g9wC) for human ITB1 [135]; while the reported structure for human ITB3 [136] (PDB:1m8oB) served as the reference for human ITB3 (residues 1–47), ITB5 (residues 1–40) and ITB6 (residues 1–45) tertiary structures, and the human ITB4 structure (PDB: 2yrzA) [137] for human ITB4 (residues 906–1007). Theoretical isoelectric points and molecular weights for vertebrate ITB-like proteins were obtained using Expasy web tools (http://au.expasy.org/tools/pi_tool.html).The UCSC Genome Browser (http://genome.ucsc.edu) [45] was used to examine GNF Expression Atlas 2 data using various expression chips for human ITB genes (http://biogps.gnf.org) [91]. Gene array expression ‘heat maps’ were examined for comparative gene expression levels among human and mouse tissues showing high (red); intermediate (black); and low (green) expression levels.The UCSC Human Genome Browser (http://genome.ucsc.edu) [45] was used to examine the comparative location, number and sequences for human CpG islands, transcription factor binding sites (TFBS) and microRNA sites located in the 3′-untranslated region (UTR) of human ITB genes in association with the TargetScan website (http://www.targetscan.org).Alignments of vertebrate ITB-like and nematode (Caenorhabditis elegans) PAT3 protein sequences were assembled using BioEdit v.5.0.1 and the default settings [137]. Alignment ambiguous regions were excluded prior to phylogenetic analysis yielding alignments of 370 residues for comparisons of vertebrate ITB sequences with the nematode PAT3 (beta-integrin homolog) sequence (Table 1). Evolutionary distances were calculated using the Kimura option [138] in TREECON [139]. Phylogenetic trees were constructed from evolutionary distances using the neighbor-joining method [140] and rooted with the nematode PAT3 sequence. Tree topology was reexamined by the boot-strap method (100 bootstraps were applied) of resampling and only values that were highly significant (≥90) are shown [141].Bioinformatic analyses of the integrin genes and proteins in vertebrates revealed a high degree of diversity in terms of their chromosome locations, alternate splicing, transcriptional and post-transcriptional regulations, and tissue specific expressions. Results suggested that the evolution of integrins within vertebrates followed a divergent path for these genes and protein structures but with common functions specializing towards adhesion, migration and transmigration of cells in succession. Our phylogenetic analysis revealed for the first time that ITB4 (encoding the β4 integrin) is the most likely ancestral form of integrin β-like genes. This subunit has inherited the ancestral role for β-integrins in forming simple adhesions (hemidesmosomes) in vertebrate cells similar to unicellular organism and is also involved in the migration of transformed (cancer) cells [7]. The subunit sharing analysis of ITB subunits reveals that β2 and β7 subunits that are expressed only in the cells of hematopoietic and immune system are possibly the most specialized forms of integrins.This project was supported by a research grant (RR017701) from the Center for Psychiatric Neuroscience, UMMC to UKR. We also acknowledge the expert assistance of Bharet Patel of Griffith University with the phylogeny studies.

Med-MDPI/biomolecules/biomolecules-01-01-00032.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Pro-inflammatory cytokines, such as tumor necrosis factor (TNF)-α, induce the expression of a wide variety of genes, including intercellular adhesion molecule-1 (ICAM-1). Ursolic acid (3β-hydroxy-urs-12-en-28-oic acid) was identified to inhibit the cell-surface ICAM-1 expression induced by pro-inflammatory cytokines in human lung carcinoma A549 cells. Ursolic acid was found to inhibit the TNF-α-induced ICAM-1 protein expression almost completely, whereas the TNF-α-induced ICAM-1 mRNA expression and NF-κB signaling pathway were decreased only partially by ursolic acid. In line with these findings, ursolic acid prevented cellular protein synthesis as well as amino acid uptake, but did not obviously affect nucleoside uptake and the subsequent DNA/RNA syntheses. This inhibitory profile of ursolic acid was similar to that of the Na+/K+-ATPase inhibitor, ouabain, but not the translation inhibitor, cycloheximide. Consistent with this notion, ursolic acid was found to inhibit the catalytic activity of Na+/K+-ATPase. Thus, our present study reveals a novel molecular mechanism in which ursolic acid inhibits Na+/K+-ATPase activity and prevents the TNF-α-induced gene expression by blocking amino acid transport and cellular protein synthesis.Pro-inflammatory cytokines, such as tumor necrosis factor (TNF)-α, induce the expression of a variety of genes essential for inflammatory responses, including intercellular adhesion molecule-1 (ICAM-1; CD54) [1]. In response to pro-inflammatory cytokines, ICAM-1 expression is induced on the surface of vascular endothelial cells and required for tissue infiltration of circulating leukocytes [2]. Thus, ICAM-1 plays a critical role in inflammatory response and its expression and function are regarded as a potential target of therapeutic intervention. ICAM-1 expression is regulated mainly at the transcription level and induced by the transcription factor nuclear factor κB (NF-κB) [3].The NF-κB signaling pathway is induced by various stimuli, including pro-inflammatory cytokines, infectious agents, and chemical/physical stress. The NF-κB dimer is sequestered in the cytosol by associating with the inhibitor of NF-κB (IκB) family of proteins [4]. Upon TNF-α stimulation, TNF receptor 1 triggers the activation of IκB kinase and phosphorylated IκB undergoes ubiquitination and subsequent hydrolysis by the proteasome [5]. The NF-κB dimer then translocates from the cytosol to the nucleus where it activates the transcription of responsive genes that regulate inflammation, proliferation, cell death, cell survival, and angiogenesis [6]. Many types of natural and synthetic small molecules have been reported to block the NF-κB signaling pathway and its downstream gene expression [7].Ursolic acid (3β-hydroxy-urs-12-en-28-oic acid) (Figure 1) is a natural pentacyclic triterpenoid that is often found as a major component of medicinal herbs, foods, and other plants. It has been reported that ursolic acid possesses a wide range of biological activities, such as anti-inflammatory and anti-carcinogenic activities [8]. In a screening for anti-inflammatory agents, we were able to isolate and investigate many plant-derived natural compounds. We found that ursolic acid most effectively inhibits the inducible expression of cell-surface ICAM-1 among ursane-, oleanane-, lupane-, and taraxasterane-type triterpenes purified from Nerium oleander [9,10]. In relation to this finding, it has been thus far reported that ursolic acid inhibits the NF-κB-dependent signaling pathway and gene expression [11,12,13,14,15,16,17,18]. However, in contrast to these reports, we unexpectedly found that ursolic acid inhibits ICAM-1 expression more effectively by targeting intracellular processes downstream of mRNA expression. In this study, the molecular mechanism by which ursolic acid inhibits ICAM-1 expression was investigated.Structure of ursolic acid.Human lung carcinoma A549 cells (JCRB0076) and human fibrosarcoma HT-1080 cells (JCRB9113) were provided by Health Science Research Resources Bank (Osaka, Japan). A549 cells, HT-1080 cells and human breast adenocarcinoma MCF7 cells were maintained in RPMI 1640 medium (Invitrogen, Carlsbad, CA, USA) supplemented with 10% (v/v) heat-inactivated fetal calf serum (JRH Biosciences, Lenexa, KS, USA) and penicillin-streptomycin mixed solution.Ursolic acid was prepared from the leaves of Nerium oleander as described previously [9] or commercially purchased from Sigma-Aldrich (St. Louis, MO, USA). Cycloheximide, MG-132 [Z-Leu-Leu-Leu-H (aldehyde)], and ouabain were obtained from Wako Pure Chemical Industries (Osaka, Japan), Peptide Institute (Osaka, Japan), and Sigma-Aldrich, respectively. Antibodies to β-actin (AC-15; Sigma-Aldrich), c-FLIP (Dave-II; Alexis, Lausen, Switzerland), ICAM-1 (clone 15.2; Leinco Technologies, St. Louis, MO, USA), ICAM-1 (clone 28; BD Biosciences, Franklin Lakes, NJ, USA), IκBα (clone 25; BD Biosciences), p65 (20/NF-kB/p65; BD Biosciences) and horseradish-peroxidase (HRP)-linked anti-mouse and anti-rat IgG antibodies (Jackson Immunoresearch, West Grove, PA, USA) were commercially obtained.A549 cells were washed twice with phosphate-buffered saline (PBS) and fixed with 1% paraformaldehyde–PBS for 15 min. After washing twice with PBS, the cells were incubated with 1% bovine serum albumin–PBS overnight. Fixed cells were treated with mouse anti-ICAM-1 IgG antibody (clone 15.2) for 60 min and then washed three times with 0.02% Tween 20–PBS. The cells were further treated with HRP-linked anti-mouse IgG antibody for 60 min and then washed three times with 0.02% Tween-20–PBS. To develop the colorimetric reaction, the cells were incubated with the substrate solution (0.2 M sodium citrate (pH 5.3), 0.1% o-phenylenediamine dihydrochloride, 0.02% H2O2) for 20 min at 37 °C. Absorbance at 415 nm was measured with a Model 680 microplate reader (Bio-Rad Laboratories, Hercules, CA, USA).A549 cells were pulsed with 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT; 500 μg/mL) for 4 h and resultant MTT formazan was solubilized with 5% sodium dodecyl sulfate overnight. Absorbance at 595 nm was measured with a Model 680 microplate reader.A549 cells, HT-1080 cells and MCF7 cells were harvested with a cell scraper, washed once with PBS, and lysed with Triton X-100 lysis buffer consisting of 50 mM Tris-HCl (pH 7.4), 1% Triton X-100, 2 mM dithiothreitol, 2 mM sodium orthovanadate, and the protease inhibitor cocktail Complete™ (Roche Diagnostics, Mannheim, Germany). Postnuclear lysate was prepared by centrifugation (10,000 × g, 5 min). Protein samples (30 μg/lane) were separated by SDS-PAGE and transferred onto Hybond-ECL nitrocellulose membranes (GE Healthcare, Piscataway, NJ, USA). The membranes were incubated with primary antibodies and then HRP-conjugated secondary antibodies, followed by analysis using ECL Western blotting detection reagents (GE Healthcare).Total RNA was extracted from A549 cells using Sepasol®-RNA I super (Nacalai Tesque, Kyoto, Japan) and reverse-transcribed to cDNA using ReverTra Ace® (Toyobo, Osaka, Japan) and oligo(dT)20 according to the manufacturer's instructions. cDNA was amplified with a 7000 Fast Real-Time PCR System (Applied Biosystems, Foster City, CA, USA) using SYBR® Premix Ex Taq™ (Takara Bio, Otsu, Japan) and KOD -Plus- Ver. 2 DNA polymerase (Toyobo). The following primer pairs were used: ICAM-1: 5′-GGGAGGCTCCGTGCTGGTGA-3′ (forward) and 5′-TCAGTGCGGCACGAGA AATTG-3′ (reverse) [19]; c-FLIPL: 5′-GCCTGTATGCCCGAGCACCG-3′ (forward) and 5′-GCAGG GGGAGCCCTGAGTGA-3′ (reverse) [20]; β-actin: 5′-GGCATCGTGATGGACTCCG-3′ (forward) and 5′-GCTGGAAGGTGGACAGCGA-3′ (reverse). PCR conditions were 94 °C for 3 min, followed by 40 cycles of 94 °C for 15 s, 58 °C for 30 s, and 68 °C for 1 min. The quantity of the initial cDNA was calculated from primer-specific standard curves. The expression level of ICAM-1 and c-FLIPL was normalized on the basis of β-actin levels.A549 cells were transfected with a plasmid encoding κB-responsive firefly luciferase reporter gene using the Lipofectamine™ 2000 transfection reagent (Invitrogen) for 24 h. The transfected A549 cells were treated with agents as indicated. The cells were washed once with PBS, lysed with Triton X-100 lysis buffer, and centrifuged (10,000 × g, 5 min). Equal amounts of cell lysates were mixed with luciferase assay solution (0.25 mM luciferin, 0.8 mM ATP, 1 mM dithiothreitol, 9 mM MgCl2, 25 mM Tris-phosphate (pH 7.8), 15% glycerol) and relative light units were immediately measured with a Lumitester K-100 Luminometer (Hamamatsu Photonics, Hamamatsu, Japan).A549 cells were pulse-labeled with [3H]L -leucine (41.66 TBq/mmol; Moravek Biochemicals, Brea, CA, USA), [3H]L -glutamine (1.628 TBq/mmol; American Radiolabeled Chemicals, St. Louis, MO, USA), [3H]thymidine (2.37 TBq/mmol; MP Biomedicals, Santa Ana, CA, USA), and [3H]uridine (0.626 TBq/mmol; Moravek) for the indicated times. The labeled cells were washed three times with PBS and then lysed with 250 mM NaOH for 15 min, followed by 1 h incubation on ice in the presence of 5% trichloroacetic acid. The precipitates and the supernatants were separated by centrifugation (10,000 × g, 5 min). Radioactivity was measured with a 1900CA TRI-CARB® liquid scintillation analyzer (Packard Instrument, Meriden, CT, USA).Luciferase control RNA was subjected to the cell-free reaction (30 °C, 90 min) for translation by the Rabbit Reticulocyte Lysate System (Promega, Madison, WI, USA). Reaction mixtures were mixed with the luciferase assay solution and relative light units were immediately measured with the Lumitester K-100 Luminometer.A549 cells were treated with agents as indicated in control incubation buffer (125 mM NaCl, 4.8 mM KCl, 1.3 mM CaCl2, 1.2 mM MgSO4, 25 mM Hepes, 1.2 mM KH2PO4, 5.6 mM glucose, pH 7.4) and pulse-labeled with [3H]L -serine (0.74 TBq/mmol; Moravek) for 5 min. This was followed by the addition of cold L-serine at the final concentration of 500 μM. The cells were immediately washed three times with ice-cold PBS and lysed with 250 mM NaOH, and this was followed by radioactivity measurement.Highly purified Na+/K+-ATPase from porcine kidney (15.6 μmol Pi/min/mg protein at 37 °C) was a generous gift from Drs. Yoshikazu Tahara and Yutaro Hayashi (Department of Biochemistry, Kyorin University School of Medicine, Tokyo, Japan) [21,22] Assay conditions for the ATP-hydrolyzing activity of Na+/K+-ATPase were described previously [23]. The colorimetric assay for inorganic phosphate was basically performed as described previously [24].Statistical significance was assessed by one-way ANOVA followed by the Tukey test for multiple comparisons. Differences of P <0.05 were considered to be statistically significant.Human lung carcinoma A549 cells were alveolar epithelial cells highly responsive to pro-inflammatory cytokines and induced to express cell-surface ICAM-1, which was easily measured by the Cell ELISA assay. A549 cells were thus used as a model cell line for screening for anti-inflammatory agents as well as elucidation of their mode of actions. In our previous studies [9,10], ursolic acid was initially identified to inhibit the cell-surface ICAM-1 expression induced by pro-inflammatory cytokines.A549 cells were preincubated with various concentrations of ursolic acid for 1 h and then incubated with TNF-α for 6 h in the presence of ursolic acid. Ursolic acid inhibited the TNF-α-induced expression of cell-surface ICAM-1 in a dose-dependent manner and to the background level at concentrations higher than 30 μM (Figure 2a). This was not caused by the induction of cell death, since ursolic acid up to 50 μM did not decrease cell viability in the presence or absence of TNF-α during the same incubation time (Figure 2b). The amount of total ICAM-1 protein as translation product was then analyzed by Western blotting. Ursolic acid inhibited the TNF-α-induced ICAM-1 expression profoundly at concentrations higher than 30 μM (Figure 2c). In ursolic acid-treated samples, ICAM-1 protein bands migrated faster than those in control samples (Figure 2c), suggesting the ursolic acid may influence post-translational modification. To further investigate the effect of ursolic acid on ICAM-1 transcription, mRNA levels were evaluated by quantitative RT-PCR. TNF-α induced a drastic induction of ICAM-1 mRNA (Figure 2d). However, ursolic acid did not inhibit the TNF-α-induced ICAM-1 mRNA expression at concentrations up to 30 μM and decreased it partially at 40-50 μM (Figure 2d), although ICAM-1 expression at the cell-surface level as well as the protein level was almost completely inhibited by ursolic acid (Figure 2a and Figure 2c). These data suggest that ursolic acid inhibits the TNF-α-induced ICAM-1 expression more effectively at processes downstream of ICAM-1 transcription.It has been shown that TNF-α-induced ICAM-1 expression is highly NF-κB-dependent in A549 cells [25]. TNF-α stimulation induced the drastic degradation of IκBα (Figure 3a) and the nuclear translocation of the NF-κB subunit p65 (Figure 3b). Ursolic acid weakly affected the TNF-α-induced IκBα degradation in A549 cells, as compared with the proteasome inhibitor MG-132 that markedly prevented the IκBα degradation (Figure 3a). In addition, unlike MG-132, ursolic acid only marginally decreased the TNF-α-induced p65 translocation to the nucleus (Figure 3b). In the absence of TNF-α stimulation, it seems that ursolic acid alone increases p65 nuclear translocation to some extent without inducing IκBα degradation (Figure 3a and Figure 3b). Consistent with a partial reduction of ICAM-1 mRNA expression by ursolic acid (Figure 2d), these data indicate that ursolic acid does not efficiently inhibit the TNF-α-induced NF-κB signaling pathway.Ursolic acid inhibits TNF-α-induced ICAM-1 expression. (a) A549 cells were preincubated with various concentrations of ursolic acid for 1 h and then incubated with (filled circles) or without (open circles) TNF-α (2.5 ng/mL) for 6 h in the presence of ursolic acid. ICAM-1 expression (%) was measured by the Cell ELISA assay. Data points represent means ± SD (n = 3). * P < 0.01, compared with control; (b) A549 cells were incubated with various concentrations of ursolic acid for 1 h and then incubated with (filled circles) or without (open circles) TNF-α (2.5 ng/mL) for 6 h in the presence of ursolic acid. Cell viability (%) was measured by the MTT assay. Data points represent means ± SD (n = 3). * P < 0.05, compared with control; (c) A549 cells were preincubated with various concentrations of ursolic acid for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 6 h. Protein expression of ICAM-1 was analyzed by Western blotting; (d) A549 cells were pretreated with indicated concentrations of ursolic acid for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 6 h in the presence of ursolic acid. ICAM-1 mRNA expression was measured by quantitative RT-PCR. The ratio of ICAM-1 mRNA relative to β-actin mRNA is shown as means ± SD (n = 3). * P < 0.05 and ** P < 0.01, compared with control.Ursolic acid does not efficiently inhibit TNF-α-induced NF-κB signaling pathway. (a) A549 cells were preincubated with indicated concentrations of ursolic acid or MG-132 (20 μM) or without any of those compounds (–) for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 15 min in the presence of those compounds. Cell lysates were analyzed by Western blotting; (b) A549 cells were preincubated with indicated concentrations of ursolic acid or MG-132 (20 μM) or without any of those compounds (–) for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 30 min in the presence of those compounds. Cell lysates were analyzed by Western blotting.We further addressed whether the inhibitory effect of ursolic acid on TNF-α-induced ICAM-1 expression (Figure 2) is common to other NF-κB-responsive genes and not restricted to ICAM-1. Ursolic acid prevented the TNF-α-induced expression of c-FLIPL, a cytosolic protein mainly induced by NF-κB (Figure 4a). By contrast, the TNF-α-induced c-FLIPL mRNA expression was not decreased but rather increased upon treatment with ursolic acid (Figure 4b). Consistently, ursolic acid inhibited the expression of an NF-κB-responsive luciferase reporter to the background level (Figure 4c). Taken together, our present results suggest that ursolic acid inhibits the TNF-α-induced gene expression more effectively by targeting biosynthetic processes from mRNA to protein.To broaden the impact of our findings, the biological activity of ursolic acid was examined in two additional cell lines. Human fibrosarcoma HT-1080 cells and human breast adenocarcinoma MCF7 cells were preincubated with ursolic acid for 1 h and then incubated with TNF-α for 6 h in the presence of ursolic acid. As observed in A549 cells, ursolic acid profoundly inhibited the protein expression of ICAM-1 and c-FLIPL at concentrations more than 20 μM (Figure 5a and Figure 5b).Ursolic acid inhibits TNF-α-induced expression of NF-κB-responsive genes. (a) A549 cells were preincubated with various concentrations of ursolic acid for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 6 h in the presence or absence of ursolic acid. Protein expression of c-FLIPL was analyzed by Western blotting; (b) A549 cells were pretreated with indicated concentrations of ursolic acid for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 6 h in the presence of ursolic acid. c-FLIPL mRNA expression was measured by quantitative RT-PCR. The ratio of c-FLIPL mRNA relative to β-actin mRNA is shown as means ± SD (n = 3). * P < 0.05 and ** P < 0.01, compared with control; (c) A549 cells were transiently transfected with the NF-κB-responsive luciferase reporter for 24 h. The cells were preincubated with various concentrations of ursolic acid for 1 h and then incubated with TNF-α (2.5 ng/mL) for 6 h in the presence of ursolic acid. Cell lysates were prepared and their luciferase activities measured. Luciferase activity (%) is shown as means ± SD (n = 3). * P < 0.01, compared with control.Ursolic acid inhibits TNF-α-induced protein expression of ICAM-1 and c-FLIPL. (a and b) HT-1080 cells (a) and MCF7 cells (b) were preincubated with various concentrations of ursolic acid for 1 h and then incubated with (+) or without (–) TNF-α (2.5 ng/mL) for 6 h in the presence or absence of ursolic acid. Protein expression of ICAM-1 and c-FLIPL was analyzed by Western blotting.To investigate the effect of ursolic acid on cellular protein synthesis, A549 cells were labeled with [3H]L-amino acids and radioactivity incorporated into the acid-insoluble fractions was measured. Ursolic acid inhibited the incorporation of [3H]L-leucine and [3H]L-glutamine in a dose-dependent manner (Figure 6a and Figure 6b). Compared with protein synthesis, ursolic acid did not significantly affect the incorporation of [3H]thymidine and [3H]uridine into DNA and RNA fractions (Figure 6c to Figure 6e). The translation inhibitor cycloheximide exerted an inhibitory profile somewhat different from ursolic acid, as it prevented protein synthesis much more strongly than ursolic acid (Figure 6c). Then, we further investigated the direct effect of ursolic acid on the cell-free translation system where luciferase mRNA was used as a template. Ursolic acid decreased the luciferase activity by 30% at 50 μM (Figure 6f), whereas other translation inhibitors, such as puromycin, prevented luciferase activity completely (data not shown). The mechanism by which ursolic acid partially affects the cell-free translation system is currently unclear. Nevertheless, these data suggest that ursolic acid, unlike cycloheximide or puromycin, directly targets not the translation machinery but the amino acid transport across the plasma membrane.Ursolic acid inhibits cellular protein synthesis. (a and b) A549 cells were incubated with various concentrations of ursolic acid for 5 h and then incubated with [3H]L-leucine (a) or [3H]L-glutamine (b) for 2 h in the presence of ursolic acid. Radioactivity incorporated into the acid-insoluble fractions was measured. Radioactivity (%) is shown as means ± SD (n = 3). * P < 0.05 and ** P < 0.01, compared with control; (c to e) A549 cells were preincubated with or without ursolic acid (50 μM) or cycloheximide (10 μM) for 1 h and then incubated with [3H]L-leucine (c), [3H]uridine (d), or [3H]thymidine (e) for 2 h in the presence or absence of those compounds. Radioactivity incorporated into the acid-insoluble fractions was measured. Radioactivity (%) is shown as means ± SD (n = 3). * P < 0.05 and ** P < 0.01, compared with control; (f) Luciferase mRNA was translated by rabbit reticulocyte lysates in the presence of indicated concentrations of ursolic acid at 30 °C for 90 min. Luciferase activity (%) is shown as means ± SD (n = 3). * P < 0.01, compared with control.We have recently shown that the Na+/K+-ATPase inhibitors, such as ouabain, inhibit cellular protein synthesis by preventing the Na+-dependent transport of amino acids in A549 cells [26]. By this mechanism of action, ouabain strongly inhibits TNF-α-induced expression of gene expression at the translational level [26]. To investigate the effect of ursolic acid on the uptake of amino acids into the cell, A549 cells were labeled with [3H]L-leucine for 2 h and radioactivity incorporated into the cells was separated into acid-soluble supernatants (free amino acid) and acid-insoluble precipitates (proteins).Ouabain strongly decreased the incorporation of [3H]L-leucine into the cells as well as into proteins (Figure 7a). By contrast, cycloheximide barely affected the incorporation of [3H]L-leucine into the cells but strongly inhibited that into proteins (Figure 7b). Similar to ouabain, ursolic acid decreased the incorporation of [3H]L-leucine into the cells as well as into proteins (Figure 7a). The incorporation of [3H]uridine and [3H]thymidine into the cells was not significantly affected by either ursolic acid or ouabain (Figure 7c and Figure 7d). These data suggest that ursolic acid selectively inhibits the uptake of amino acids into the cells. To investigate the effect of ursolic acid on Na+-dependent amino acid transport, A549 cells were preincubated with ursolic acid for 1 h and then pulse-labeled with [3H]L-serine, which was shown to be incorporated into A549 cells via a Na+-dependent manner [26]. Ursolic acid was found to inhibit the uptake of [3H]L-serine (Figure 7e). Thus, these data reveal that ursolic acid inhibits Na+-dependent amino acid transport.Ursolic acid selectively inhibits the incorporation of amino acids into the cell. (a to d) A549 cells were preincubated with or without ursolic acid (50 μM) or ouabain (10 μM) for 1 h and then incubated with [3H]L-leucine for 2 h in the presence or absence of those two compounds (a); A549 cells were preincubated with or without cycloheximide (10 μM) for 1 h and then incubated with [3H]L-leucine for 2 h in the presence or absence of cycloheximide (b); A549 cells were preincubated with or without ursolic acid (50 μM) or ouabain (10 μM) for 1 h and then incubated with [3H]uridine (c) or [3H]thymidine (d) for 2 h in the presence or absence of those two compounds. Radioactivity incorporated into the cell was separated into acid-soluble supernatants (Sup) and acid-insoluble precipitates (Ppt). Total radioactivity incorporated into the cell (Sup+Ppt; gray bars) as well as radioactivity incorporated into the precipitates (Ppt; filled bars) is shown as means ± SD (n = 3). * P < 0.05 and ** P < 0.01, compared with control; (e) A549 cells were preincubated with various concentrations of ursolic acid for 1 h and then pulse-labeled with [3H]L-serine for 5 min in the presence of ursolic acid. Radioactivity incorporated into the cell was measured. Data points represent means ± SD (n = 3). * P < 0.01, compared with control.Na+/K+-ATPase pumps Na+ out and K+ in via the hydrolysis of ATP and plays a major role in maintaining Na+ and K+ gradients that couple amino acid transport. To investigate the direct effect of ursolic acid on Na+/K+-ATPase activity, Na+/K+-ATPase highly purified from porcine kidney was incubated in the presence of ursolic acid and ATP, and this was followed by measurement of inorganic phosphates released by hydrolysis. Ursolic acid inhibited ATPase activity in a dose-dependent manner and at the IC50 value of 24.7 μM (Figure 8).Ursolic acid inhibits Na+/K+-ATPase activity. Na+/K+-ATPase highly purified from porcine kidney was preincubated with various concentrations of ursolic acid for 15 min and then incubated with ATP for 30 min in the presence of ursolic acid. The inorganic phosphates produced by ATP hydrolysis were measured. Data points represent means ± SD (n = 3). * P < 0.01, compared with control.Pro-inflammatory cytokines activate the NF-κB signaling pathway and induce the expression of a wide range of NF-κB-responsive genes, such as ICAM-1. In our screening for anti-inflammatory agents, ursolic acid was initially identified to inhibit cell-surface ICAM-1 expression induced by pro-inflammatory cytokines, such as TNF-α [9,10]. It has been reported that ursolic acid inhibits the NF-κB signaling pathway constitutively activated or induced by pro-inflammatory cytokines or chemotherapeutic agents [11,12,13,14,15,16,17,18]. However, in contrast to these previous reports, we found that ursolic acid only partially diminishes TNF-α-induced NF-κB signaling pathway but more effectively inhibits the TNF-α-induced NF-κB-responsive gene expression by targeting cellular protein synthesis at least in A549 cells. The ability of ursolic acid to inhibit the NF-κB signaling pathway may be cell-type-specific or influenced by experimental conditions such as incubation periods. In fact, relatively long pre-incubation periods (8 to 48 h) with ursolic acid (10 to 100 μM) were used to block constitutive and inducible NF-κB activation in human cancer cell lines including A549 cells [11,12,15,18].Cardiac glycosides, such as ouabain, are known to be highly specific inhibitors of Na+/K+-ATPase activity. We have recently shown that ouabain and odoroside A inhibit the ATP-hydrolyzing activity of porcine kidney Na+/K+-ATPase at the IC50 values of 1.6 μM and 1.2 μM, respectively [26]. Using the same experimental systems, ursolic acid was found to inhibit Na+/K+-ATPase activity at the IC50 value of 24.7 μM. Thus, ursolic acid can be regarded as an inhibitor of Na+/K+-ATPase activity with 10- to 20-fold weaker potency than cardiac glycosides. In agreement with this, it has been recently shown that ursolic acid inhibits the activity of Na+/K+-ATPase purified from cerebral cortex at the IC50 value of 76.7 μM [27] as well as Na+/K+-ATPase activity in the crude mitochondrial fractions of human cell lines [28]. Moreover, a theoretical modeling study shows that ouabain and ursolic acid bind to Na+/K+-ATPase in a distinct manner [27,29]. In agreement with our previous study using cardiac glycosides [26], our present study for the first time reveals a link between the inhibition of Na+/K+-ATPase activity by ursolic acid and the prevention of NF-κB-inducible gene expression at the post-transcriptional level.Membrane transport of amino acids in mammalian cells is accomplished by a number of discrete systems and has distinct substrate specificity [30]. As a primary transporter, Na+/K+-ATPase pumps Na+ out and K+ in by hydrolyzing ATP and maintains Na+ and K+ gradients across plasma membranes [31]. Some secondary amino acid transporters are coupled with electrical and chemical gradients by primary active transporters [30]. We have recently shown that ouabain selectively inhibits both Na+-dependent membrane transport of amino acids and cellular protein synthesis by blocking Na+/K+-ATPase activity [26]. In accord with this notion, ursolic acid was found to inhibit the Na+-dependent membrane transport of L -serine as well as to diminish cellular protein synthesis much more strongly than DNA/RNA syntheses. Therefore, it seems most likely that ursolic acid mainly inhibits cellular protein synthesis by targeting Na+/K+-ATPase in the same mechanism as ouabain. However, as described previously [11,12,13,14,15,16,17,18], it should be noted that ursolic acid may interfere with intracellular proteins other than Na+/K+-ATPase as primary targets and modulate the NF-κB signaling pathway in different experimental settings.Our present study provides a novel molecular mechanism in which ursolic acid inhibits the translation process during the TNF-α-induced NF-κB-dependent gene expression by preventing Na+/K+-ATPase. Pro-inflammatory cytokines and their activation of the NF-κB signaling pathway play a critical role in chronic inflammation, which is known to be associated with autoimmune diseases, cancer, and metabolic disorders. Ursolic acid is a constituent of popular herbs and foods. Thus, it seems plausible that ursolic acid is a beneficial natural compound that exerts anti-inflammatory activity and is effective for the prevention of chronic inflammation.The authors declare no conflict of interest.We are very grateful to Yoshikazu Tahara and Yutaro Hayashi for the gift of Na+/K+-ATPase. This work was supported partly by a Grant-in-Aid for Scientific Research (KAKENHI) from Japan Society for the Promotion of Science (JSPS).

Med-MDPI/biomolecules/biomolecules-01-01-00048.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Glycosylation improves the solubility and stability of proteins, contributes to the structural integrity of protein functional sites, and mediates biomolecular recognition events involved in cell-cell communications and viral infections. The first step toward understanding the molecular mechanisms underlying these carbohydrate functionalities is a detailed characterization of glycan structures. Recently developed glycomic approaches have enabled comprehensive analyses of N-glycosylation profiles in a quantitative manner. However, there are only a few reports describing detailed O-glycosylation profiles primarily because of the lack of a widespread standard method to identify O-glycan structures. Here, we developed an HPLC mapping method for detailed identification of O-glycans including neutral, sialylated, and sulfated oligosaccharides. Furthermore, using this method, we were able to quantitatively identify isomeric products from an in vitro reaction catalyzed by N-acetylglucosamine-6O-sulfotransferases and obtain O-glycosylation profiles of serum IgA as a model glycoprotein.Glycosylation is one of most ubiquitous post-translational modifications. Carbohydrate moieties, which are typically found on asparagine or serine/threonine residues, are associated with an increase in solubility and stability of proteins, structural integrity of protein functional sites, and mediation of biomolecular recognition events involved in cell-cell communications and viral infections [1,2]. Since glycans affect the serum half-life of proteins and functional protein–protein interactions, glycosylation is currently considered to be a crucial factor in the design and development of biopharmaceuticals [3,4,5]. To address the detailed molecular basis of the functional roles of protein glycosylation, the first step is identifying the glycan structures expressed on the proteins [6,7,8,9,10]. Recently developed glycomic approaches using chromatographic and mass spectrometric (MS) techniques have enabled comprehensive analyses of N-glycosylation profiles [11,12]. For example, a multidimensional HPLC mapping method has been developed for quantitative N-glycosylation profiling at molecular, cellular, and tissue levels, enabling isomeric N-glycan structures [13,14]. In this method, identification of individual N-glycans is based on their elution positions on three types of HPLC columns. The accumulated HPLC data of approximately 500 different N-glycans are now available by using the web application GALAXY (http://www.glycoanalysis.info/galaxy2/ENG/index.jsp) [15], and the applicability of this method has been extended to sialylated, glucronylated, and sulfated oligosaccharides [16,17,18].However, few reports describe the detailed O-glycosylation profiles with linkage information due to the lack of widespread standard methods for unambiguous identification of O-glycan structures [19,20,21]. The HPLC elution conditions employed in the current GALAXY protocols are not applicable to the profiling of O-glycans, because they frequently include smaller saccharides, e.g., mono- and di-saccharides, in contrast to the generally larger N-glycans. In view of this situation, we herein attempted to develop HPLC-based profiling of O-glycans for their detailed structural identification. By chromatographic and mass spectrometric analyses in conjunction with several exoglycosidase treatments in vitro, we successfully collected HPLC data for 27 different O-glycans including neutral, sialyl, and sulfated oligosaccharides, which were isolated from natural sources and/or by in vitro enzymatic reactions. By applying this extended HPLC map, we have obtained O-glycosylation profiles of serum immunoglobulin A (IgA) as a model glycoprotein. Furthermore, we characterized branch specificity in the sulfation reaction catalyzed by human N-acetylglucosamine-6O-sulfotransferases (GlcNAc6ST)-1.Materials used for the experiments were purchased from the sources indicated below: Glycoamidase A from sweet almond, β-galactosidase and β-N-acetylhexosaminidase from jack bean were purchased from Seikagaku Kogyo Co. (Tokyo, Japan). α-Galactosidase from green coffee bean was purchased from Oxford Glycosystems Inc. (Bedford, MA, USA) (currently available in Prozyme (Hayward, CA, USA)). α-Sialidase from Arthrobacter ureafaciens was purchased from Nacalai Tesque (Kyoto, Japan). α2,3-Sialidase from Salmonella typhimurium was purchased from Takara Bio Inc. (Otsu, Japan). Recombinat α2,3-sialyltransferase and α2,6-sialyltransferase were purchased from Calbiochem (San Diego, CA, USA). Colostrum IgA, porcine stomach mucin, trypsin, and chymotrypsin were purchased from Sigma Chemical Co. (St. Louis, MO, USA).2-Aminopyridine-derivatized (PA) isomalto-oligosaccharides were prepared from glucose oligomers (1-20) (Seikagaku Kogyo Co., Tokyo, Japan), fucose (Fuc), galactose (Gal), N-acetylgalactosamine (GalNAc) (Seikagaku Kogyo Co., Tokyo, Japan), glucose (Glc), N-acetylglucosamine (GlcNAc), mannose (Man), Galβ1-3GalNAc (Calbiochem, Schwalbach, Germany), and Galβ1-3(Fucα1-2)GalNAc. Four types of O-glycosylated peptides—Galβ1-4GlcNAcβ1-6(Neu5Acα2-3Galβ1-3)GalNAcα 1-peptide (AHGVT*SAPDTRK; asterisks indicate glycosylation sites)-FAM (5-carboxyfluorescein), Galβ1-4GlcNAcβ1-6(Galβ1-3)GalNAcα1-peptide-FAM, GlcNAcβ1 -6(GlcNAcβ1-3Galβ 1-3)GalNAcα1-peptide-FAM, and Neu5Acα2-6(Galβ1-3)GalNAc-peptide-FAM were purchased from GlycoGene (Tsukuba, Japan).Human serum (1 mL) was diluted in 10 mL of 0.01 M phosphate buffer (pH 7.4) containing 0.15 M NaCl and absorbed on jacalin-agarose columns (1 mL) (Thermo Scientific, San Jose, CA, USA). After the column was thoroughly washed with 10 mL of 0.01 M phosphate buffer (pH 7.4) containing 0.15 M NaCl and 0.8 M glucose, lectin-binding proteins were eluted with 10 mL of 0.01 M phosphate buffer (pH 7.4) containing 0.1 M melibiose as described previously [22]. After the eluate was concentrated using an AMICON Ultra-15 centrifugal filter unit (Millipore, Billerica, MA, USA), serum IgA was purified with a Superose 12 gel filtration column (GE Healthcare, Uppsala, Sweden) equilibrated with 0.01 M phosphate buffer (pH 7.4) containing 0.15 M NaCl. The purified IgA was desalted with a PD-10 column (GE Healthcare) according to the manufacturer's instructions and then lyophilized for glycosylation profiling.COS7 cells grown in 75 cm2 culture flasks (Corning, Corning, NY) were transfected with 10 μg of relevant plasmid, pcDNA-GlcNAc6ST-1 [23] using Lipofectamine Plus (Invitrogen, Carlsbad, CA, USA) according to manufacturer's instructions. After 24 h of culture in Dulbecco's modified Eagle's medium (DMEM) containing 10% fetal calf serum, the medium was replaced with DMEM containing 2% IgG-free fetal calf serum. The cells were further cultured for 96 h. Subsequently, the culture medium was collected and concentrated to 1 mL using Amicon Ultra-15 (Millipore). The recombinant protein A-fused GlcNAc6ST-1 expressed in the medium was adsorbed to IgG-Sepharose (20 μL resin/1 mL of culture medium) at 4 °C for 3 h. The resin was collected by centrifugation and washed three times with 50 mM Tris-HCl (pH 7.5). Finally, the resin was suspended in 20 μL of 50 mM Tris-HCl (pH 7.5) and used for enzymatic reaction. The glycopeptide GlcNAcβ1-6(GlcNAcβ1-3Galβ1-3)GalNAcα1-peptide-FAM was utilized as an acceptor substrate. The standard reaction mixture was composed of 1 μmol of Tris-HCl (pH 7.5), 0.4 μmol of MnCl2, 0.08 μmol of AMP, 24 μmol of NaF, 50 pmol of glycopeptide, 300 pmol of 3′-phosphoadenosine 5′-phosphosulfate, 0.1% Triton X-100, and 20 μL of the fusion protein suspension in a final volume of 40 μL. After incubation at 37 °C for 1, 5, 24, and 48 h, the individual reaction mixtures were applied to a TSK gel ODS-80s HPLC column (TOSOH, Tokyo, Japan) at a flow rate of 1.0 mL/min at 25 °C using two solvents: G and H. G comprised water containing 0.05% trifluoroacetic acid and H comprised acetonitrile-2-propanol (2:1, v/v) containing 0.05% trifluoroacetic acid. The column was equilibrated with 90% solvent G and 10% solvent H. The time for gradient elution was 0–40 min with a linear gradient of 10%–15% D. The glycopeptides were detected by fluorescence using excitation and emission wavelengths of 492 and 520 nm, respectively.The O-glycans were released from glycoproteins and glycopeptides by β-elimination using hydrazine for a convenient modification with 2-aminopyridine. Lyophilized glycoproteins (∼250 μg) or glycopeptides (∼5 μg) were dissolved in 1 mL of distilled anhydrous hydrazine with a water content of less than 1% (v/v) in 10 mL glass tube, incubated at 60 °C for 6 h and quenched by 9 mL of 50 mM ammonium acetate buffer (pH 7) with slight modification of the previous literature [24]. The excess hydrazine, peptides, and other reagents were removed and N-acetylated using a graphite carbon column (GL-Pak Carbograph, GL Science, Tokyo, Japan) according to the literature [25]. The hydrazine solution was mixed with 3 mL of 50 mM ammonium acetate buffer (pH 7) and loaded onto the GL-Pak Carbograph column. After the column was washed with 15 mL of 50 mM ammonium acetate buffer (pH 7.0), the released glycans were eluted with 5 mL of a mixture of 50 mM ammonium acetate buffer (pH 7.0):acetonitrile containing 2% acetic anhydride (40:60).The released O-linked saccharides, as well as those commercially obtained, were labeled with 2-aminopyridine as described previously [26]. Ten volumes of acetonitrile were added to one volume of reaction mixture. The excess PA reagents were removed with a MonoSpin NH2 desalting column (GL Science). After the column was equilibrated with 200 μL of acetonitrile, the PA reaction mixture was loaded onto the column. The column was washed with acetonitrile three times. Then, the PA-saccharides were eluted with 100 μL of water and subsequently dried under vacuum.Three types of HPLC columns were used for the separation of PA-saccharides. In each step, PA-saccharides were detected by fluorescence using excitation and emission wavelengths of 310 and 380 nm, respectively. The PA-saccharide mixture was firstly separated on a Mono-Q HR 5/5 anion-exchange column (GE Healthcare) at 30 °C with a flow rate of 1.0 mL/min using two solvents, A and B. Solvent A was aqueous ammonia (pH 9.0) and solvent B was a 50 mM ammonium acetate solution (pH 9.0). The column was equilibrated with solvent A. The gradient elution parameters were 0–3 min, linear gradient 0%–12% B; 3–17 min, linear gradient 12%–40% B; 17–22 min, linear gradient 40%–100% B. Each oligosaccharide was separated according to its anionic charges.In the second step, each fraction separated from the Mono-Q column was collected, evaporated, and then applied to a Decelosil C30-HG-5 (C30) column (Nomura Chemical Co., Ltd., Seto, Japan). Elution was performed at a flow rate of 1.0 mL/min at 25 °C using two solvents, C and D. Solvent C was 0.1 M ammonium acetate buffer (pH 6.0) containing 0.01% 1-butanol and solvent D was 0.1 M ammonium acetate buffer (pH 6.0) containing 1% 1-butanol. The column was equilibrated with solvent C. The gradient elution parameters were 0–51 min, linear gradient 0%–50% D and 51–63 min, linear gradient 50%–100% D.In the third step, individual peak fractions from the C30 column were isolated using a TSK gel amide-80 size fractionation column (TOSOH). In this system, two solvents were used at 25 °C. Solvent E was composed of 3% acetic acid in water with triethylamine (pH 7.3) and acetonitrile, 15:85 by volume. Solvent F was composed of 3% acetic acid in water with triethylamine (pH 7.3). The column was equilibrated with solvent E. The gradient elution parameters were 0–5 min, linear gradient 0%–20% F and 5–17 min, linear gradient 20%–44% F.The HPLC elution times were represented by glucose units (GUs) on the columns calibrated with a PA-derivatized isomalto-oligosaccharides mixture. The structures of the PA-saccharides were characterized by HPLC mapping in conjunction with exoglycosidase treatments and matrix-assisted laser desorption/ionization time of flight (MALDI-TOF-MS) analysis using a MALDI-TOF-MS spectrometer (AXIMA-CFR; Shimadzu, Kyoto, Japan). Collision-induced dissociation spectra of PA-oligosaccharides were acquired using a MALDI-quadrupole ion trap (QIT)-TOF-MS spectrometer (AXIMA-QIT; Shimadzu). All analytical procedures used in this work, including sulfation, sialylation, several glycosidase treatments, and MALDI-TOF-MS analyses have been described previously [16,17,27,28,29].First, we attempted to make an HPLC map of the standard PA-saccharide. The PA tag was attached to the commercially obtained saccharides Fuc, Gal, GalNAc, Glc, GlcNac, Man, Galβ1-3GalNAc, and Galβ1-3(Fucα1-2)GalNAc. In addition, four types of O-glycosylated peptides were treated with hydrazine, and the released oligosaccharides were labeled with 2-aminopyridine. The PA-saccharides thus prepared were subjected to amide and C30 columns to record their elution times (Table 1). Furthermore, pyridylaminated O-glycans were prepared from colostrum IgA and porcine stomach mucin, and their structures were identified by chromatographic analyses combined with exoglycosidase treatments and MALDI-TOF-MS. The structural identification of these O-glycans would be exemplified by a glycan derived from mucin. Since no sialylated O-glycans were detected in the glycosylation profile of mucin on a Mono-Q column (data not shown), the PA-glycans were directly applied to an amide column. Figure 1a shows the O-glycosylation profile of the mucin on the amide column in which two major O-glycans were found. Then, fraction A was applied to a C30 column, giving rise to several peaks including B (Figure 1b). The elution times of the PA-O-glycan in fraction B are represented as 3.2 GU on the amide column and 4.6 GU on the C30 column. The molecular mass of this glycan was determined by MALDI-TOF-MS analysis as 665 Da, which corresponds to (Hex)1(HexNAc)2PA (Figure 1c). The fragment ions indicated that the PA-O-glycan exhibits the branching structure Hex-(HexNAc-)HexNAc-PA. Finally, the glycan corresponding to fraction B was treated with β1,3-galactosidase and then applied to a C30 column, giving rise to a new fraction. The elution time of the glycan corresponding to this fraction coincided with that of the reference PA-glycan, GlcNAcβ1-6GalNAc-PA. On the basis of all these data, we concluded that the glycan corresponding to fraction B was Galβ1-3(GlcNAcβ1-6)GalNAc-PA.HPLC and mass sprctrometric (MS) data of PA-O-glycans.Average mass calculated from the m/z values of [M+H]+, [M+Na]+, and [M-H]– ions for PA-saccharides.Isolation and identification of an O-glycan derived from porcine stomach mucin. (a)Chromatogram of PA-glycans derived from mucin on an amide column; (b) Chromatograms of the PA-glycans corresponding to fraction A on the C30 column; (c) MALDI-QIT-TOF-MS/MS spectra of the PA-glycan corresponding to fraction B. Precursor ion was m/z 666 as protonated ion; (d) Chromatograms of the PA-glycan corresponding to fraction B on the C30 column (upper) before and (lower) after β1,3-galactosidase treatment. The asterisk indicates the fractions containing no detectable PA-saccharide.With similar methodology, we identified eight types of O-glycans derived from mucin and IgA glycoproteins and recorded their elution times (Table 2). Using the HPLC data as a guide, these O-glycans could be strategically collected from glycoproteins and further derivatized by glycosidase glycosyltransferase and sulfotransferase treatments in vitro, giving rise to a variety of standard PA-oligosaccharides. For example, the mono-sialyl PA-oligosaccharide Galβ1-3(Neu5Acα2-6)GalNAc-PA was treated with α2,6-silayltransferase, giving rise to di-sialyl PA-oligosaccharide Neu5Acα2-6Galβ1-3(Neu5Acα2-6)GalNAc-PA, which was eluted differently from the reaction precursor on the C30 column (Figure 2). Similarly, we collected the HPLC data of seven kinds of PA-O-glycans (Table 3). Finally, we made an HPLC map containing 16 neutral, seven sialylated, and four sulfated O-glycans (Figure 3).HPLC and MS data of PA-O-glycans derived from colostrum IgA and porcine stomach mucin.Average mass calculated from the m/z values of [M+H]+, [M+Na]+, and [M-H]– ions for PA-saccharides.Identification of the disialyl PA-saccharide produced by the reaction catalyzed by α2,6-sialyltransferase. Chromatograms of (a) the precursor Galβ1-3(Neu5Acα2-6)GalNAc-PA; and (b) the reaction product Neu5Acα2-6Galβ1-3(Neu5Acα2-6)GalNAc-PA on the C30 column.HPLC and MS data of PA-O-glycans produced in vitro by derivatisation of the neutral O-glycans listed in Table 1 and Table 2.Average mass calculated from the m/z values of [M+H]+, [M+Na]+, and [M−H]− ions for PA-saccharides.HPLC map of O-glycans. •, neutral; ▪, sialylated; ♦, sulfated glycans. Key: Gal, galactose; Glc, Glucose; GlcNAc, N-acetylglucosamine; GalN, N-acetylgalactosamine; Man, mannose; Fuc, fucose; S, sulfate; Neu, N-acetylneuraminic acid.The HPLC map thus established facilitates the quantitative O-glycosylation profiling with discriminating isomeric structures of O-glycans, which would be difficult to perform by MS-based approaches. The HPLC-based O-glycosylation profiling methods so far reported need much longer elution times (more than 2 h) or employ different elution conditions between neutral and acidic O-glycans [19,20]. Our developed HPLC map is able to deal with neutral and anionic O-glycans (including sulfated O-glycans whose HPLC data have not been reported previously) with the same protocol using a shorter elution time (within 1 h) and therefore would be advantageous in comparison to previously reported methods.The HPLC map thus established will facilitate structural identification of sulfated O-glycans. To date, five types of sulfotransferases (termed GlcNAc6STs) have been reported to catalyze a sulfate group on the C6 position of GlcNAc residues [30]. Although spatio-temporal expression patterns of these enzymes have been extensively characterized, their reaction specificities are not fully understood. We herein applied the developed HPLC data to examination of the branch specificity of enzymatic sulfation catalyzed by human GlcNAc6ST-1, which was expressed by COS7 cells as a fusion protein with protein A [23].Figure 4a shows the time-dependent change of reverse-phase HPLC elution profiles for the reaction mixture of the in vitro sulfation catalyzed by this recombinant enzyme using a fluorescent glycopeptide, GlcNAcβ1-6(GlcNAcβ1-3Galβ1-3)GalNAcα1-peptide-FAM, as an acceptor. MALDI-TOF-MS indicated that two reaction products, A and B, were isomeric glycopeptides that possessed a single sulfate group. For unambiguous identification of the isomeric structures of sulfated O-glycans, fractions A and B were subjected to the developed HPLC mapping. As a result, the PA-glycans derived from fractions A and B were identified as GlcNAcβ1-6(HSO3-GlcNAcβ1-3Galβ1-3)GalNAc-PA and HSO3-GlcNAcβ1-6(GlcNAcβ1-3Galβ1-3)GalNAc-PA, respectively. This result clearly indicates that GlcNAc6ST-1 selectively catalyzed sulfation at the β1-6-linked GlcNAc residue in comparison with the remaining β1-3-linked GlcNAc residue, consistent with the previous report that preferential sulfation occurs with core 2 (GlcNAcβ1-6(Galβ1-3)GalNAc) rather than core 3 (GlcNAcβ1-3Galβ1-3GalNAc) as the acceptor [23]. Thus, our HPLC map is a useful tool for detailed characterization of substrate and reaction specificities of glycosyltransferases and sulfotransferases, leading to a better understanding of their detailed functional roles.We also applied our HPLC map to O-glycosylation profiling of human serum IgA, which possesses nine O-glycosylation sites at the hinge region [31,32]. Galactose depletion of O-glycans at the IgA hinge has been observed in the serum of patients with IgA nephropathy [31,32]. Figure 5a shows a typical elution profile on a Mono-Q column of the PA-O-glycans derived from the IgA sample, which were separated according to the degrees of sialylation. Each fraction was further separated on a C30 column as shown in Figure 5b. Individual fractions separated by the C30 column were further separated on an amide-silica column. The PA-oligosaccharides were identified on the basis of coincidence of the elution data with those in the HPLC map established in the present study. The incidence of O-glycan structures derived from serum IgA is indicated in Figure 5b. To date, IgA O-glycosylation has been characterized by lectin blotting, mass spectroscopy, and chromatographic separation [21,31,32,33]. These studies have described O-glycan structures Galβ1-3GalNAc, Neu5Acα2-3Galβ1-3GalNAc, Galβ1-3(Neu5Acα2-6)GalNAc-PA, and Neu5Acα2-3Galβ1-3(Neu5Acα2-6)GalNAc. To the best of our knowledge, the present study is the first to identify the di-sialyl O-glycan Neu5Acα2-6Galβ1-3(Neu5Acα2-6)GalNAc in serum IgA. The HPLC map developed in the present study enables us to distinguish the isomeric structures of sialyl O-glycans, offering quantitative information for O-glycosylation profiling.HPLC-based characterization of branch specificity of GlcNAc6ST-1. (a) Time-dependent change of the HPLC profiles on the ODS column for products resulting from glycopeptides possessing two terminal GlcNAc residues during the sulfation reaction catalyzed by the recombinant GlcNAc6ST-1. The asterisk indicates the fractions containing the substrate O-glycosylated peptide: GlcNAcβ1-6(GlcNAcβ1-3Galβ1-3)GalNAcα1-peptide-FAM; (b) Time course of the amounts of the resultant glycopeptides corresponding to fraction A (solid line) and B (dashed line). The O-glycan structures of fractions A and B were identified as GlcNAcβ1-6(HSO3-GlcNAcβ1-3Galβ1-3)GalNAc and HSO3-GlcNAcβ1-6(GlcNAcβ1-3Galβ1-3)GalNAc, respectively.HPLC profiles of PA-O-glycans derived from human serum IgA. (a) Chromatogram of PA-O-glycans derived from serum IgA on a Mono-Q column. The PA-glycan mixture was separated according to sialic acid contents. N, neutral; MS, mono-sialyl; DS, di-sialyl; (b) Chromatograms of the neutral, mono-sialyl, and di-sialyl fractions on a C30 column. The structures of PA-O-glycans in each fraction were identified on the basis of the HPLC map. Molar percent of the O-glycan content in the IgA sample was calculated based on the peak areas. The structure and incidence of major O-glycans are shown on the profiles. Asterisks indicate the peaks derived from melobiose used for IgA purification.In the present study, we developed an HPLC mapping method for detailed structural identification of O-glycans in neutral, sialylated, and sulfated forms. Furthermore, using this method, we were able to quantitatively identify isomeric products from an in vitro reaction catalyzed by human GlcNAc6ST-1 and obtain O-glycosylation profiles of human serum IgA as a model glycoprotein. The HPLC map will provide a glycomics tool for unambiguous identification and quantitative profiling of O-glycans expressed on a variety of proteins of physiological and pathological interest.This work was supported in part by Grants-in-Aid for Young Scientists B (H.Y.) and Scientific Research B (K.K.) from the Ministry of Education, Culture, Sports, Science and Technology (MEXT), and the Program for Promotion of Fundamental Studies in Health Sciences of the National Institute of Biomedical Innovation (NIBIO) (K.K.). We thank Kazuhiro Ikenaka (National Institute for Physiological Sciences) for his helpful remarks and suggestions about the purification of PA-O-glycans. We also thank Yoshiki Yamaguchi (RIKEN) for his support in the MS/MS analyses. We also appreciate Uchimura (Research Institute, National Center for Geriatrics and Gerontology) and Kannangi (Aichi Medical University) for the gift of the expression vector of GlcNAc6ST-1 and for the useful discussion. Glycosylation profiling of human serum IgA has been approved by the ethics committee of Nagoya City University.

Med-MDPI/biomolecules/biomolecules-02-01-00001.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Accurately predicting essential genes is important in many aspects of biology, medicine and bioengineering. In previous research, we have developed a machine learning based integrative algorithm to predict essential genes in bacterial species. This algorithm lends itself to two approaches for predicting essential genes: learning the traits from known essential genes in the target organism, or transferring essential gene annotations from a closely related model organism. However, for an understudied microbe, each approach has its potential limitations. The first is constricted by the often small number of known essential genes. The second is limited by the availability of model organisms and by evolutionary distance. In this study, we aim to determine the optimal strategy for predicting essential genes by examining four microbes with well-characterized essential genes. Our results suggest that, unless the known essential genes are few, learning from the known essential genes in the target organism usually outperforms transferring essential gene annotations from a related model organism. In fact, the required number of known essential genes is surprisingly small to make accurate predictions. In prokaryotes, when the number of known essential genes is greater than 2% of total genes, this approach already comes close to its optimal performance. In eukaryotes, achieving the same best performance requires over 4% of total genes, reflecting the increased complexity of eukaryotic organisms. Combining the two approaches resulted in an increased performance when the known essential genes are few. Our investigation thus provides key information on accurately predicting essential genes and will greatly facilitate annotations of microbial genomes.Essential genes are defined as those that, when disrupted, confer a lethal phenotype to microorganisms under defined conditions. As such, the essentiality of a gene is the indispensability of this gene’s product to the survival of a microorganism. A complete understanding of gene essentiality is important in multiple facets of biology, medicine and bioengineering. For example, because of the lethal consequences of their disruption, essential genes are often attractive targets of antibiotics [1]. Essential genes of an organism also constitute its minimal gene set, a key concept in the emerging field of synthetic biology [2,3]. Furthermore, studying gene essentiality is a crucial step toward unraveling the complex relationship between genotype and phenotype [4], a fundamental question in genetics. Systematic genome-wide interrogations of essential genes have been conducted by single gene knockouts [5,6,7,8], transposon mutagenesis [9,10,11,12,13,14,15], or antisense RNA inhibitions [16,17]. Although the efficiency of gene deletion has improved, performing large-scale experiments to knock out each gene encoded in an organism’s genome, usually in the magnitude of thousands, is still a daunting task. The work of experimentally identifying essential genes in an organism is even more formidable than was once thought as researchers have found that growth conditions can significantly alter the spectrum of essentiality in bacteria [18,19,20,21,22] and yeast [23]. Therefore, computational methods for predicting essential genes become an appealing option for circumventing the expense and difficulty of experimental screens. A computational prediction is especially useful when the organism is either unculturable, such as Pneumocystis carinii, or difficult to perform gene disruption on, such as Aspergillus fumigatus.In our previous research, we developed a machine-learning based algorithm that predicts essential genes by integrating diverse types of information encoded in a microorganism’s genome that are potentially associated with gene essentiality [24]. We tested this algorithm in four bacterial species whose essential genes have been well characterized: Escherichia coli (EC), Pseudomonas aeruginosa (PA), Acinetobacter baylyi (AB) and Bacillus subtilis (BS). Ten-fold cross-validations in each organism showed a high predictive accuracy (AUC: ~0.9). We also reported that gene essentiality can be reliably transferred using features trained and tested in a distantly related microorganism (AUC: 0.69–0.89). Cross-organism predictions significantly outperformed homology mapping. Our algorithm thus significantly extended our ability to predict essential genes beyond orthologs by providing two alternative approaches: We can learn the characteristics underlying the subset of known essential genes in one organism and predict the essentiality of the rest of the genes in the same organism. Alternatively, we can transfer the gene essentiality from its most closely related model organisms where a complete set of essential genes is available. However, to determine the essential gene set in an understudied microbe, both approaches have potential limitations. The first approach is limited by the often low number of known essential genes, while the second approach is limited by the availability of model organisms and the evolutionary distance to the target organism. Although our previous work demonstrated that both approaches are capable of producing accurate predictions, further study is needed to determine the most suitable situation each approach can be employed.The current study represents a significant progress since our previous work by aiming to determine an optimal strategy for predicting essential genes in an understudied microbe by examining these potential limitations with regard to the above-mentioned approaches and a third approach that combines the two approaches. We performed our investigations on two pairs of microbes with well-characterized essential genes: two prokaryotes, Escherichia coli K-12 (EC) and Acinetobacter baylyi ADP1 (AB) and two eukaryotes, Saccharomyces cerevisiae S288c (SC) and Neurospora crassa OR74A (NC). We withheld different fractions of known essential genes in each organism and evaluated the predictive performance. Through these simulations, we were able to reveal the conditions under which each approach is most suitable for predicting essential genes in a microbe with respect to the size of known essential genes. The results obtained from our study will greatly facilitate the annotations of microbial genomes and provide valuable information to synthetic biology. E. coli K-12 sequence data were downloaded from Comprehensive Microbial Resource (CMR) database at http://cmr.jcvi.org. It contains 4289 protein sequences in total [25]. The essential genes of E. coli K-12 were downloaded from the PEC database [7]. The Kato dataset contains 302 essential genes from gene deletion experiments. A. baylyi ADP1 sequences were collected from the Magnifying Genomes database (http://www.genoscope.cns.fr/agc/mage). Of a total of 3308 genes, 499 are essential genes from de Berardinis et al. [6]S. cerevisiae S288c sequences were downloaded from Saccharomyces Genome Database at: http://downloads.yeastgenome.org/sequence/genomic_sequence/. It contains 5885 ORFs. The essential gene list was downloaded from Giaever et al. [26]. This dataset contains 1049 essential genes from targeted mutagenesis experiments. N. crassa OR74A ORFs were downloaded from Neurospora crassa database at Broad Institute at http://www.broadinstitute.org/annotation/genome/neurospora/MultiDownloads.html. Dubious ORFs and pseudogenes were excluded from this list. The essential gene dataset was kindly provided by K. Borkovich at UC Riverside from the systematic genome deletion project in N. crassa. This list contains 7172 experimental verified essential/nonessential genes, and 1251 of them are essential genes.Gene expression data in these organisms were downloaded from NCBI GEO [27], ArrayExpress [28], and the gene-expression profiles of microarray data from Gasch et al. [29].Based on our previous research, we considered three main types of features: (A) those intrinsic to a gene’s sequence (e.g., GC content, length); (B) those derived from genomic sequence (e.g., localization signals and codon adaptation measures); and (C) experimental functional genomics data (e.g., gene-expression microarray data) (Table 1). The detailed descriptions of these features and their biological implications can be found in the supplemental materials as well as in Deng et al. [24]. For example, domain enrichment score (DES) reflects the conservation of local domains rather than the entire gene, which is calculated by the ratio of the domain’s occurrence frequencies in essential genes vs. in total genes in a given organism. In another example, phylogenetic score (PHYS) measures the evolutionary conservation of a gene, which is calculated by counting the number of genomes that have orthologous hits. Such conservation has been shown to correlate well with the indispensability of a gene. These diverse types of features suggest that gene essentiality is likely determined not solely by the genomic sequence, but by multiple aspects of biology coinciding. Thirty-five considered features.*—Class A: Sequence-based intrinsic features; Class B: Sequence-derived intrinsic features; Class C: Context-dependent features; **—Features used in the training and testing in each organism are in bold.We evaluated these features based on their predictive power following a procedure described in Deng et al. [24]. To briefly summarize, we performed a logistic regression analysis and ranked all features according to the cover length of log-odds ratio. A longer overall coverage length indicates greater contribution of the corresponding feature to the gene essentiality. Because we were more interested in predicting essential genes rather than non-essential genes, the features with a positive coverage length were our candidate features. We also considered prior biological information to remove feature redundancy. The training data included the attribute values for each feature and the class assignments. Each gene was assigned a Boolean value regarding its essentiality (1—essential; 0—non-essential). The training data were divided into 10 equal parts. Nine parts were used to train the classifiers and the remaining part was used for testing. The control training set was generated by randomly assigning essential labels to all genes. The same number of random “essential genes” as the number of true essential genes was used in the training and testing frame.For each of the four organisms (i.e., EC, AB, SC and NC), we withheld different fractions of known essential genes to simulate the situation that only partial true essential genes were known. These known essential genes were selected through random sampling and comprised of our “gold standard” positive set. Because there are more non-essential genes than essential genes (10:1 in prokaryotes and 5:1 in eukaryotes), we constructed our training datasets with the same essential vs. non-essential ratio to resemble the situation in nature. That is, for a “gold standard” positive set of size N, we randomly selected xN (x = 10 for prokaryotes, and 5 for eukaryotes) genes from the non-essential genes as the “gold standard” negative set. We then solved the problem of imbalanced training set through data re-sampling, where we extracted a smaller set of non-essential genes while preserving all the essential instances. This method modifies the prior probability of the non-essential and essential classes to obtain a more balanced training set. Similar approaches have been used in other studies [30,31]. We trained our model using this training set. Each time we repeated the random process 200 times to obtain a reliable result. As described in Deng et al. [24], when predicting essential genes in each of the four organisms, the training set is the complete gene set of its paired organism. For example, when we predict essential genes in EC, the training set is the complete gene set in AB, where the complete AB essential genes compose the “gold standard” positive set and the remaining AB non-essential genes consist of the “gold standard” negative set.For each of the four organisms, the training set was constructed as the combination of the training sets in the same-organism approach and cross-organism approach. Meanwhile, we assigned different weights to each model organism based on the evolutionary distance to the target organism. For example, when we predicted essential genes in EC, the “gold standard” positive set consisted of a randomly selected fraction of essential genes in EC together with the complete set of essential genes in AB, where genes from EC were assigned weights w (w > 1), and those from AB were assigned a weight of 1. Similarly, the “gold standard” negative set consisted of the same fraction of randomly selected non-essential genes from EC together with the complete set of non-essential genes in AB, with weights w and 1 respectively.We used a logistic regression classifier to train and test the model. All classifiers were implemented using the Orange software package (http://www.ailab.si/orange/). To train and test our classifier, features were first extracted where available for each ORF and annotated with known essentiality values, thereby creating our “gold standard” data set. Then the “gold standard” dataset was divided into 10 equal parts. Nine parts were used to train the classifiers and the remaining part was used for testing.Then we applied the model to the target organism, and predicted the probability of essentiality for each gene in that organism. Based on the true gene labels and the predicted probability, we were able to calculate the AUC (Area Under Curve) of the Receiving Operation Curve (ROC) and the Sensitivity (number of correctly predicted essential genes/total essential genes) of the prediction. AUC and Sensitivity were then used to evaluate the performance of the model.EC is a gram-negative bacterium commonly found in the lower intestine of warm-blooded organisms. It is one of the most well-studied prokaryotic model organisms and has the best-characterized essential genes. We compared three approaches using our previously developed integrative algorithm (Table 2): (1) the same-organism approach, where we learned traits among the partially known essential genes in EC and predicted the rest of the essential genes; (2) the cross-organism approach, in which we learned traits among the known essential genes in AB, a closely-related model organism, and tried to predict the essential genes in EC; and (3) the combined approach, in which we learned traits among the known essential genes in AB as well as the partially known essential genes in EC and tried to predict the rest of the essential genes in EC. Because in our previous research we have shown that our cross-organism approach outperforms homology mapping [24], we did not compare homology mapping in this study.Summary of the three approaches (see Experimental Section for details).Among the total characteristic features that we considered, we have identified 13 that are potentially associated with gene essentiality in EC with relatively weak correlations among themselves (Table 1). Among these 13 features, we previously identified the domain enrichment score (DES) as the strongest [24], suggesting that gene essentiality is likely preserved through the function of protein domains or domain combinations rather than through the conservation of the entire genes. To show its efficiency, in our model construction process, we separated this dominant feature from the remaining 12 features. First, we used 12 features excluding DES to build the “no-DES” model. Next, we compiled the DES feature with the other features to form the “with-DES” model. We first built the “no-DES” model in EC (see Experimental Section). The 12 selected features were used as input variables in the logistic regression classifier. The classifier generated a probability score of essentiality for each gene of the entire target organism (both “gold standard” set and prediction set (Table 2)). Combining this probability score and the true essentiality information of each gene, we generated the ROC curve. The ROC was then evaluated by the AUC score. We gradually increased the size of known essential genes in our model. The result showed that the AUC score increased from 0.84 to 0.88 before the size of known essential genes reached 2% of the total genes in the genome. At this point, the model had already performed very closely to its optimal, achieving over 95% of its best performance. Beyond this point, the AUC score increased slowly from 0.88 to 0.89 even with a substantial increase of known essential genes (Figure 1a, red curve). Comparison of three approaches in EC. (a) The distribution of AUC along with the different sizes of known essential genes in EC: red curve: same-organism approach “with no-DES”; black curve: same-organism approach “with DES”; blue curve: combined approach; green curve: the DES feature only dashed line: cross-organism approach. The bar chart of the correctly classified essential genes among the top 400 predictions with respect to the different sizes of known essential genes in EC using (b) “no-DES” model; (c) “with-DES” model; and (d) combined model. The black bar shows the correctly classified essential genes in the “gold standard” set.Besides the AUC score, we were also interested in the number of genes we successfully classified. Using 10% as a cutoff, the top 400 genes with the largest probability scores were predicted as essential genes. Those 400 genes came from two parts, the “gold standard” set (Figure 1b, black bar) and the prediction set (Figure 1b, white bar). Figure 1b showed that the performance was nearly stable if the known essential genes took up more than 2% of the total genes in EC. Next, we compiled the DES feature with the other 12 features and built model by the same process used for the “no-DES” set. Compared with the “no-DES” model, the results were significantly improved (Figure 1a, black curves and Figure 1c). We can see that the AUC reached 0.94 if we knew about 2% of total genes to be essential. Figure 1c also suggested that the performance of the classification is stable if more than 2% of the total genes are known to be essential. They both decrease quickly as less essential information is given. We also applied our model only using the DES feature and compared the predictions with both the “no-DES” and “with-DES” sets (Figure 1a). The comparison showed that the DES alone is not enough to make optimal predictions, suggesting that including more features is necessary to achieve the optimal prediction performance. AB is a gram-negative bacterium commonly found in aquatic and soil environments. It belongs to the same class of gram-negative proteobacteria as EC. A set of 499 AB essential genes has been identified by targeted mutagenesis. We were able to use AB essential genes set to predict essential genes in EC [24], and the direct prediction yielded an ROC with the AUC score of 0.92. In Figure 1a, the dashed line shows the AUC of the prediction from AB, and the black curve dominates it when 1.5% of the total genes are known to be essential. This suggested that knowing 1.5% or more genes of the total genes to be essential in EC is sufficient to achieve a prediction better than transferring annotations from AB.Based on the above results, we had a new question: If we combine both AB and the fraction of known genes with essential information in EC as the new “gold standard” set and try to predict the rest of the essential genes in EC, could the result be significantly improved? To answer this question, we randomly chose a fraction of genes (we gradually increased the number of known genes from 10% to 90%) from EC and combined them with AB dataset (see Experimental Section). In the model training process, we assigned different weights to the two gene sets to obtain a more reliable result. Here, the partially known genes with essential information from EC have been set to have 4:1 weights vs. the AB genes. We trained the model on this combined “gold standard” set. Each time we also repeated the random process 200 times to estimate the variance. The results (Figure 1a, blue curve) showed that the combined approach outperformed the same-organism approach at the beginning. However, the black curve quickly outperformed the blue curve as the known essential genes in EC increased. The correctly predicted genes in Figure 1d also supported this result.In AB, we identified 11 features that are potentially associated with gene essentiality and have relatively weak correlations among themselves [24] (Table 1). We followed the same analysis procedure as in EC. In the same-organism approach, we first used 10 features excluding DES as the input of the classifier to build the “no-DES” model, and then including DES to build the “with-DES” model. The model generated a probability score of gene essentiality for each gene of the entire target organism. Combining this probability score and the true essentiality information of each gene, we were able to evaluate the performance. In Figure 2a, the red and black curves showed the distribution of the AUC scores of the results output from the “no-DES” and “with-DES” models respectively. Both curves increase rapidly before 2% (66/3308) of total genes are known to be essential, achieving more than 95% of the best performance. Compared with “no-DES” results, the results of “with-DES” were significantly improved. Also, the dashed line in Figure 2a shows the AUC of the cross-organism approach using EC essential genes, suggesting that knowing 2% of total genes to be essential is “sufficient” to lead to a prediction better than transfer from EC. Figure 2b and c show the bar charts of the correctly classified essential genes using the “no-DES” and “with-DES” models respectively. For AB, we adopted a similar percentage as the cutoff to predict essential genes as in EC, and the top 400 genes with the largest probability scores were predicted as essential genes. In both Figure 2b and 2c, the performance is nearly stable if the known essential genes take up more than 2% of the total genes in AB. In the combined approach, we combined both the EC essential genes with increasing numbers of known AB essential genes by assigning different weights. The blue curve (Figure 2a) shows the combined approach outperforming the same-organism approach only at the beginning. Compared with Figure 1, the difference between the combined approach and the same-organism approach in AB was less significant than in EC. The green curve in Figure 2a shows the performance of DES feature only. This suggests that the integration of different features is able to make more accurate predictions than using DES alone. Comparison of three approaches in AB. (a) The distribution of AUC along with the different sizes of known essential genes in AB: red curve: same-organism approach “with no-DES”; black curve: same-organism approach “with DES”; blue curve: combined approach; dashed line: cross-organism approach. The bar chart of the correctly classified essential genes among the top 400 predictions with respect to the different sizes of known essential genes in AB using (b) “no-DES” model; (c) “with-DES” model; and (d) combined model. The black bar shows the correctly classified essential genes in the “gold standard” set.Our results suggested that essential genes are highly predictable by learning the characteristics underlying gene essentiality in prokaryotes. To test whether the same trend can also be observed in eukaryotic species, we chose SC and NC as our test candidate species. SC is an important eukaryotic model organism in cell biology and is one of the most thoroughly studied eukaryotic microorganisms. There are 1049 essential genes identified by the systematic deletion project [26]. Using the same-organism approach in SC, we identified 14 features potentially associated with gene essentiality (Table 1). Domain enrichment score (DES) was found to be a strong feature in predicting essential genes in eukaryotes as well. This suggests that, much as in prokaryotes, gene essentiality in eukaryotes is likely preserved through the function of protein domains or domain combinations rather than through the conservation of entire genes. First, we used 13 features excluding DES as the input of the classifier. After the 10-fold cross-validation, each gene of the target organism received a probability score of essentiality. Combining this probability score and the true essentiality information of each gene, we were able to evaluate the performance. Figure 3a (red curve) showed the AUC curve of the “no-DES” results. It gradually increases along with the increase of the known essential genes and reaches stable at around 4% point on the x-axis, achieving 95% of the best performance. Besides the AUC curve, we also plotted the bar chart of correctly predicted essential genes (Figure 3b). Since essential genes comprise of about 20% of a eukaryotic genome, we used 1200 as the cutoff, i.e., the 1200 genes with the highest essential scores were predicted as SC essential genes. The performance increased as we increased the size of the training dataset, and the saturation point was at 4%. Figure 3a (green curve) shows that, similar to in prokaryotes, DES is a strong feature to the prediction of gene essentiality and incorporating it with other functional and genomics features is able to achieve an optimal performance. Comparison of three approaches in SC. (a) The distribution of AUC along with the different sizes of known essential genes in SC: red curve: same-organism approach “with no-DES”; black curve: same-organism approach “with DES”; blue curve: combined approach; dashed line: cross-organism approach. The bar chart of the correctly classified essential genes among the top 1200 predictions with respect to the different sizes of known essential genes in SC using (b) “no-DES” model; (c) “with-DES” model; and (d) combined model. The black bar shows the correctly classified essential genes in the “gold standard” set.Next, we added the DES feature into the model. Figure 3a (black curve) and Figure 3c show a similar trend, except the values are significantly higher than those of the “no-DES” results. This further supports the notion that the DES feature has strong power in predicting essential genes in eukaryotic species. Moreover, we note that the saturation occurred at 4% point in both figures. Thus, knowing 4% or more of the total genes is essential to building a reliable prediction. In the combined approach, we used both NC and the partially known essential genes in SC as the new training set. Would the result be significantly improved again? We followed the same scheme as described above. The results were consistent: As shown in Figure 3a, the performance of the same-organism approach (black curve) dominates the performance of the combined approach (blue curve) from about 1.5% on the x-axis. Although the saturation point of the prediction is different, the dominating points are almost the same as those in EC and AB. NC is an ascomycete, the red bread mold. Like all fungi, it reproduces by spores. It is used as a eukaryotic model organism because it is easy to grow and has a haploid life cycle which makes genetic analysis easier. There are 1250 essential genes in NC produced by the systematic gene deletion project. We identified 14 features potentially associated with gene essentiality in NC (Table 1). Following the same procedure as above, we analyzed the “no-DES” and “with-DES” dataset of the same-organism approach separately. We assigned the top 1500 genes as the predicted essential genes. Figure 4a shows that when given about 4% of total genes to be essential, the prediction achieves stable AUC with over 95% best performance. Compared with the red curve, the black curve is significantly improved. The blue curve also showed the performance of the combined approach using SC and partial NC known essential genes. The conclusion is similar to that in SC: The same-organism approach in NC (black curve) dominates the combined approach (blue curve) after at least 1.5% of the total genes are known to be essential.Comparison of three approaches in NC.  (a) The distribution of AUC along with the different sizes of known essential genes in NC: red curve: same-organism approach “with no-DES”; black curve: same-organism approach “with DES”; blue curve: combined approach; dashed line: cross-organism approach. The bar chart of the correctly classified essential genes among the top 1500 predictions with respect to the different sizes of known essential genes in NC using (b) “no-DES” model; (c) “with-DES” model; and (d) combined model. The black bar shows the correctly classified essential genes in the “gold standard” set.Our results suggest that, in prokaryotes, when the number of known essential genes is greater than 2% of total genes, it will achieve over 95% of the best performance, recovering >68% of total essential genes at the given cutoff. For example, for an understudied organism with 3000 genes, we need to know ~60 essential genes in order to accurately predict the majority of its ~300 essential genes. In contrast, in eukaryotes, achieving the same level of performance requires more than 4% of total genes, reflecting the increased complexity of eukaryotic organisms. The complexity comes from different aspects. One possibility is that eukaryotes have more complex genome structures than prokaryotes, such as the expanded protein domain repertoire. In fact, EC and AB contain 5468 and 4204 unique domains, respectively, while SC and NC contain 6023 and 7031 unique domains, respectively, according to the Interpro database. In addition, higher organisms have larger and more complex cellular structure as well as perform more diversified functions, which also require them to have more essential genes. We found that the required number of known essential genes was surprisingly small for both prokaryotes and eukaryotes, suggesting that the distribution of genomic features extracted from this small subset already provided a close approximation to the distribution of those extracted from the entire essential gene set. This showed the advantage of predicting essential genes using machine-learning approaches. We also noticed that as the model reaches saturation, there are still parts of essential genes (i.e., 32% in prokaryotes) that cannot be correctly predicted as essential. We further explored these incorrectly predicted essential genes by plotting the distributions of their associated features. Here we defined the essential genes that were correctly predicted as true positives (TPs) and those that were incorrectly predicted as false negatives (FNs). Figure 5 shows the boxplot of the two parts of genes in AB. The features for which the distributions between the two sets of genes differed most widely are DES and PHYS, followed by CAI, Nc and Aromo, all of which were derived from genomic sequences. This suggests that in order to correctly predict the FNs, relying on features based on genomic sequences is no longer enough. Other strong functional genomics features have to be discovered and incorporated into predictions. This observation also supports the notion that gene essentiality is likely determined not solely by genomic sequence, but by multiple aspects of biology, from sequence to function. The distribution of features among true positives (TPs) and false negatives (FNs) in AB.We then performed functional analysis of the FN genes by categorizing them according to the clusters of orthologous groups (COGs) proteins classification. In COGs, genes can be generally classified into four broad functional categories: information storage & processing, cellular processes & signaling, metabolism and poorly characterized. Previous work has shown that essential genes are overrepresented in the category of information storage and processing with basic cellular functions such as RNA processing and modification and DNA replication [32]. Essential genes involved in this category are often well conserved across species. On the other hand, the species-specific essential genes are mainly distributed in cellular processes and metabolic categories, which often reflects a microbe’s unique life style and living environment. Figure S1a and Figure S1b) show the distributions of FN genes across different functional categories in EC and SC respectively. We can see in EC the FN genes are enriched in the metabolic category while in SC these FN genes are enriched in cellular processes and signaling category.Comparing different sets of features used between the prokaryotes (EC, AB) and eukaryotes (SC, NC) in Table 1, the common features they shared are: Nc, L_aa, PHYS, PA, DES and FLU. These features cover all three categories described in Section 2.2. This supports our conclusion that the computational integration of different genomic and functional features is able to accurately predict essential genes in both prokaryotes and eukaryotes. However, there are some differences of features used between them, such as those sub-cellular localization features. For example, Nucleus, Plasma and PredHel are used only by SC and NC while Inner member is used only by EC and AB. These reflect the differences in cellular structure between prokaryotes and eukaryotes—the eukaryotic cells are much larger and more complex than prokaryotic cells.Through our analysis, we realize that the evolutionary distance between the understudied organism and the model organism may affect the thresholds observed in our study. Nevertheless, our results suggest that an organism’s own known essential genes usually contain more information about its unique physiology and are a better representative set of its total essential genes.Logistic regression was chosen in this study mainly because of its simplicity and ease of interpretation of results. Other machine-learning methods could have been used. However, most alternative techniques suffer from their own limitations, e.g., missing value problems or being prohibitively time-consuming, which prevent them from being used in this study. Nevertheless, we expect our conclusions from this investigation are unlikely to change if a different machine-learning technique is used. Since the four species we studied are all microorganisms, the conclusions from this study may not be applicable to more complex systems, such as mouse and human. Finally we believe the results obtained from our study provided important information on accurately predicting essential genes and will greatly facilitate the annotations of microbial genomes. In this study, we investigated the performance of three approaches for predicting essential genes under conditions where information on different numbers of known essential genes is given. Our results suggest that when determining the best strategy for predicting essential genes, unless the number of known essential genes is small, i.e., less than 1.5% of total genes, learning from the known essential genes in the target organism usually outperforms transferring essential gene annotations from a related model organism. This is consistent in both prokaryotes and eukaryotes. Moreover, when the known essential genes are few (i.e., <1.5% of total genes), and a closely related organism is available, combining these two sources of information results in a significantly increased performance over either the same-organism approach or the cross-organism approach. On the other hand, when the target organism has a sufficiently large number of known essential genes, combining the annotations from a model organism often results in a reduced performance as compared with using its own known essential genes, reflecting the slight differences of the underlying properties of essential genes between different organisms.The authors would like to thank four anonymous reviewers for valuable suggestions. LJL designed the research, JD and LT implemented the research. XL and YL offered critical suggestions. JD, XL and LJL contributed in writing and revising the article. This research was supported by Cincinnati Children’s Hospital Medical Center (CCHMC) Trustee Grant and a grant from The Midwest Center for Emerging Infectious Diseases (MI-CEID) awarded to LJL. To create a training dataset for our classifier, features are extracted where available for each ORF in each organism and annotated with known essentiality values from the essential gene datasets. Our study considers three main types of features: (A) those intrinsic to a gene’s sequence (e.g., GC content, length); (B) those derived from genomic sequence (e.g., localization signals and codon adaptation measures) and (C) experimental functional genomics data (e.g., gene-expression microarray data).A-1. Genomic sequence properties: We use CodonW (http://bioweb.pasteur.fr/seqanal/interfaces/codonw.html) to calculate the following properties associated with genomic sequences: Kyte and Doolittle’s grand average of hydropathicity (GRAVY) [1], protein length (amino acids), GC content, and two measures of codon usage: effective Nc [2,3] and CAI [4]. B-1. Predicted subcellular localization: We used the PA-SUB Server v2.5 to obtain these features [5]. Gram-negative bacteria (EC, PA and AB) have five predicted localizations: Inner membrane, Extracellular, Cytoplasm, Periplasm, Outer membrane. Gram-positive bacteria (BS) have three predicted localizations: Extracellular, Cytoplasm, Plasma membrane. B-2. Transmembrane helices for each ORF: The putative transmembrane helices are calculated by TMHMM Web server v2.0 [6,7].B-3. Evolutionary conservation of a gene: We used the RBH method to search orthologs in multiple complete genomes for each gene of the target organism (PA, EC, AB and BS). The number of genomes that have orthologous hits was used as a measure of evolutionary conservation of a gene. Such conservation has been shown to correlate well with the dispensability of a gene [8]. B-4. Paralogy: Duplicated genes in an organism are often referred to as paralogs. An all-against-all FASTA search was conducted for the whole set of ORFs in the target organism (PA, EC, AB and BS) to identify the paralogs with an E-value threshold of 10−20.B-5 Domain enrichment: For each individual domain, we collected its occurrence in each organism (PA, EC, AB and BS) using the Pfam database (http://pfam.sanger.ac.uk). Then we estimated the domain enrichment score according to the ratio of occurrence frequencies between essential gene sets and the total genes in the target organism: , here ness and nnon-ess represent a domain’s occurrence frequency in the essential and non-essential dataset, respectively. Ness and Nnon-ess is the size of the essential and non-essential dataset, respectively.C-1. Fluctuation in gene-expression: The mRNA expression levels of essential genes often vary, on average, within a narrower range, whereas the expression of nonessential genes fluctuates more widely [9]. Gene expression data in these bacteria were downloaded from NCBI GEO [10], ArrayExpress [11], as well as the gene-expression profiles of microarray data from Gasch et al. [12]. The variance of each gene was calculated from these gene expression profiles as a measure of the fluctuation of gene expression.C-2. Topology in gene co-expression network: From gene expression microarray data, a gene-expression cooperativity graph is constructed as Gg(D) = (Vg,Eg), with the vertex set  and the edge set  for  and | rij | ≥ 0.7. Each vertex represents a gene and each edge represents a gene pair whose gene expression profiles correlation coefficient | rij | is greater than 0.7. This cutoff value of | rij | is determined based on our previous work [13]. The hubs (nodes with high degrees) and bottlenecks (nodes with high betweenness or shortest paths occurrence) have been found to have correlations with gene essentiality [14]. The network statistics are calculated using tYNA (http://www.gersteinlab.org/tyna).Functional distribution of false negative genes according to the orthologous groups of proteins (COGs) classification in EC (a) and SC (b) respectively.

Med-MDPI/biomolecules/biomolecules-02-01-00023.txt

@@ -0,0 +1 @@
+Present address: Otsuka Maryland Medicinal Laboratories, Inc., 9900 Medical Center Drive, Rockville, MD 20850, USA.licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Hydrodynamic tail vein (HTV) delivery is a simple and rapid tail vein injection method of a high volume of naked plasmid DNA resulting in high levels of foreign gene expression in organs, especially the liver. Compared to other organs, HTV delivery results in more than a 1000-fold higher transgene expression in liver. After being bitten by malaria-infected mosquitoes, malaria parasites transiently infect the host liver and form the liver stages. The liver stages are known to be the key target for CD8+ T cells that mediate protective anti-malaria immunity in an animal model. Therefore, in this study, we utilized the HTV delivery technique as a tool to determine the in vivo cytotoxic effect of malaria antigen-specific CD8+ T cells. Two weeks after mice were immunized with recombinant adenoviruses expressing malarial antigens, the immunized mice as well as naïve mice were challenged by HTV delivery of naked plasmid DNA co-encoding respective antigen together with luciferase using dual promoters. Three days after the HTV challenge, non-invasive whole-body bioluminescent imaging was performed. The images demonstrate in vivo activity of CD8+ T cells against malaria antigen-expressing cells in liver.In a complex plasmodial cycle, the liver stage, in which parasites reside in the liver, represents an important stage for the cell-mediated immunity to take place in the host. The role of CD8+ T cells in protective immunity against the liver stages has been well established in a rodent model. Evidence from in vivo depletion of CD8+ T cells in mice clearly demonstrated its protective role against the liver stages [1,2]. Furthermore, adoptive transfer studies corroborated the findings of protective CD8+ T cell functions [3,4]. Finally, we have previously shown that a single immunizing dose of a recombinant adenovirus (rAd) expressing a circumsporozoite (CS) antigen of Plasmodium yoelii, AdPyCS, induced a robust PyCS-specific CD8+ T cell response; moreover, a strong protective anti-malaria immunity, which is mediated by CD8+ T cells [5]. We have chosen two pre-erythrocytic antigens, as model antigens, to test the in vivo function of CD8+ T cells in the current study. The CS protein, which is a major surface protein of malarial sporozoite, has been well-characterized and shown to mediate protective immunity against malaria [3,4,5,6]. CS protein-based vaccine, called RTS,S is undergoing Phase III trial [7]. Cell traversal protein of Plasmodium ookinetes and sporozoites (CelTOS), another pre-erythrocytic antigen, is shown to be recognized by T cells of at least 50% of human volunteers immunized with irradiated sporozoites of P. falciparum [8]. This CelTOS is a microneme protein secreted by sporozoites and shown to mediate protective anti-malaria immunity in a mouse model [9].After the gene expression in muscle was observed upon intramuscular injection of naked plasmid DNA [10], a non-viral delivery of nucleic acids by injecting a large volume of solutions was performed [11,12]. Then, a simple technique called Hydrodynamics-based gene transfection was developed in the late 90s [13,14]. Using this technique, a rapid tail vein injection of a high volume of naked plasmid DNA was performed, leading to high levels of foreign gene expression in organs, especially the liver. This delivery method, called Hydrodynamic Tail Vein (HTV) delivery, is simple and achieves 40% of liver transfection [14]. In our studies, we established a tool to measure the in vivo cytotoxic effect of malaria specific CD8+ T cells using a Hydrodynamic Tail Vein (HTV) injection. For this purpose, mice were first immunized with rAd expressing malaria antigens, described above, to mount malaria-specific CD8+ T cells. A group of immunized mice, as well as naïve mice, then received plasmid DNA encoding respective antigens together with luciferase gene through HTV delivery. Finally, all the mice challenged with the DNA by HTV delivery were subjected to non-invasive whole body bioluminescent imaging to determine the level of luciferase expression in their livers and assess the function of malaria-specific CD8+ T cells response in vivo. The genes coding for P. yoelii CS protein and P. yoelii CelTOS protein were first codon optimized, as shown in Figure 1, and then linked to the luciferase gene using a linker (sequence: PGILASQSTCRHASLRPIQ). The constructs were then amplified and cloned into a vector plasmid, pCMV-MCS (Agilent Technologies, Stratagene Products Division, La Jolla, CA, USA). The final constructs were verified by sequencing. The plasmid DNA, named DNAPyCS-Luc and DNAPyCelTOS-Luc, were purified with a Midi purification kit (Qiagen, Valencia, CA, USA). Codon optimized sequences of PyCS and PyCelTOS.Recombinant Adenoviruses (rAds) expressing P. yoelii CS protein, AdPyCS, and P. yoelii CelTOS protein, AdPyCelTOS, were generated previously [15]. Briefly, after both PyCS and PyCelTOS genes were codon optimized (Figure 1), the optimized fragments were cloned into a shuttle vector, pShuttle-CMV5, and the PmeI linearized shuttle vector was introduced into E. coli strain of BJ5183 that harbored the adenoviral backbone vector, pAdEasy-1 (Agilent Technologies, Stratagene Products Division, La Jolla, CA, USA). Recombinant Ad plasmids were transfected into AD-293 cells (Stratagene, Cedar Creek, TX, USA) to generate rAds. Finally, rAds were amplified and subsequently purified by CsCl gradient ultracentrifugation, previously described [16]. Virus particle (v.p.) was calculated based on O.D.260 (1 OD260 = 1.25 × 1012 v.p./mL).Eight-ten weeks old female BALB/c mice were obtained from Taconic Farms (Hudson, NY, USA). Animals were maintained in the Laboratory Animal Research Center of the Rockefeller University following standard provisions. The animal protocols, #8065 and #10095, were approved by the Institutional Animal Care and Use committee at the Rockefeller University. Group of three mice were immunized intramuscularly with 1 × 1010 virus particles (v.p.) of AdPyCS or AdPyCelTOS. Two weeks later, the mice received the plasmid DNA by HTV injection in less than 5 seconds after the plasmid DNA was diluted in PBS in a total volume of 2 mL. Group of three mice were administered with monoclonal antibodies against CD8+ T cells (YTS 169) (Harlan Bioproducts For Science Inc, Madison, WI, USA) to deplete CD8+ T cell from previously immunized mice. Briefly, 500 μg of YTS 169 diluted in PBS was injected intra-peritoneally at 3 and 1 day prior to the DNA challenge. We confirmed that this anti-CD8+ T cell antibody administration regimen resulted in the depletion of more than 95% of CD8+ T cells among splenocytes by FACS analysis (data not shown). Three days after the HTV injection, the images of the luciferase expression in mouse liver was monitored using Caliper Life LifeSciences IVIS®Lumina/Living Image (Caliper LifeScience, Hopkinton, MA, USA). Briefly, after anesthetizing the mice, 200 μL of 15 mg/mL D-luciferin (Gold Biotechnology, St Louis, MO, USA) was injected intra-peritoneally, and the whole body in vivo imaging analysis was performed for 30 sec to 2 min, using in vivo imaging system (IVIS®Lumina). Luciferase expression data were then quantified using the Living Image software (Caliper LifeScience) in a fixed region of interest (ROI) in terms of photons/sec/cm2/sr. Statistical analysis of experimental and control data was evaluated by Student’s t-test. A value of P < 0.01 was considered statistically significant.In order to determine the level of the luciferase expression in the liver after HTV injection of different doses of plasmid DNA, we injected various doses, 2 μg, 10 μg and 50 μg, of plasmid DNA co-encoding PyCS and luciferase (DNAPyCS-Luc), or plasmid DNA co-encoding PyCelTOS and luciferase (DNAPyCelTOS-Luc). Three days post HTV injection, we performed a non-invasive whole body bioluminescent imaging using IVIS (Caliper LifeSciences) and found that the HTV injection of 50 μg of DNAPyCS-Luc (Figure 2a and 2c), as well as 10–50 μg of DNAPyCelTOS-Luc (Figure 2b and 2d) induced the highest level of luciferase expression in the liver. It is noteworthy that the luciferase expression after DNAPyCS-Luc challenge is lower than that after DNAPyCelTOS-Luc challenge. This difference may be due to two reasons. Firstly, the translation efficiency of the PyCS coding region may be decreased compare to that of PyCelTOS, due to the unstable nature (it contains a highly repetitive sequence in the middle) and the PyCS gene is twice the size of the PyCelTOS gene. Secondly, luciferase is fused with PyCS antigen with a short linker, therefore the structure of PyCS may possibly affect the level of luciferase expression. In fact, we observed a lower expression of PyCS protein compared to PyCelTOS protein upon in vitro transfection of the corresponding plasmids. Nevertheless, in the subsequent experiments, we decided to choose 30 μg and 3 μg of DNAPyCS-Luc and DNAPyCelTOS-Luc, respectively, for the HTV injection. Luciferase expression in the liver after Hydrodynamic tail vein (HTV) injection of various doses of each plasmid DNA, co-encoding genes for a malaria antigen and luciferase. (a) and (b) Non-invasive bioluminescence imaging depicts the luciferase expression in the liver after HTV injection of 2, 10, and 50 μg of plasmid DNA co-encoding genes for PyCS antigen and luciferase in (a) and plasmid DNA encoding genes for PyCelTOS antigen and luciferase in (b). Three day after HTV injection, mice were anesthetized and injected with D-luciferin, and the luciferase intensity was optically imaged. The numbers below indicate the bioluminescent signal intensity in the region of interest (ROI), quantified as photons/sec/cm2/sr. (c) and (d) The graphs show the bioluminescent signal intensity in the ROI, as calculated by photons/sec/cm2/sr, of the same mice observed in (a) and (b), respectively.The level of luciferase expression upon HTV injection of DNAPyCS-Luc was determined in mice immunized with AdPyCS or AdPyCelTOS, compared to those in naïve mice. For this purpose, we immunized a group of mice with AdPyCS or AdPyCelTOS, and 2 weeks later, we challenged the immunized mice, as well as naïve mice, with HTV injection of DNAPyCS-Luc. Non-invasive bioluminescent images have shown the complete inhibition of luciferase expression in the liver of a group of mice receiving a single immunizing dose of AdPyCS, but not AdCelTOS (Figure 3a and 3b). This indicates that the level of luciferase expression induced by DNAPyCS-Luc was inhibited by an antigen-specific fashion. Luciferase expression in AdPyCS-immunized mice, AdPyCelTOS-immunized mice, or naïve mice, upon HTV injection of DNAPyCS-Luc. (a) Noninvasive bioluminescence image shows the inhibition of DNAPyCS-Luc induced luciferase expression in mice by a prior single immunizing dose of AdPyCS, but not of AdPyCelTOS. Three day after the HTV injection with DNAPyCS-Luc, mice were anesthetized and injected with D-luciferin, and the luciferase intensity was optically imaged; (b) Quantification of the bioluminescent signal intensity in AdPyCS-immunized mice, AdPyCelTOS-immunized mice, or naïve mice, upon HTV injection of DNAPyCS-Luc. In order to determine whether the antigen-specific inhibition of the level of luciferase expression is unique to the PyCS antigen, we also determined the luciferase expression upon HTV injection of DNAPyCelTOS-Luc in the second set of experiments. Briefly, we first immunized a group of mice with AdPyCS or AdPyCelTOS, and two weeks later, we challenged these mice, as well as naïve mice, with HTV injection of DNAPyCelTOS-Luc. In corroboration with our results from HTV injection of DNAPyCS-Luc (Figure 3), only a group of mice receiving a single immunizing dose of AdPyCelTOS, but not AdPyCS, could inhibit the level of luciferase expression induced by DNAPyCelTOS-Luc, as shown in Figure 4a and 4b. Thus, the inhibition of luciferase expression induced by plasmid DNA coencoding malaria antigen and luciferase is due to the immune response elicited by immunization of a rAd expressing the same antigen. Luciferase expression in AdPyCS-immunized mice, AdPyCelTOS-immunized mice, or naïve mice, upon HTV injection of DNAPyCelTOS-Luc. (a) Noninvasive bioluminescence image shows the inhibition of DNAPyCelTOS-Luc induced luciferase expression in mice by a prior single immunizing dose of AdPyCelTOS, but not of AdPyCS. Three days after the HTV injection with DNAPyCelTOS-Luc, mice were anesthetized and injected with D-luciferin, and the luciferase intensity was optically imaged; (b) Quantificationof the bioluminescent signal intensity in AdPyCS-immunized mice, AdPyCelTOS-immunized mice, or naïve mice, upon HTV injection of DNAPyCelTOS-Luc.In this study, we have shown that a single immunization of rAd expressing a malarial antigen could inhibit the level of luciferase expression in the liver of mice upon HTV injection with a plasmid DNA co-encoding the antigen and luciferase (Figure 3 and Figure 4). We have previously shown that a single immunizing dose of AdPyCS could elicit a robust CD8+ T cell response that can attack the liver stages of rodent malaria [5]. Therefore, we hypothesized that the inhibition of luciferase expression induced by the HTV injection with the DNA co-encoding the malaria antigen and luciferase could be due to malaria-specific CD8+ T cells elicited by the immunization with rAd expressing the same malaria antigen. To determine if this is the case, we depleted the CD8+ T cell subset from mice immunized with AdPyCS or AdPyCelTOS prior to challenge with DNAPyCS-Luc and DNAPyCelTOS-Luc by HTV injection, respectively. As shown in Figure 5a and 5b, the inhibition of luciferase expression in the liver of DNAPyCS-Luc-challenged mice observed after the AdPyCS immunization was completely abolished and the luciferase expression was recovered in mice depleted of CD8+ T cells. Similarly, the inhibition of luciferase expression induced by the HTV challenge with DNAPyCelTOS-Luc observed in mice immunized with AdPyCelTOS was abrogated by the CD8+ T cell depletion. These results clearly indicate that malaria antigen-specific CD8+ T cells induced by immunization with rAd expressing the antigen are responsible for inhibiting the luciferase expression in the liver of mice challenged with HTV injection of DNA co-encoding the antigen and luciferase. CD8+ T cell-mediated inhibition of luciferase expression in mice upon HTV injection of plasmid DNA, co-encoding genes for a malaria antigen and luciferase, by immunization with rAd expressing the antigen. (a) CD8+ T cell population was depleted from malaria vaccine-immunized mice before HTV injection with plasmid DNA co-encoding a malaria antigen and luciferase, and luciferase expression was assessed by noninvasive bioluminescence imaging; (b) The bioluminescent signal intensity was quantified in various groups of mice upon HTV injection of plasmid DNA co-encoding a malaria antigen and luciferase. The groups include naïve mice, malaria vaccine-immunized mice, and malaria vaccine-immunized mice that were depleted of CD8+ T cells in vivo.A non-viral gene delivery platform has been shown to be the simple and efficient method of gene expression in various organs using rodent model [10,11,12,13,14,17,18]. With the novel concept of hydrodynamics-based delivery [13,14], the transgene expression has been observed in liver, lung, heart, kidney and also various species [17]. However, the highest gene expression has been achieved only in liver [13,14]. The hepatic transfer of DNA via HTV injection is a physical process, called hydroporation, in which membrane pores are generated by highly pressured solution in liver [17,18]. Hepatocytes are known to be non-professional antigen-presenting cells (APCs), and since hepatocytes do not express sufficient co-stimulatory molecules, they are not efficient at priming and inducing CD8+ T cells like professional APCs, including Kupffer cells and dendritic cells. However, when hepatocytes are infected with hepatotrophic virus, such as HCV, they can act as a target for the CD8+ T cells [19]. In fact, both human and murine CD8+ T cells are shown to recognize endogenously synthesized and processed virus proteins in association with MHC-class I molecules, and eliminate virus-infected hepatocytes [20,21]. Therefore, having taken advantage of the natures of hepatocytes that are capable of processing and presenting CD8+ epitopes and being recognized by epitope-specific CD8+ T cells, we decided to deliver foreign genes that carry CD8+ T cell epitopes into hepatocytes in vivo by the HTV administration.In this study, we have utilized this HTV delivery technology to express a high level of luciferase together with a malaria antigen of interest in the mouse liver. Then, upon a single immunizing dose of rAd expressing a malaria antigen, an immunization regimen known to elicit a robust CD8+ T cell response [5], we examined the in vivo function of malaria antigen-specific CD8+ T cells by the inhibition of the luciferase expression in the liver, as measured by a non-invasive whole body bioluminescent imaging analysis.We found that a single immunizing dose of AdPyCS and AdPyCelTOS could almost completely inhibit the expression of luciferase in the mouse liver, induced upon HTV delivery of DNAPyCS-Luc and DNAPyCelTOS-Luc, respectively. Furthermore, we were able to determine that the inhibition of luciferase expression by prior immunization with AdPyCS or AdPyCelTOS was mainly due to CD8+ T cells, as the depletion of this T cell subset in vivo abolished the inhibition of luciferase expression. This is an interesting finding in view of our recent observation that although a single immunizing dose of both AdPyCS and AdPyCelTOS could induce a high level of antigen-specific CD8+ T cell response, only PyCS-specific CD8+ T cell response, but not PyCelTOS-specific CD8+ T cell response, was able to inhibit the parasite load in the liver of malaria-challenged mice [15]. Our current results indicate that PyCelTOS-specific CD8+ T cells induced by AdPyCelTOS immunization is indeed functional, and that the failure of PyCelTOS-specific CD8+ T cells to attack the liver stages of malaria may be due to other factor(s) rather than the function of the CD8+ T cells. In order for antigen-specific CD8+ T cells to recognize and kill the infected hepatocytes, efficient processing and presentation of malaria antigens onto MHC class I by malaria parasite-infected hepatocytes is necessary. Hence, we speculate that after challenge with sporozoites, PyCelTOS antigen is not sufficiently processed and presented by MHC class I molecules expressed on the infected hepatocytes, thereby preventing CD8+ T cells from attacking liver stage malaria parasites. Alternatively, the liver stages of malaria parasites may simply express a lower amount of PyCelTOS. It is important to clarify this issue in the future studies. We believe that our current study demonstrates the successful utilization of a non-viral gene delivery platform that has led to the establishment of a novel method that can assess the in vivo function of antigen-specific CD8+ T cells in a mouse model. Because of the handling of a non-viral gene, it is a rather simple and safe method, which should be readily applicable for the in vivo assessment of the function of antigen-specific CD8+ T cells, not only limited to the liver stages of malaria parasites, but also other liver-specific pathogens and cancers. The authors declare no conflict of interest.This work was supported by a grant from NIH R56 AI073658, and Otsuka Pharmaceutical Co. Ltd.

Med-MDPI/biomolecules/biomolecules-02-01-00034.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Oligosaccharides, sequences of carbohydrates conjugated to proteins and lipids, are arguably the most abundant and structurally diverse class of molecules. Fucosylation is one of the most important oligosaccharide modifications involved in cancer and inflammation. Recent advances in glycomics have identified several types of glyco-biomarkers containing fucosylation that are linked to certain types of cancer. Fucosylated alpha-fetoprotein (AFP) is widely used in the diagnosis of hepatocellular carcinoma because it is more specific than alpha-fetoprotein. High levels of fucosylated haptoglobin have also been found in sera of patients with various carcinomas. We have recently established a simple lectin-antibody ELISA to measure fucosylated haptoglobin and to investigate its clinical use. Cellular fucosylation is dependent upon fucosyltransferase activity and the level of its donor substrate, guanosine diphosphate (GDP)-fucose. GDP-mannose-4,6-dehydratase (GMDS) is a key enzyme involved in the synthesis of GDP-fucose. Mutations of GMDS found in colon cancer cells induced a malignant phenotype, leading to rapid growth in athymic mice resistant to natural killer cells. This review describes the role of fucosylated haptoglobin as a cancer biomarker, and discusses the possible biological role of fucosylation in cancer development.Glycosylation is involved in a variety of biological phenomena including birth, differentiation, growth, inflammation, carcinogenesis, and cancer metastasis. Each oligosaccharide structure consists of several kinds of monosaccharides, the addition of which is catalyzed by specific glycosyltransferases. Among approximately 10 kinds of oligosaccharide modifications, fucosylation isone of the most important types in cancer. Hakomori et al. first reported the involvement of fucosylation in cancer, comparing the fucosylation levels of glycolipids in hepatoma cells and normal hepatocytes [1]. Fucose (6-deoxy-L-galactose) is a monosaccharide that is found in glycoproteins and glycolipids present in vertebrates, invertebrates, plants, and bacteria. Fucosylation, which comprises the transfer of a fucose residue to oligosaccharides and proteins, is regulated by many types of molecules, including fucosyltransferases, guanosine diphosphate (GDP)-fucose synthetic enzymes, and GDP-fucose transporter(s). Fucosylation levels in normal liver and colon are relatively low, but increase during carcinogenesis. Most target glycoproteins undergoing fucosylation are secretary proteins or membrane proteins on the cell surface. Increased fucosylated proteins in sera of patients with cancer are dependent on cellular fucosylation of cancer tissues and/or changes in fucosylation states in the liver. Glycomics, the systematic study of glycans and glycan-binding proteins in various biological systems, is an emerging field in the post-genomic and post-proteomic era. Change in fucosylation of glycoproteins such as fucosylated alpha-fetoprotein (AFP) is one of the most representative types of glycan-related cancer biomarkers. Increases in fucosylated AFP in sera of patients with hepatocellular carcinoma (HCC) was reported by Breborowicz et al. [2,3]. They characterized microheterogeneity of AFP in several liver conditions, and found increases in α1-6 fucosylation (core fucosylation) of AFP using lectin affinity electrophoresis [4,5]. AFP is a well-known tumor marker for HCC, but sometimes also increases in benign liver diseases such as chronic hepatitis and liver cirrhosis. In contrast, AFP with core fucosylation is a very specific marker for HCC [6,7]. AFP with core fucosylation was known as AFP-L3, because it was detected as the L3 fraction on Lensculinaris agglutinin (LCA) lectin electrophoresis. Core fucosylation involving attachment of fucose to the innermost N-acetylglucosamine in N-glycans is synthesized by α1-6 fucosyltransferase (Fut8) [8]. However, expression of Fut8 is increased in both HCC tissues and surrounding tissues with liver cirrhosis [9]. Therefore, complicated molecular mechanisms might exist in the production of fucosylated glycoproteins in cancer. There have been a number of methods to measure AFP-L3. The second generation AFP-L3 assay was based on liquid phase binding of the AFP-L3 glycoform with LCA and two specific monoclonal antibodies [10]. More recently, an automated immunoassay system for AFP-L3 has been developed [11]. This assay was based on an on-chip electrokinetic reaction and separation by affinity electrophoresis. It showed higher sensitivity than conventional methods. Since the limit of detection is 0.1 ng/mL AFP, the assay is available for early HCC detection. Block et al. have reported certain types of novel fucosylated glycol-biomarkers for HCC [12]. In a blind test, AFP had a better area under the receiver operating characteristic (ROC) curve than Des-gamma carboxy-prothrombin (DCP) and AFP-L3 in a total of 836 patients with chronic liver diseases and HCC [13]. Although AFP is one of the best standards for detection of HCC, AFP has limited utility for detecting HCC [14], suggesting that a combined assay of AFP, DCP, and AFP-L3 would be recommended as a follow-up for patients with chronic liver diseases. Loss of fucosylation has significant biological consequences. Lack of IgG core fucosylation results in 50–100 times higher activity of antibody-dependent cellular cytotoxicity [15]. These data have been confirmed by several groups, and as a result, genetically modified antibodies are now used in clinical trials. Complete loss of fucosylation was found in colon cancer cell line, HCT116 [16]. The loss was due to a GDP-mannose-4,6-dehydratase (GMDS) mutation, resulting in production of undetectable levels of GDP-fucose, a donor substrate for fucosyltransferases. Similar genetic mutations of GMDS were found in various cancer cell lines and human cancer tissues.However, FX (GDP-4-keto-6-deoxy-mannose-3,5-epimerase-4-reductase) deficient mice showed severe defects in the immune system and died shortly after birth [17]. Both FX and GMDS are rate-limiting enzymes in the de novo pathway producing GDP-fucose. Theoretically, FX deficient mice should show more severe abnormalities than Fut8 deficient mice, which lack only core fucose and not total fucose. However, HCT116 cells can grow rapidly under normal conditions, when growth factor receptors in the cells lack fucosylation. Therefore, there may be many genetic mutations which affect the signaling pathway of growth factor receptors in HCT116 cells. For example, autophosporylation of growth factor receptors without ligand stimulation might exist in these cells. In this review, we describe novel types of fucosylated cancer biomarkers, possible mechanisms for the production of fucosylated proteins, and biological functions of fucosylation and defucosylation.Fucosylated haptoglobin (Fuc-Hpt) was first found in sera of patients with advanced ovarian cancer and breast cancer [18,19]. Ulex europaeus agglutinin (UEA) lectin, which mainly recognizes α1-2 fucose residues, was used to detect Fuc-Hpt. A recent study showed that Fuc-Hpt, present in sera of patients with pancreatic cancer, involved the addition of fucose residues through the α 1-3/1-4 linkage [20]. We found Fuc-Hpt in sera of patients with pancreatic cancer as shown in Figure 1. In addition, fucosylated glycoproteins are recognized by several types of lectins. These lectins include Aleuria aurantia lectin (AAL), UEA, LCA, and Aspergillus oryzae lectin (AOL). AAL recognizes α1-3/α1-4 and α1-6 fucose, UEA recognizes α1-2 fucose, LCA recognizes the native form of α1-6 fucose with a mannose arm, and AOL recognizes a specific form of α1-6 fucose [21]. Recently, a more specific lectin for α1-6 fucose, called Pinellia ternata lectin or PTL, has been isolated from mushrooms (submitted for publication). This lectin could assist in cancer diagnosis. Western blotting of serum samples from patients with pancreatic cancer, using the AAL lectin, showed that a protein of approximately 40 kDa was highly fucosylated. The N-terminal sequence revealed that this protein was the haptoglobin β chain [22]. The fucosylated haptoglobin was found in 60–80% of the patients, and the prevalence increased progressively with stage of the disease. Increased fucosylated haptoglobin levels have been observed in several types of cancer (20–40%). Haptoglobin is produced in the liver and exhibits a low level of fucosylation, since the expression of Fut8 and GDP-fucose synthesis enzymes such as FX and GMDS is quite low in the normal liver. The ectopic expression of haptoglobin is observed in special conditions such as infections, inflammation and cancer. Fucosylated haptoglobin is a novel cancer biomarker for differential diagnosis and predicted prognosis. (A) Lectin blot using aleuria aurantia lectin (AAL) detected a protein of approximately 40 kDa from sera of patients with pancreatic cancer. Coomassie Brilliant Blue staining showed no changes in levels of this protein. This figure is cited from reference [22] with slight modification; (B) Establishment of lectin-antibody ELISA kit to measure Fuc-Hpt. Schematic system is shown; (C) Representative data of the Fuc-Hpt ELISA kit. Seventy-two cases of patients with pancreatic cancer and 22 healthy volunteers were assayed with 25 times dilution of sera. This data is cited from reference [28] with slight modification; (D) Combination assay of Fuc-Hpt and carcinoembryonic antigen is a marker for poor prognosis in patients with colorectal cancer after operation. This data is cited from reference [23] with slight modification.Where is fucosylated haptoglobin produced in patients with pancreatic cancer? A pancreatic cancer cell, PSN-1, expresses haptoglobin mRNA and produces fucosylated haptoglobin in conditioned medium. However, this situation is rare in comparison with the prevalence of fucosylated haptoglobin (60–80%). To answer this question, we transplanted a colon cancer cell line, HCT 116 which lacks fucosylation due to GMDS mutation, into athymic mice and investigated serum levels of Fuc-Hpt during tumor development [23]. HCT 116 cells were studied in two different pathways by using intrasplenic and subcutaneous injections. All animals after intrasplenic injection were positive for Fuc-Hpt with macro/micro liver metastasis. In contrast, two of six mice injected subcutaneously were positive for Fuc-Hpt. The positive animals exhibited intraperitoneal invasion of tumor cells. These data suggested that metastatic lesions of cancer could be one of the major sources of Fuc-Hpt production. In the case of HCT116 cells, cancer cells themselves cannot produce Fuc-Hpt, indicating that their surrounding normal cells such as hepatocytes and lymphocytes produced Fuc-Hpt. After we reported that Fuc-Hpt is a promising biomarker for pancreatic cancer, many researchers have reported increases in Fuc-Hpt in sera of patients with different cancers such as colon cancer [24], liver cancer [25], lung cancer [26], and pancreatic cancer [27].When evaluating Fuc-Hpt using Western blotting and AAL lectin, two problems occur in examining many samples at the same time. The first is whether fucosylated 40 kDa protein is really haptoglobin and the second is the difficulty in quantitating levels of Fuc-Hpt. To overcome these obstacles, we established a lectin-antibody ELISA to measure Fuc-Hpt. Since most oligosaccharides on IgG are fucosylated, the Fab fragment of IgG is used in this ELISA, after treatment with papain to remove the Fc part of IgG. Detailed procedures have been previously described [28]. When the conditioned medium from PK8 pancreatic cell line, transfected with an expression vector of human haptoglobin, was used for the standard curve of lectin-antibody ELISA, Fuc-Hpt was quantitatively measured in relative units. The cut-off index of Fuc-Hpt was 539 unit/mL, according to the ROC curve, and the sensitivity and the specificity of Fuc-Hpt for diagnosis of pancreatic cancer were 70% and 77%, respectively. In addition, a few cases of pancreatic cancer showed high levels of Fuc-Hpt, even at an early clinical stage. There is a possibility that these cases involved micrometastasis of the liver or lymph node. It is difficult to correlate Fuc-Hpt levels and the prognosis after operation, because the prognosis of pancreatic cancer is quite poor. Therefore, we compared the level of Fuc-Hpt before the operation with the prognosis after the operation in 63 cases of colorectal cancer [23]. An assay using Fuc-Hpt and carcinoembryonic antigen (CEA), a conventional cancer biomarker for colorectal cancer, was a marker for poor prognosis. Both CEA and Fuc-Hpt positive cases showed a poorer prognosis than CEA or Fuc-Hpt positive or both CEA and Fuc-Hpt negative. This data could be directly applied to clinical laboratory examinations. The reason for poor prognosis of colorectal cancer, which produces Fuc-Hpt, could be the micrometastasis as suggested from the results of experiments using HCT116 cells as described above.As mentioned in the introduction, an increase in cellular fucosylation is dependent on the induction of several glycosyltransferases, GDP-fucose production, and up-regulation of the GDP-fucose transporter. Although expression of these parameters is up-regulated in liver cirrhosis, serum fucosylated proteins are not dramatically increased as in HCC. Fucosylated AFP in HCC is an example. While the GDP-fucose level is up-regulated in HCC tissues compared to surrounding tissues, the level was only 2–3 times higher [29]. Another mechanism might exist to increase fucosylated AFP in serum of HCC patients.Fucosylated glycoproteins produced in hepatocytes are secreted into the bile, which is on the apical side of hepatocytes [30]. When oligosaccharide structures of bile and serum glycoproteins were compared using lectin blotting or 2D mapping, dramatic increases in fucosylation were observed for bile glycoproteins. In the human liver, Fut6 is involved in the synthesis of Lewis types of fucosylation, and hepatic glycoproteins with this oligosaccharide structure are present in the bile. In mice, Fut6 is a pseudo-gene, and the secretion of certain kinds of hepatic glycoproteins into the bile is disrupted in Fut8 knockout mice. The Fut8 knockout mice show decreased levels of hepatic glycoproteins such as α1-acid glycoprotein and α1-antitrypsin in their bile, suggesting that fucosylation regulates the secretion of certain types of hepatic glycoproteins into the bile. The disruption of this system could be one of the mechanisms underlying the increases in fucosylated protein levels, including AFP-L3 in the serum of patients with HCC. It is possible that HCC cells lose their polarity because they are rapidly proliferating. There are often no bile duct structures in HCC tissues. Therefore, selective secretion of fucosylated glycoproteins was not observed in HCC tissues, which led to the production of AFP-L3. AFP-L3 has also been detected in severe acute hepatitis [31]. These hypotheses were confirmed, using a rodent model of hepatocarcinogenesis [32]. Fucosylation of glycoproteins in bile is increased with progression of the liver disease. This could be due to an inflammation. A representative cytokine involved in inflammation, IL6, induces expression of fucosylation regulatory genes [33]. In contrast, serum fucosylation is not significantly increased in chronic liver diseases but increased in HCC. Interestingly, target glycoproteins for fucosylation in HCC are limited in sera and bile, suggesting that fucosylation is a signal for secretion into bile and this selective secretion is dependent on the characteristics of each glycoprotein. The structure as well as the numbers of oligosaccharides attached to proteins might be key factors for selective secretion via glycosylation. These data in part were confirmed with experiments using a human hepatoma cell line, HepG2, which has cellular polarity [34].While many studies have revealed that fucosylation is closely associated with cancer biology through modulation of signal transduction and the cell-cell adhesion pathways, we recently provided new evidence that fucosylation affects tumor immune surveillance via another signaling pathway, tumor necrosis factor-related apoptosis-inducing ligand (TRAIL) signaling [16]. When we examined the global fucosylation level in several colon cancer cells using AAL, which recognizes fucosylated oligosaccharides, little binding to AAL lectin was found in a human colon cancer cell line, HCT116. Further analysis revealed that HCT116 cells had a deleted GMDS transcript, which eliminated their ability to synthesize GDP-fucose, and resulted in an almost complete absence of fucosylation (Figure 2A). Transfection of the wild-type GMDS gene into HCT116 cells restored the cellular fucosylation. GMDS-rescued cells showed dramatically suppressed tumor formation and metastasis compared with mock cells when they were inoculated into athymic nude mice (Figure 2B). Depletion of natural killer (NK) cells stimulated tumor growth of the GMDS-rescued cells, but not that of the mock cells, indicating that a deficiency of fucosylation leads to escape from NK cell-mediated tumor immune surveillance. TRAIL is expressed mainly on the surface of immune cells, where it functions in T-cell homeostasis and NK cell-mediated killing of virally infected or oncogenically transformed cells [35,36]. Binding of TRAIL receptors by the ligand leads to apoptosis through a specific signaling cascade. Subsequent studies revealed that the GMDS-rescued cells were significantly more susceptible to TRAIL-induced apoptosis, which resulted in increased sensitivity of GMDS-rescued cells towards NK cells (Figure 2C). Aberrant transcripts of the GMDS gene were found in three other cancer cell lines as well as several colon and ovarian cancer tissues. Thus, loss of GMDS might be a common mechanism for cancer cells to evade TRAIL-mediated killing. While the increase in fucosylation is important at an early stage of carcinogenesis, defucosylation through genetic mutation in certain types of advanced cancers would lead to escape from NK-cell mediated tumor surveillance, and the acquisition of more malignant characteristics because of their ability to kill cancer cells. Optimized soluble recombinant human TRAIL or agonistic antibodies targeting TRAIL receptors are undergoing phase I or II clinical evaluation as promising proapoptotic antitumor therapeutic agents in patients with several types of tumors [37]. However, it has now become clear that many types of tumor cells are resistant to TRAIL [38,39]. Thus, studies are now underway to identify and characterize potential biomarkers of sensitivity to TRAIL. When several kinds of cancer cell lines are treated with a demethylation agent, zebularine, to induce DNA hypomethylation, expression of fucosylation regulatory genes is up-regulated, resulting in susceptibility to TRAIL-induced apoptosis [40].Combination therapy of TRAIL and demethylation agents could be a promising immunotherapy. In contrast, we have recently found that loss of fucosylation induced the inhibition of complex II formation which is an important component in apoptosis signaling [41] (Figure 2D). However, complex II, comprised of caspase-8 and cellular Fas-associated death domain [FADD]-like interleukin-1 beta-converting enzyme [FLICE] inhibitory protein, are not glycoproteins. Further studies should clarify how fucosylation affects complex formation. After answering this question, TRAIL therapy could be a more effective means of treatment.To use glyco-cancer biomarkers in clinical diagnosis, simple and repeatable methods are required. However, the mechanisms underlying the production of tumor specific oligosaccharide changes should first be investigated. While proteomic/glycomic approaches can produce a variety of candidates for cancer biomarkers, their target glycoproteins for characteristic oligosaccharides should be identified. Although lectins are powerful tools for simple oligosaccharide analyses, their cross-reactivity and low affinity for characteristic oligosaccharides present problems. An antibody for characteristic oligosaccharides or an antibody for characteristic glycol-peptides is a more ideal tool. We have investigated biological functions of cellular fucosylation and its clinical application. Although there are many fucose-specific lectins, their cross reactivity to the linkage of fucose-binding is a problem. Recently, we have found a novel lectin, which can bind to core-fucose, but not other types of fucose (manuscript submitted). This lectin can be used for immunohistochemistry. Since core-fucosylation regulates cell surface glycoproteins such as epidermal growth factor and transforming growth factor beta receptors directly [42,43], and loss of fucosylation brings dramatic changes in TRAIL signaling as described above, fucosylation is a promising target for cancer diagnosis and therapeutics. Deficiency of GDP-mannose-4,6-dehydratase (GMDS) leads to escape from natural killer (NK) cell-mediated tumor surveillance through modulation of tumor necrosis factor-related apoptosis-inducing ligand (TRAIL) signaling. (A) After transfection of wild-type GMDS gene into HCT116 cells, Western blot analysis of GMDS and AAL was performed. The binding to AAL was restored in transfected cells (WT-GMDS). Each number represents an independent clone; (B) Tumor growth of GMDS-rescued cells on the backs of athymic nude mice was significantly suppressed compared to mock cells. Rejected tumors at one month after transplantation were photographed; (C) The higher susceptibility of the GMDS-rescued cells to recombinant human TRAIL was confirmed by Western blotting of cleavedpoly (adenosine diphosphate-ribose) polymerase (PARP), which is a signal of apoptosis; (D) Complex II formation after TRAIL treatment was inhibited in HCT 116 cells which lacked cellular fucosylation. Immunoprecipitaion of FADD followed by Western blotting of caspase 8, c-Flip, and cleaved PARP were performed on HCT 116 clones treated with TRAIL. Detailed procedures are described in reference [29]. All data were derived from references [11,29].All authors have no conflict of interest in this study.This study was supported by a grant from a Grant-in-Aid for Scientific Research (A) No. 21249038 from the Japan Society for the Promotion of Science, and the Global COE Program of Osaka University funded by the Ministry of Education, Culture, Sports, Science, and Technology of Japan.

Med-MDPI/biomolecules/biomolecules-02-01-00046.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Complement is an essential part of innate immunity as it participates in host defense against infections, disposal of cellular debris and apoptotic cells, inflammatory processes and modulation of adaptive immune responses. Several soluble and membrane-bound regulators protect the host from the potentially deleterious effects of uncontrolled and misdirected complement activation. Factor H is a major soluble regulator of the alternative complement pathway, but it can also bind to host cells and tissues, protecting them from complement attack. Interactions of factor H with various endogenous ligands, such as pentraxins, extracellular matrix proteins and DNA are important in limiting local complement-mediated inflammation. Impaired regulatory as well as ligand and cell recognition functions of factor H, caused by mutations or autoantibodies, are associated with the kidney diseases: atypical hemolytic uremic syndrome and dense deposit disease and the eye disorder: age-related macular degeneration. In addition, factor H binds to receptors on host cells and is involved in adhesion, phagocytosis and modulation of cell activation. In this review we discuss current concepts on the physiological and pathophysiological roles of factor H in light of new data and recent developments in our understanding of the versatile roles of factor H as an inhibitor of complement activation and inflammation, as well as a mediator of cellular interactions. A detailed knowledge of the functions of factor H in health and disease is expected to unravel novel therapeutic intervention possibilities and to facilitate the development or improvement of therapies.Innate immunity is a first-line defense system, essential for the protection of the host against invading pathogens, acting immediately after infection and without previous antigen contact [1]. The innate immune system comprises several cellular and humoral components and utilizes germ-line encoded receptors for the recognition of microorganisms. A major humoral component of innate immunity is the complement system, which was established early in evolution and is present in invertebrates lacking an adaptive immune system. Originally identified as a serum component that ‘complements’ the antibody response towards pathogens, it is now known as a system of more than forty proteins. These proteins form a complex network of various recognition, effector, regulatory and receptor molecules that act in a finely tuned fashion, allowing complement to safely exert its functions [2]. It is well acknowledged that the functions of the complement system go far beyond the elimination of invading microbes. In addition to protecting the host from infections by destroying pathogens and promoting their elimination, complement has important roles in maintaining the integrity of the body by discriminating between healthy and injured tissue, in participating in the disposal of immune complexes, apoptotic/necrotic cells and cellular debris, as well as in inflammatory processes, angiogenesis and tissue regeneration. Moreover, complement is involved in the induction and regulation of both innate and adaptive cellular immune responses [3,4].Complement can be activated via three major pathways that merge at the central component C3 (Figure 1). The classical pathway is initiated by C1q binding to immune complexes, pentraxins or other targets such as apoptotic cells, the lectin pathway by binding of mannan-binding lectin (MBL) to repetitive carbohydrate residues, or by binding of ficolins to carbohydrate or acetylated groups on target surfaces, and the alternative pathway is spontaneously autoactivated by the hydrolysis of the internal thioester group of C3. Activation of the three pathways leads to the generation of the classical/lectin pathway C3 convertase (C4b2b) and the alternative pathway C3 convertase (C3bBb). These enzymes cleave C3, resulting in the two main fragments C3a, a potent inflammatory mediator, and C3b, an opsonin that deposits on the surface of target cells and particles and promotes phagocytosis. C3b can also form additional C3 convertases and in this way amplify complement activation. Furthermore, when C3b binds to an already existing C3 convertase, the new complex is termed a C5 convertase as it gains the ability to cleave C5 and thus to activate the terminal complement pathway. This potentially leads to inflammation (via C5a generation) and target cell lysis through the formation of membrane pores by the C5b-9 membrane attack complex. Whereas these activation processes are strongly favored on microbial surfaces, they are potentially destructive to host cells. To prevent host tissue damage, the activation of the complement system is strictly regulated by membrane-bound and plasma regulatory molecules. Since the alternative pathway is constitutively activated at a background level and is also activated secondary to classical/lectin pathway activation (through the ‘amplification loop’), its regulation is particularly important. Factor H is a major plasma regulator acting in the alternative complement pathway at the level of the central C3b component, facilitating C3b inactivation and the dissociation of the C3/C5 convertases and thus blocking further activation of the complement cascade (Figure 1).Complement activation pathways and the regulatory role of factor H.Complement activation is initiated by the recognition molecules of the classical (C1q) and lectin (MBL, ficolins) pathways or by the hydrolysis of the C3 thioester bond (alternative pathway). All three activation cascades lead to the assembly of C3 convertase enzymes that cleave the central C3 molecule into C3a and C3b. C3b deposits to nearby surfaces and, if not inactivated, it forms additional C3 convertases (amplification loop). The binding of C3b to an existing C3 convertase results in a C5 convertase that cleaves C5 into C5a, a potent anaphylatoxin, and C5b. C5b binds to surfaces and by binding C6, C7, C8 and several C9 molecules a terminal membrane attack complex is formed that allows target cell lysis. The system is controlled by several fluid phase and membrane-bound regulators that act at various steps of the activation cascade. Factor H is the major fluid-phase regulator of the alternative pathway, as it prevents the formation of the C3 and C5 convertases, facilitates the disassembly of already formed convertases and acts as a cofactor for the inactivation (enzymatic cleavage) of C3b.Disturbances in this complex system caused by complement gene mutations, autoantibodies or exogenous triggers may tip the balance between complement activation and inhibition, resulting in an attack on self tissue [4,5]. Complement deficiencies and malfunctions in the complement system are associated with various infectious, inflammatory and (auto)immune diseases. Factor H gene mutations and polymorphisms, as well as anti-factor H autoantibodies, are associated with several diseases that are characterized by complement dysregulation, e.g., the eye disorder age-related macular degeneration (AMD) and the rare kidney diseases atypical hemolytic uremic syndrome (aHUS) and dense deposit disease (DDD) [6,7,8]. Significant progress has been made in recent years in clarifying the roles of factor H and complement in these pathological conditions. Novel ligands and functions of factor H have been identified. Beside its role as a complement regulator, factor H has been shown to mediate cellular interactions by binding to receptors on various cells. A detailed description of the physiological and pathophysiological roles of factor H and its intricate interactions with other plasma and cell membrane molecules are necessary for our understanding of the underlying pathomechanisms of inflammatory, autoimmune and infectious diseases where factor H is implicated. This could lead to improved diagnostics and to the development of more effective treatments for affected patients. This review summarizes the current concepts of the roles of factor H in health and disease and discusses open questions for future research.Factor H is the main soluble regulator of the alternative complement pathway [9,10]. The factor H gene (CFH) is located on chromosome 1q32 in the regulators of complement activation (RCA) gene cluster, adjacent to the five genes that code for the factor H-related proteins (CFHRs). Factor H is constitutively expressed in the liver and is distributed systemically in body fluids. Reported factor H plasma concentrations vary, depending on the study population, age, genetic and environmental factors, as well as on the method used for quantification [11,12,13,14]. Of note, previous studies overestimated factor H concentration: a result of measuring both factor H and factor H-related proteins. While early reports suggested that the plasma factor H concentration ranges 265–684 µg/mL [11] and 116–562 µg/mL [12], current studies using monoclonal antibodies have established mean factor H concentrations of 233 µg/mL (in young adults), 269 µg/mL (in elderly individuals) [13], and 263 µg/mL [14], in different control populations. Thus, normal factor H concentrations in human plasma correspond to approximately 1–2 µM. In addition, factor H is produced extrahepatically by different cell types such as monocytes [15], fibroblasts [16], endothelial cells [17], keratinocytes [18], platelets [19] and retinal pigment epithelial cells [20]. Locally released factor H in tissues may help to limit complement activation and maintain an anti-inflammatory environment.Factor H is a single-chain, 150-kDa plasma glycoprotein composed of 20 domains. These are termed short consensus repeats (SCRs) or complement control protein modules (CCPs). Each of these autonomously folding globular domains is composed of approximately 60 amino acids and is stabilized by two internal disulfide-bonds [21]. Factor H regulates complement activation by (i) inhibiting the assembly of the alternative pathway C3 and C5 convertase enzymes via competition with factor B for C3b binding; (ii) facilitating the disassembly of the convertases by displacing bound factor Bb (‘decay accelerating activity’); and (iii) acting as a cofactor for the serine protease factor I in the cleavage and inactivation of C3b (‘cofactor activity’) [22,23,24]. These regulatory activities are mediated by the four N-terminal domains SCRs 1–4 [25,26], while the C-terminal domains SCRs 19-20 are responsible for target recognition (Figure 2) [27,28]. One of the important targets for factor H binding in the vicinity of C3b on host cells are polyanionic surface molecules, such as glycosaminoglycans and sialic acid, which increase the affinity of factor H for C3b [29,30]. Thus, in addition to its regulatory activities in the fluid phase, factor H is also able to control complement activation on self-surfaces (Figure 2) [31,32,33]. In contrast, host-like polyanionic molecules are normally not present on the surface of pathogens, rendering them susceptible to complement attack. The schematic structure of factor H. (a) Factor H is composed of 20 short consensus repeat (SCR) domains. Two major functional regions are located at the N- and C-termini of the molecule. SCRs 1-4 mediate the complement regulatory activities of factor H; (b) The SCRs 19-20 allow the attachment of factor H to host cells so that it can also inhibit complement activation directly at the cell surface.In addition to C3b and polyanionic molecules (such as surface glycosaminoglycans), factor H interacts with further endogenous ligands (Table 1), including the pentraxins C-reactive protein (CRP) and pentraxin 3 (PTX3), the extracellular matrix (ECM) proteins fibromodulin, osteoadherin and chondroadherin, prion protein, adrenomedullin, DNA, annexin-II and histones. These interactions allow factor H to inhibit complement on certain host surfaces (such as the glomerular basement membrane, the extracellular matrix, and late apoptotic cells), that are otherwise not properly protected due to a reduced expression or the lack of membrane-anchored complement regulators (i.e., membrane cofactor protein, decay accelerating factor and CD59). Several of these ligands and structures activate complement via interactions with C1q, MBL or ficolins. The simultaneous binding of both complement activating molecules (i.e., C1q, MBL) and complement inhibiting molecules (i.e., factor H, C4b-binding protein) on such ligands and cells may facilitate their opsonization and safe removal, but at the same time prevents an exaggerated complement activation which could lead to inflammation, cell lysis and subsequent tissue damage [34]. Factor H also binds to nonhost ligands, such as certain surface proteins of microbes, which hijack host factor H in order to protect themselves from complement attack (see section 6.1.), (reviewed e.g., in [35]). Furthermore, factor H can bind to receptors on host cells and mediate functions unrelated to its regulatory activity in the complement system (see section 7).Factor H ligands, binding sites and potential relevance of the interactions.The central complement protein C3b is the main host ligand of factor H. The C3b-factor H interaction is of particular importance for the pivotal functions of factor H, namely complement regulation and host surface recognition. In factor H, four binding sites were reported for C3b and its fragments, each with a different binding preference, affinity and functional relevance [36,37,38]. These binding sites are located in the SCR domains 1-4, 6-8, 12-14 and 19-20. Solid evidence supports the two major C3b binding sites in SCRs 1-4 and 19-20, whereas evidence for additional binding sites remains inconclusive. The N-terminal SCRs 1-4 mediate factor H binding to intact C3b, SCRs 12-14 to the C3c part of C3b (i.e., binds both C3b and the C3c fragment), and SCRs 19-20 to the C3d part of C3b (i.e., binds both C3b and the C3d fragment) [37]. Surface plasmon resonance analyses indicate that the main binding sites are in SCRs 1-4 and SCRs 19-20, with the latter having the higher affinity [38], whereas additional domains may contribute to C3b binding. The crystal structure of the complex of C3b and factor H SCRs 1-4 showed that all four N-terminal domains of factor H are involved in this interaction [54], and that these domains are necessary and sufficient for both cofactor and decay accelerating activities [25,54].Self-nonself discrimination by factor H is mediated by SCRs 19-20 that bind to surface-bound C3b/C3d and host surface glycosaminoglycans or sialic acids [28,38]. Recent structural data provided insights into this host recognition mechanism. SCR19 contains a main C3d (and C3b) binding site, whereas glycosaminoglycan binding is mediated by SCR20 [55,56]. Thus, the deposited C3b and the surface glycosaminoglycans (the latter lacking on microbes) together allow factor H SCRs 19-20 the recognition of host cells. Furthermore, the data of Kajander et al. raises the possibility that in addition to host polyanionic molecules, factor H is recruited by previously deposited and degraded C3b (i.e., C3d) via the SCR20 domain [55].In addition to membrane-bound regulators, host cells require soluble complement inhibitors, particularly factor H, which provide effective protection from unwanted complement-mediated damage, especially under conditions with strong complement activation. This is exemplified by the attachment of factor H to endothelial cells via cell surface glycosaminoglycans and C3b, as discussed above, which is impaired in aHUS and is associated with endothelial damage and acute renal failure [32,33,55,56]. Also, factor H can bind to cell surface polyanionic molecules in the absence of C3b, although this binding is weak and not readily detectable in physiological buffers [32]. Notably, this interaction is distinct from the binding to specific cellular receptors (see section 7).Host cells, so-called nonactivators of complement, possess cell surface polyanionic molecules that allow for factor H binding [29]. In contrast, complement activators such as microbes or rabbit erythrocytes that lack sialic acids and host-like glycosaminoglycans do not allow significant factor H binding and thus complement activation can proceed unchecked. Heparin is generally used in studies as a model of host glycosaminoglycans, and the major heparin binding sites were located in SCRs 7 and 19-20, and a possible third site in SCR13 for which evidence remains inconclusive [38,57,58,59]. Polymorphisms or mutations in SCRs 7 and 19-20 may affect interactions of factor H with host cells and basement membranes, and are implicated in the diseases AMD and aHUS (as discussed in section 5).To maintain tissue homeostasis, old and damaged cells must be removed and replaced by new ones. This is facilitated by apoptosis, a programmed mechanism of cell death, which involves changes such as nuclear and cellular fragmentation, chromatin condensation and cell shrinkage. Changes in the cell membranes facilitate an efficient recognition and safe clearance of apoptotic cells by phagocytes. Complement proteins (e.g., C1q and MBL) and pentraxins (CRP, PTX3) can bind to apoptotic cells and thereby enhance their uptake by phagocytes via specific receptors (i.e., complement and Fcγ receptors) [60]. Moreover, pentraxins can enhance the binding of C1q, which may further increase the deposition of complement-derived opsonins. However, this could potentially lead to the activation of the terminal pathway. Gershov et al. showed that complement activation does not proceed to the terminal pathway on apoptotic cells, which is partly due to the binding of factor H [44]. The expression of membrane-bound complement regulators is down-regulated on apoptotic cells, which would increase the susceptibility of these cells to complement-mediated lysis. The loss of membrane-bound regulators on apoptotic cells is in part compensated by the acquisition of the soluble complement regulators factor H and C4b-binding protein, protecting against complement attack which would otherwise lead to the release of potential autoantigens from the cells [45]. The factor H binding site for apoptotic/necrotic cells is located within SCRs 6-20, which is outside the complement regulatory region. Thus surface-bound factor H is able to regulate complement activation. This binding is in part mediated by annexin-II, DNA and histones, which become exposed on the surface of apoptotic cells [46]. In addition, factor H may be recruited by monomeric CRP, and this interaction further facilitates the removal of apoptotic cells in a non-inflammatory way [41]. Notably, native pentameric CRP does not enhance the binding of factor H to apoptotic or necrotic cells [41,45].Pentraxins are recognition molecules of the innate immune system [61]. The classical short pentraxins CRP and serum amyloid P component circulate as pentamers in human plasma. The long pentraxins, including PTX3, PTX4 and neuronal pentraxins, display a more complex structure. The functions of pentraxins in innate immune defense and beyond are reviewed elsewhere [61,62]; here, we focus on their interaction with factor H.C-reactive protein. Human CRP is an acute-phase protein, whose synthesis by hepatocytes is up-regulated in response to inflammatory stimuli. Its plasma concentration can increase dramatically from below 1 µg/mL to more than 500 µg/mL following the initiation of an acute phase reaction [63]. The main effector functions of CRP are the activation of the complement system and the stimulation of phagocytosis [64]. CRP activates the classical and lectin complement pathways by binding C1q and ficolins and thus can lead to an enhanced opsonization of target surfaces and cells [65,66]. CRP can be recognized directly by Fcγ receptors on leukocytes, thereby activating phagocytosis [67]. Thus, CRP and complement can collaborate and have synergistic functions, for example in the removal of apoptotic cells and particles [44,60]. However, there are contradictory reports regarding CRP interactions and functions that are in part explained by different conformations of the used CRP [63]. CRP circulates in plasma as a pentamer, whose conformation is stabilized by calcium ions. In vitro CRP immobilization on surfaces, such as microtiter plates used for ELISA or chips used for surface plasmon resonance studies, or the use of inappropriate buffer conditions (e.g., low calcium content), may cause denaturation or aggregation of the protein, leading to artefacts or results with no physiological relevance [42,68]. On the other hand, it cannot be excluded that under certain conditions, for instance the lower pH of inflammatory sites, or by binding of the native pentamer to certain ligands and surfaces, conformational changes may occur leading to the exposure of novel epitopes or to the dissociation into the monomeric form, termed mCRP [69,70,71]. These considerations are relevant for assessing the in vivo significance of the interaction of factor H with CRP.A direct binding of factor H to CRP was described [40,72], suggesting regulation of CRP-mediated complement activation on self surfaces. As we and others have shown, factor H mainly interacts with the monomeric or denatured form of CRP [41,68,73,74], although an interaction of factor H with native pentrameric CRP at acute phase concentrations was also demonstrated by analytical ultracentrifugation [42]. Thus, it is unlikely that under normal conditions CRP would interact with factor H to a significant extent. However, under infection/inflammatory conditions (e.g., during the acute phase response or at sites of tissue damage and local inflammation), CRP can bind factor H and locally focus its complement inhibitory activity. The interaction of factor H with mCRP leads to factor H recruitment, which limits complement activation but increases phagocytosis of apoptotic cells and reduces the release of inflammatory cytokines by macrophages [41]. Such regulation is reduced by the common factor H variant 402H, which shows a reduced binding to mCRP on self surfaces [71]. Pentraxin 3. Aside from the short pentraxin CRP, factor H was also shown to bind to the long pentraxin PTX3 [43]. PTX3 has a pentraxin domain, homologous to that of the short pentraxins CRP and serum amyloid P, and has an additional unique N-terminal domain. In contrast to the short pentraxins that are mainly produced in the liver and thus act systemically in the body, PTX3 is produced locally by various cell types, such as vascular endothelial cells, fibroblasts, monocytes, macrophages, myeloid dendritic cells and neutrophil granulocytes [61,62]. PTX3 expression is increased upon inflammatory stimuli and exposure to pathogens, leading to significantly elevated PTX3 plasma levels (200-800 ng/mL, compared with approximately 2 ng/mL in normal plasma) [62]. PTX3 can activate the classical and lectin complement pathways by binding to C1q, MBL, M-ficolin and L-ficolin [75,76,77,78]. In addition to these complement-activating molecules, PTX3 binds the complement regulators factor H and C4b-binding protein [43,79]. The binding of factor H to PTX3 requires the presence of calcium and is mediated by at least two binding sites in factor H. The primary binding site located within SCRs 19-20 of factor H interacts with the N-terminal domain of PTX3, whereas a secondary binding site on SCR 7 binds to the C-terminal pentraxin domain [43]. PTX3 recruits both factor H and C4b-binding protein to the surface of apoptotic cells, which prevents excessive complement activation and cell lysis [43,79]. We also found that ECM-bound PTX3 increases the recruitment of factor H and C4b-binding protein, resulting in enhanced local complement regulation [79,80]. According to preliminary data, certain factor H mutations in SCR20 and aHUS-associated autoantibodies impair the binding of factor H to PTX3 and this may result in a reduced local complement inhibition [81].In summary, these in vitro studies suggest that factor H binding to pentraxins is important to limit complement activation and inflammation locally. However, further studies are warranted to determine the relevance of factor H-pentraxin interactions in vivo.Extracellular matrices are important components of the extracellular space, providing a scaffold for residing cells, mechanically supporting cellular movements and binding various biomolecules derived from body fluids or nearby cells. ECM components may be exposed to complement during pathological processes, such as injury to the endothelium or to the cartilage in the joints. Since the ECM can activate complement, e.g., via the binding of C1q to certain ECM components, a proper regulation is important to prevent inflammation [34]. Because ECMs lack the membrane-bound complement regulators that normally protect host cells, attachment of fluid phase regulators to ECM is considered important. It was shown that the ECM proteins fibromodulin, chondroadherin and osteoadherin can bind both C1q and the regulators factor H and C4b-binding protein, which maintains a balance between complement activation and inhibition [34,47,49]. Exaggerated complement activation in turn may lead to inflammatory disease.We have recently analyzed complement activation and the roles of factor H and C4b-binding protein on endothelial cell-derived ECM in vitro. ECM-bound factor H and C4b-binding protein acted as cofactors for the inactivation of C3b and C4b, respectively. Furthermore, their binding and thus cofactor activity were enhanced by PTX3 [79,80]. The possibility that factor H may play a role in the regulation of local inflammation at the ECM in vivo is supported by reports showing factor H binding to the Bruch’s membrane in the retina [82] and an association of factor H with aortic ECM [83].The factor H protein family includes factor H, factor H-like protein 1 (CFHL1) and five factor H-related proteins (CFHRs). All these proteins are composed of different numbers of SCR domains, each exhibiting varying degrees of sequence similarity and displaying diverse, but also overlapping, biological functions (reviewed in [8]).CFHL1 is a 43-kDa protein that derives from an alternative splice product of the CFH gene. It shares the seven N-terminal SCRs with factor H and has four additional amino acids. Due to the broad sequence overlapping, CFHL1 possesses the same complement regulatory activities mediated by the N-terminus of factor H. CFHL1 may also play a role in age-related macular degeneration as the SCR7 harbors the Y402H polymorphism. Consequently, the CFHL1 402H variant has impaired ligand-binding capacity, similar to that exhibited by the factor H 402H variant [71,84]. However, an expression pattern differing from factor H and a distinct role in mediating cell adhesion have been reported for CFHL1 [85,86].The five factor H-related proteins are derived from separate genes (CFHR1 to CFHR5) which are located adjacent to the CFH gene. CFHR proteins lack the complement regulatory activities of factor H, but some of them have weak cofactor or decay accelerating activity, or modulate the complement regulatory activity of factor H [8]. These proteins have both redundant and nonredundant ligands and functions with factor H. CFHR1, CFHR3, CFHR4 and CFHR5 were shown to bind to C3b, CFHR1, CFHR3 and CFHR5 to heparin, and CFHR4 and CFHR5 to CRP. In spite of similar cell and ligand binding properties, however, CFHR1 was reported to regulate the terminal complement pathway [87]. In addition, we showed that both factor H and CFHR1 enhance neutrophil adhesion and activation during host cell-pathogen contact [88]. Similar to factor H, CFHR4 binds to late apoptotic and necrotic cells, but in contrast to factor H, CFHR4 binds to the native pentameric form of CRP [68]. Even though a comprehensive knowledge of CFHR functions is lacking, the available data indicate that CFHR proteins may be relevant in regulating local inflammatory processes and could modulate the functions of factor H, e.g. through competition [8].Complement regulatory defects due to factor H mutations or anti-factor H autoantibodies have been described in certain pathological conditions, and CFH polymorphisms have also been associated with disease. Here, we briefly review the three diseases where the role of factor H has been best studied. AMD is a leading cause of visual impairment in elderly, western populations. In recent years, complement gene mutations and polymorphisms have been found to be associated with AMD, pointing to a role of the complement system in the pathogenesis of the disease [89]. Although the underlying pathomechanism is not yet fully known, a role of complement-mediated inflammation in the eye is postulated. Correspondingly, several therapeutic compounds targeting the complement system are currently evaluated in clinical trials [89].The common factor H polymorphism 402H has been identified as a major genetic risk factor for developing AMD [90,91,92,93]. In addition, a protective CFH haplotype associated with the deletion of the CFHR1 and CFHR3 genes in AMD has been described [94]. Functional analyses of the factor H 402Y and 402H variants revealed a reduced binding of the AMD-associated 402H variant to mCRP [48,71,84,95,96,97,98]. Since residue 402 in SCR7 is involved in the glycosaminoglycan-binding site of factor H, there are also subtle differences between the variants in their interaction with heparin and glycosaminoglycan-analogs [97,99,100]. In contrast, the 402H variant has a higher affinity for DNA and necrotic cells compared to the 402Y variant [48]. No difference in binding to retinal pigment epithelial cells was found [98], but the disease-associated variant binds less efficiently to both the extracellular matrix protein fibromodulin [48] and the Bruch’s membrane in the retina [82]. In addition, malondialdehyde, a lipid peroxidation product has been described as a novel ligand of factor H on apoptotic/necrotic cells, and shown to bind the 402H variant less strongly, thus adversely affecting the anti-inflammatory role of factor H [50]. Very recently, the rare factor H variant R1210C, previously described in aHUS patients, has been linked to AMD [101]. This variant was shown to affect factor H interactions with C3b and cell surfaces [102]. Altogether, these data suggest a factor H-associated defect in the proper, non-inflammatory handling of cellular waste and in the control of complement activation and inflammation locally at the surfaces of the Bruch’s membrane and damaged retinal pigment epithelial cells. It is still unknown how these factor H defective functions cause or contribute to the late-onset disease AMD in affected individuals, and which other factors (genetic, environmental, life-style) influence the role of factor H. Recent data indicate that common polymorphisms in factor H, C3 and factor B act collaboratively in determining complement activity and the risk to disease [103].A further functional impairment of the 402H variant is a reduced binding to streptococcal M6 protein [96,98]. Functional studies showed a decreased binding of the 402H factor H variant to Streptococcus pyogenes, resulting in increased C3b deposition and phagocytosis [104]. A genetic association study suggested that the 402H variant is protective against streptococcal tonsillitis [105]. These results are highly interesting and indicate that the 402H variant has been established in the human population due to selection pressure by pathogenic microbes.A study investigating the prevalence of anti-factor H autoantibodies in AMD showed that it is decreased in the patient group compared with age-matched controls. Analysis of the antibody binding sites demonstrated recognition of several parts of factor H, including both the N- and C-terminal domains of the molecule, a pattern different from that seen in atypical hemolytic uremic syndrome (see below) [106]. Therefore, it is unlikely that autoantibodies to factor H have pathological significance in this disease.The rare kidney disease aHUS is characterized by hemolytic anemia, low platelet count and impaired renal function [107]. Its pathomechanism is related to dysregulation of the alternative complement pathway, caused by polymorphisms, mutations and deletions in complement genes, or due to factor H autoantibodies (reviewed in [108]).Factor H mutations affect approximately 30% of aHUS patients. More than 100 factor H mutations have been described in aHUS patients and can be searched in an online database (http://www.fh-hus.org) [109]. In most cases, these are heterozygous mutations affecting various domains of factor H. However, most of the mutations affect the C-terminal SCRs 19-20. Functional analyses of several of these mutants showed an altered interaction with C3b, heparin and endothelial cells [102,110,111]. Furthermore, gene conversion and gene deletions leading to hybrid factor H proteins with functionally affected C-terminal domains have been reported [112,113,114]. These data show a disturbance in the physiologic interaction of factor H with host endothelial cells. Recent structural studies have provided new insights into how these mutations impair the function of factor H in host-nonhost discrimination [55,56]. Certain mutations in SCR20 were also found to reduce the binding of factor H to CRP [41] and, according to our preliminary data, to PTX3 [81]. For many mutations, however, there is no functional effect known to date. Anti-factor H IgG autoantibodies are detected in approximately 10% of aHUS patients [115,116]. This form of aHUS affects mainly children and young patients. As a result of its autoimmune nature, it requires a therapy that addresses the elimination or suppression of the autoantibody producing cells. A further characteristic of this patient group is that 90% of the affected individuals lack the CFHR1 gene, indicating that this genetic defect predisposes to the development of factor H autoantibodies [116,117,118,119]. These autoantibodies can also occur together with mutations in the CFH, CFI, C3 or MCP genes [119]. We and others have determined the antibody binding sites in several patients using recombinant factor H fragments and found that the autoantibodies mainly bind to SCRs 19-20 of factor H, although in some cases reactivity with other domains, such as SCRs 8-11, was also observed [116,119,120]. In three patients anti-factor H IgA autoantibodies were found that similarly recognized SCRs 19-20 [121]. Our functional studies indicated that the autoantibodies interfere with the recognition functions of factor H, namely, impairing its interaction with surface-bound C3b and inhibiting the factor H complement regulatory activity on host surfaces [120,121,122]. This reduced protection from complement-mediated damage is likely to be involved in the endothelial injury associated with aHUS. Due to the similar C-terminal SCRs of factor H and CFHR1, most of the studied autoantibodies recognize both host complement regulators. CFHR1 in fact can hijack autoantibodies and rescue host cells when added to anti-factor H autoantibody-positive plasma [121].Altogether, these genetic and acquired abnormalities affecting factor H allow a normal fluid-phase regulation, but result in an impaired cell binding and cell surface protection from complement attack, which apparently contribute to the endothelial damage and microvascular thrombus formation in aHUS. However, further studies are needed to understand the role and relevance of mutations affecting other domains of factor H in aHUS. A recent study showed that some of the mutations do not lead to any known functional effect on factor H, thus care should be taken when interpreting genetic data and advising patients [123].Dense deposit disease (DDD), also termed membranoproliferative glomerulonephritis type II, is a rare renal disease that progresses to end-stage renal failure in about 50% of patients. It is a disease associated with uncontrolled alternative pathway activation in plasma that generates C3 activation fragments depositing in the glomeruli [124]. In the majority of patients, autoantibodies against the C3 convertase (C3 nephritic factor) can be detected that stabilize the convertase and thus cause enhanced complement activation. In some cases, factor H mutations have been identified in these patients [6,125]. These mutations may lead to factor H deficiency and thus an insufficient plasma complement control [126]. Mutations in cysteine residues that are important for forming the disulfide bonds within the single SCR domains can result in a defective protein folding and a factor H secretion defect [127]. In one report, a C431Y exchange was described that caused aggregation of factor H and likely resulted in a reduced protein half-life [128]. Moreover, a mutation in SCR4 was described, where the mutant factor H protein was inefficient in its cofactor activity, while cell-binding functions remainedunaffected [129]. All these cases led to a defective C3 activation control in plasma, either due to a quantitative factor H deficiency or dysfunctional factor H.Anti-factor H autoantibodies have also been described in DDD. So far, only one case has been published where the autoantibody was characterized in detail. The isolated factor H ‘mini-autoantibody’ consisted of lambda light-chain dimers that bound to SCR3 of factor H, i.e., within the complement regulatory region of the molecule. Functional assays demonstrated that the autoantibody inhibited the factor H-C3b interaction and caused an increased C3 turnover due to a blockade of the complement inhibitory activity of factor H [130,131]. Due to the lack of systematic screening for such autoantibodies in DDD patients, at present the prevalence and the characteristics of DDD-associated anti-factor H autoantibodies are not known.Besides its physiologic interactions with host cells and ligands, the binding of factor H to several pathogenic and non-pathogenic microbes was demonstrated as a process believed to help microorganisms in complement evasion. In other words, these microbes hijack factor H to camouflage themselves as host-like cells and thus are protected from complement attack. Furthermore, certain tumor cells can also exploit factor H to increase their protection from the complement system.Since complement plays an important role in protection against infections, it is not surprising that numerous viruses, bacteria, fungi and parasites have acquired the ability to sequester host complement regulators such as factor H. A detailed discussion of these mechanisms is beyond the scope of this review and there are excellent overviews of this topic throughout the literature (e.g., [35]). As a common theme, it appears that many of the microbial factor H binding proteins interact with the positively charged surfaces on SCRs 7 and 19-20, i.e., the factor H domains relevant for host cell recognition, although for example the factor H binding protein (fHbp) of Neisseria meningitidis binds to SCR6 [132]. Interestingly, recent studies indicate that the factor H 402H polymorphism may relate to a better resistance from certain bacterial infections. Haapasalo et al. showed that the AMD-associated factor H 402H variant has a lower binding affinity to various streptococci compared to the 402Y variant, resulting in a more efficient opsonization and phagocytosis [104,105]. These data point to a pathogen-driven establishment of this common polymorphism in the human population, with evolutionary advantage against bacterial infection at the expense of late-age adverse effect in developing AMD. This example raises the possibility that other factor H polymorphisms have similarly spread in the human population because of the evolutionary race between humans and their pathogens.While sequestration of host factor H is described for many pathogens, a biologically relevant role under physiological settings has rarely been demonstrated [105]. Both pathogenic and non-pathogenic microbes can bind factor H, thus it does not necessarily play a decisive role in microbial immune escape. However, microbes may use factor H for other purposes than complement inhibition, such as mediating entry into host cells (discussed below, see section 7).Tumor cells have various means to escape from the control of the immune system. Protection from the attack of the complement system is part of this immune evasion repertoire [133]. In addition to modulating the expression of membrane-anchored complement regulatory proteins and degrading complement components, tumor cells may also use factor H for efficient protection [134]. Several tumor cells were reported to express and release increased amounts of factor H, thus reducing complement activity in their microenvironment [135,136,137,138]. The increased factor H expression may even be used as a diagnostic marker for certain cancers [139,140]. Furthermore, in certain cancers the increased expression of the factor H binding proteins: bone sialoprotein, osteopontin and dentin matrix protein-1 may confer enhanced protection from complement [141]. In addition, anti-factor H autoantibodies have been described in non–small cell lung cancer [142]; it is not yet known, however, whether this is due to an increased production of factor H and if and how these autoantibodies differ from those described in aHUS and DDD. These data clearly indicate that tumor cells can exploit factor H to their advantage. Therefore, a targeted inhibition of factor H overexpression by these cells may enhance their sensitivity to complement-mediated lysis.When bound to the surface of host cells via C3b/C3d and polyanionic molecules, the major task of factor H is to control complement activation. However, factor H can also bind to cell surface receptors and modulate cell activation and cellular functions. To date, studies mainly focused on factor H interactions with soluble (plasma) ligands and their role in complement regulation, but there is a lack of detailed information on factor H interactions with host cell receptors. However, there is support for such non-canonical roles (i.e., beyond its complement regulatory function) of factor H.Early studies have shown that factor H binds to human B lymphocytes and stimulates a calcium-dependent factor I release from these cells [143], and that factor H also stimulates murine B cells and triggers blastogenesis [144]. Tsokos et al. demonstrated that factor H blocks the differentiation, but not the proliferation, of B cells [145]. Attempts to identify the B cell factor H receptor resulted in the description of a putative receptor consisting of three subunits (each of ca. 50 kDa) [146] and in a 140-kDa single polypeptide chain protein [147]. However, the nature of these receptors at the molecular level remains unresolved.Factor H binds to human neutrophil granulocytes via complement receptor type 3 (CR3; CD11b/CD18, αMβ2 integrin, Mac-1) [148,149]. Factor H was shown to support neutrophil adherence and to enhance the release of reactive oxygen species in primed neutrophils [149]. Recently, we characterized the interaction of factor H with neutrophils in the context of host-pathogen interaction [88]. Factor H, when bound on the human-pathogenic yeast Candida albicans, served as a bridging molecule to enhance the adherence and antimicrobial activity of neutrophils. We confirmed CR3 as the major neutrophil factor H receptor, but the data also indicated that CR4 likely binds factor H. Although the latter is expected to play only a minor role in the case of neutrophils, because of low CR4 expression, it might be relevant on other cells, such as macrophages and dendritic cells which express significant amounts of CR4. In addition, we have identified the factor H SCR7 and SCRs 19-20 domains as major binding sites for CR3. The two factor H family proteins CFHL1 and CFHR1 also bound to CR3 and supported neutrophil adhesion [88]. In parallel, other groups have reported that binding of factor H facilitates the entry of pathogens into host cells. The factor H-CR3 interaction could enhance Streptococcus pneumoniae adherence and uptake by epithelial cells and neutrophils [150]. Similarly, factor H facilitated the adherence of Neisseria gonorrhoeae to human CR3-transfected cells [151]. These data indicate a role for factor H in cellular adhesion by interacting with CR3, even beyond pathogen-host cell interactions. For example, factor H was shown to bind to heparan sulfate proteoglycans in amyloid-β plaques and to colocalize with CR3, suggesting that factor H may facilitate the recognition of amyloid-β plaques by microglia in the Alzheimer’s disease brain [152].Factor H was shown to stimulate the respiratory burst [153] and to induce the secretion of IL-1β in monocytes [154]. Thrombin-cleaved factor H was described as a monocyte chemotactic factor in a delayed-type hypersensitivity model [155]. A monocyte chemotactic effect of factor H was also demonstrated by Nabil et al. [156]. Furthermore, factor H was shown to induce the release of prostaglandin E and thromboxane from guinea pig peritoneal macrophages [157]. In these studies, however, no specific factor H receptor was identified. Since these cells express CR3, it is likely that at least some of the observed effects are mediated via factor H binding to CR3. A recent study and our unpublished data also show that CR3 is a factor H receptor on monocytes [158,159].Furthermore, the binding of factor H to L-selectin was reported, and immobilized factor H (but not soluble factor H) induced TNFα release from leukocytes [160].Finally, factor H binds to resting platelets and this binding is increased when platelets become activated [161]. Factor H binds to platelets both through thrombospondin and directly via the platelet integrin αIIbβ3 [162].Altogether these data attest to potential non-canonical roles of factor H in mediating cellular functions and interactions. These functions and interactions are likely to be important under local inflammatory conditions, such as the regulation of neutrophil adherence and migration, and in the activation of macrophages and dendritic cells. In addition, the interaction of factor H with B cells indicates a potential direct modulatory role in adaptive immunity. However, our present knowledge of factor H cellular functions is very limited. Further studies using highly purified and recombinant factor H are needed either to confirm and extend or disprove these observations. In order to better understand the roles of factor H in health and disease, a comprehensive characterization of factor H-host cell interactions and an assessment of their biological relevance are necessary. It is possible, for example, that known polymorphisms or mutations influence these cellular functions of factor H, providing new insights into the pathomechanisms of factor H-associated disease and revealing novel therapeutic intervention points.More than 45 years after its initial discovery, factor H still has many secrets to reveal. The last decade has brought with it a wealth of new information on factor H structure and function, and its pivotal role in host-nonhost discrimination is now well appreciated. We have started to understand how CFH variants and autoantibodies are involved in diseases such as AMD, aHUS and DDD. In addition, factor H may play an important role in other common diseases such as asthma [163]. A therapeutic use of factor H in such conditions needs to be evaluated in future studies, which could be facilitated by the recently described recombinant production of biologically active factor H protein [164,165]. In addition to its role as a main regulator of the complement alternative pathway, recent studies show a role for factor H in modulating classical pathway activation by competing with C1q for binding to the same ligands such as phospholipids or Escherichia coli [166]. Here again factor H appears as a downregulator of inflammatory responses. Importantly, factor H is not only a complement regulator, but also a direct modulator of cellular functions by binding to receptors. In this role factor H influences cellular adhesion, phagocytosis and antimicrobial activities [88,158]. These aspects of factor H functions warrant further attention as it is likely that they are relevant for innate resistance against infections, the handling of apoptotic cells and debris, modulation of adaptive immunity and cellular interactions with the ECM (potentially including tumor cells). Further efforts to identify and characterize factor H ligands, cellular activities and its roles in diseases will bring us closer to a better understanding of the versatile roles played by factor H in both health and disease (Figure 3), with the hope of applying this knowledge for the benefit of all.Overview of main factor H functions and their implication in pathological conditions. Factor H is a major soluble complement regulator that inhibits activation of the alternative complement pathway in body fluids. In addition, factor H is also able to control complement activation on self-surfaces and thereby protects host cells and tissues from complement attack. Factor H binds to host cells and basement membranes via interactions with glycosaminoglycans and deposited C3b. The binding of factor H to apoptotic cells and extracellular matrices is in part mediated by the pentraxins CRP and PTX3, and is of particular importance as these structures and cells are otherwise not well protected from complement. Furthermore, factor H interacts with host cells via specific receptors and thus modulates cellular functions, including adhesion and phagocytosis. Impaired functions of factor H are associated with several diseases such as DDD, aHUS and AMD. Mutations in the CFH gene (examples of affected domains are shown in red) or autoantibodies directed against factor H (“mini-autoantibody” in DDD and anti-factor H IgG in aHUS are shown in red) cause defective complement regulation and ligand recognition leading to improperly controlled inflammation and tissue damage. On the other hand, the regulatory functions of factor H are abused by several pathogens and tumor cells in order to protect themselves from complement attack and thus to evade the host immune response.The authors declare no conflict of interest.We thank Josephine Losse for her contribution to the functional analyses of factor H and factor H-related proteins in neutrophil activation, Stefanie Strobel and Harald Seeberger for their work on anti-factor H autoantibodies, and Seána Duggan for proofreading the manuscript. We are also grateful to all our collaborators. The work of the authors was supported in part by the German Research Foundation (Deutsche Forschungsgemeinschaft) JO 144/1-1. A.K. was supported in part by the Jena School for Microbial Communication, a graduate school of the University of Jena. Due to the large scientific literature on factor H, it was not possible to cite every relevant article in this review. We apologize to the authors whose important contributions were not cited due to space limitations and the focus of this review.

Med-MDPI/biomolecules/biomolecules-02-01-00076.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Oxysterols are oxidized 27-carbon cholesterol derivatives or by-products of cholesterol biosynthesis, with a spectrum of biologic activities. Several oxysterols have cytotoxic and pro-apoptotic activities, the ability to interfere with the lateral domain organization, and packing of membrane lipids. These properties may account for their suggested roles in the pathology of diseases such as atherosclerosis, age-onset macular degeneration and Alzheimer’s disease. Oxysterols also have the capacity to induce inflammatory responses and play roles in cell differentiation processes. The functions of oxysterols as intermediates in the synthesis of bile acids and steroid hormones, and as readily transportable forms of sterol, are well established. Furthermore, their actions as endogenous regulators of gene expression in lipid metabolism via liver X receptors and the Insig (insulin-induced gene) proteins have been investigated in detail. The cytoplasmic oxysterol-binding protein (OSBP) homologues form a group of oxysterol/cholesterol sensors that has recently attracted a lot of attention. However, their mode of action is, as yet, poorly understood. Retinoic acid receptor-related orphan receptors (ROR) α and γ, and Epstein-Barr virus induced gene 2 (EBI2) have been identified as novel oxysterol receptors, revealing new physiologic oxysterol effector mechanisms in development, metabolism, and immunity, and evoking enhanced interest in these compounds in the field of biomedicine.Oxysterols are 27-carbon oxidized derivatives of cholesterol or by-products of the cholesterol biosynthetic process with multiple biological activities. Several major oxysterols arise as intermediates in the pathways converting cholesterol to bile acids or steroid hormones, and their roles as readily transportable forms of sterol are well established [1,2,3]. The common oxygen-containing modifications of cholesterol in oxysterols are hydroxyl, keto, hydroperoxy, epoxy, and carboxyl moieties. In general, oxysterols have a drastically shorter biologic half-life than cholesterol—they can therefore be considered a way to route the cholesterol molecule for catabolism. The physiologically most important oxysterols are generated in cells by mitochondrial or endoplasmic reticulum cholesterol hydroxylases belonging to the cytochrome P450 family [4,5]. Of these species, the most abundant in human serum are 27-, 24(S)-, 7α-, and 4β-hydroxycholesterol (OHC). Of these, 24(S)-OHC originates from neurons in the central nervous system, the sterol homeostasis of which depends on the synthesis of this oxysterol catalyzed by the cholesterol hydroxylase CYP46A1 [6], 7α- and 27-OHC are synthesized by the liver by CYP7A1 and CYP27A1 as the first intermediates of classic and acidic bile acid synthetic pathways, respectively [7]; CYP27A1 however is functional also in non-hepatic cells [8]. 4β-OHC is generated by the hepatic drug metabolizing enzyme CYP3A4, which is markedly induced by certain anti-epileptic pharmaceuticals [9]. Oxysterols also arise in vivo or during food processing through non-enzymatic, free radical, lipid peroxide, or divalent cation-induced oxidative processes, often termed cholesterol autoxidation [10]. The most abundant oxysterols generated through autoxidation are modified at the 7-position of the cholesterol B-ring. These include 7-ketocholesterol (7-KC) and 7β-OHC with prominent cytotoxic and pro-apoptotic properties [11]. Structures of the most abundant oxysterols and the routes of their synthesis are depicted in Figure 1 [2], and the nomenclature of these oxysterols is specified in Table 1.Nomenclature of the oxysterols [3] discussed in this review.Structure and origin of selected common oxysterols. Most of the oxysterol species displayed are generated by enzymes that belong to the cytochrome P450 family (CYP). CH25H, cholesterol 25-hydroxylase, is a di-iron enzyme. The enzymatically derived species are indicated with green, products of cholesterol autoxidation with red, and a species derived from a shunt of the cholesterol biosynthetic process with blue print (Modified from [2] with kind permission from Springer Science+Business Media B.V.). Oxysterols are present in mammalian tissues at very low concentrations, as mixtures accompanied by a high excess of cholesterol. However, they are found enriched in pathologic structures such as macrophage foam cells, atherosclerotic lesions, cataracts, and gall stones. The cytotoxic and pro-apoptotic actions of oxysterols are suggested by play a role in the disease processes involved [12,13,14,15]. Moreover, oxysterols have been implicated in the pathology of degenerative diseases such age-onset macular degeneration and Alzheimer’s disease [16,17]. The above findings, as well as the potent regulatory activity that several oxysterols have on cellular cholesterol homeostatic machineries (reviewed by Gill et al. [18]), have prompted intensive research of oxysterols in the context of lipid homeostasis and atherosclerosis. However, novel physiologic activities of these compounds have emerged. Oxysterols are believed to act as endogenous regulators of gene expression in lipid metabolism and as signaling molecules with key roles in developmental, differentiation, and inflammation processes. Recently, novel oxysterol receptors responsible for these activities have been identified, opening new perspectives to oxysterol function and evoking novel interest in these compounds in the field of biomedical research. The present review focuses on the newly identified receptors and cellular effector pathways of oxysterols.The pioneering work by D. Mangelsdorf’s group revealed that the orphan nuclear receptors liver X receptor (LXR) α (NR1H3) and β (NR1H2) are activated by oxysterol ligands [19]. The physiologically most important endogenous LXR ligands are most likely 24(S),25-EPOX, 24(S)-OHC, 22(R)-OHC, 20(S)-OHC, and 27-OHC [19,20,21]. LXRα is expressed at highest amounts in liver, but also in adipose tissue, intestine, kidney, and macrophages, while LXRβ is ubiquitously expressed. The LXRs form heterodimers with retinoid X receptor (RXR) and play central roles in sterol absorption in the intestine, the reverse cholesterol transport process, bile acid synthesis, biliary neutral sterol secretion, hepatic lipogenesis, and synthesis of nascent high-density lipoproteins [22]—the LXRs have also been established as suppressors of inflammatory gene expression in macrophage [23,24], and recent studies have established an increasing number of LXR functions in immune regulation (see 3.4.). Studies employing atherosclerosis-prone mouse models have revealed anti-atherogenic function of the LXRs [25,26]. However, LXR stimulation with synthetic agonists such as T0901317 also leads, via a lipogenic activity, to hepatic steatosis and increased serum triglyceride levels, which represents an atherosclerosis risk factor [27]. This caveat can be avoided via the use of intestine-specific LXR activation leading to reduced cholesterol absorption, as demonstrated by genetic and pharmacologic approaches [28,29]. While there is some controversy as to whether oxysterols truly act as endogenous ligands of the LXR [1]: The work by Chen et al. [30] provided evidence that this indeed is the case: Overexpression of an oxysterol catabolic enzyme, cholesterol sulfotransferase, was shown to inactivate LXR signaling in several cultured mammalian cell lines but did not alter receptor response to the nonsterol LXR agonist T0901317. Moreover, triple-knockout mice deficient in the biosynthesis of three oxysterol ligands of LXRs, 24(S), 25-, and 27-OHC, responded to dietary T0901317 by inducing LXR target genes but showed impaired responses to dietary cholesterol, supporting the view that conversion of cholesterol to oxysterols is important for LXR activation in vivo. On the other hand, the recent report by Shafaati et al. [31] showed that overexpression of CYP46A1 in transgenic mice under the β-actin promoter, resulting in significantly elevated 24(S)-OHC levels, failed to induce LXR target genes in either brain or liver. This data argues against a role of 24(S)-OHC, which is in vitro a potent LXR activator, as an important endogenous LXR agonist.The cellular machinery for cholesterol biosynthesis and uptake, as well as for fatty acid biosynthesis, is controlled by transcription factors designated sterol regulatory element binding proteins (SREBPs) [32,33]. The intracellular localization and proteolytic maturation of SREBPs are regulated by a cholesterol-sensor protein called the SREBP cleavage activating protein (SCAP). The SREBPs are synthesized as precursors anchored to ER membranes and form complexes with SCAP. Under low-cholesterol conditions, SREBP-SCAP complexes are transported by a coat protein complex II (COPII)-dependent mechanism to the Golgi apparatus, where SREBPs are proteolytically processed to release a basic helix-loop-helix leucine zipper transcription factor, which enters the nucleus and binds to sterol regulatory elements (SRE) in the promoters of target genes. When sterol builds up in cells, SCAP senses cholesterol in the ER membranes and interacts with Insig (Insulin-induced gene) proteins, and as a result the SREBP-SCAP complex is retained in the ER. The transport of SREBP to the Golgi complex is sensitive to both cholesterol and several oxysterols, of which 25-OHC is most commonly used in experimental set-ups. While SCAP senses ER cholesterol, it was not found to bind 25-OHC [34,35], raising the question of how the inhibitory effect of oxysterols could be mediated. Intriguingly, the Insig proteins were found to directly bind 25-OHC and to mediate the regulatory effect of this oxysterol on SREBP processing. Binding of the Insig-25-OHC complex to SCAP elicits a conformational change similar to that induced by cholesterol binding to SCAP. This precludes the interaction of SCAP with COPII proteins and thereby the transport and proteolytic activation of SREBP [36,37]. Interestingly, Lange et al. [38] found that the mitochondrial cholesterol 27-hydroxylase (CYP27A1) is required for the rapid inactivation of 3-hydroxy-3-methylglutaryl coenzyme A reductase (HMGCoAR) in response to high levels of cholesterol. This is most likely due to 27-OHC interaction with Insig, resulting in the ubiquitination and degradation of HMGCoAR (see [33]). Vertebrates thus have two sterol sensors that control SREBP activity, enabling cells to down-regulate their sterol biosynthesis upon the build-up of either cholesterol or its oxidized derivatives. 24(S),25-epoxycholesterol [24(S),25-EPOX] is generated as a side product of the cholesterol biosynthetic process by the same enzymes that catalyze the synthesis of cholesterol [39]. 24(S),25-EPOX is a potent feedback regulator of cholesterol biosynthesis, suppressing SREBP-2 processing [40] and the cellular HMGCoAR activity [41]. It is also a highly potent oxysterol activator of the LXRs [20,30,42]. The cellular level of 24(S),25-EPOX can be manipulated either by using inhibitors of the cholesterol biosynthetic pathway enzyme 2,3-oxidosqualene cyclase (OSC), resulting in elevation of cellular 24(S),25-EPOX levels, or by overexpressing this enzyme to reduce the cellular content of the oxysterol. Using these tools the groups of A. Brown and M. Huff [43,44,45,46,47] have significantly added to our understanding of the physiologic role of 24(S),25-EPOX. Beyea et al. [43] demonstrated that partial inhibition of OSC in THP-1 macrophages reduced cholesterol synthesis and increased the expression of several LXR target genes, ABCA1, ABCG1, and APOE. Importantly, OSC inhibition did not stimulate lipoprotein lipase (LPL) or fatty acid synthase (FAS), and the observed induction of the lipogenic transcription factor SREBP-1c was counteracted by a block in its conversion to the active nuclear form, supporting the notion that OSC inhibition might have therapeutic potential [44]. However, this idea is undermined by the finding that 24(S),25-EPOX strongly inhibits cholesterol efflux from macrophage foam cells, possibly due to inhibition of cholesterol ester hydrolase function [48]. The findings of Wong et al. [45,46] suggest that synthesis of 24(S),25-EPOX parallels that of cholesterol and fine-tunes the acute control of cellular cholesterol homeostasis, thus protecting cells against accumulation of newly synthesized cholesterol. Interestingly, 24(S),25-EPOX is also suggested to have a specific role in CNS sterol homeostasis. Wong et al. [47] demonstrated that this oxysterol is produced by astrocytes and is taken up by neurons in which it impacts gene expression. The authors suggested that 24(S),25-EPOX may act as a signal from astrocytes that reduces the energetically costly cholesterol biosynthesis by neurons, enabling the neurons to direct resources to other processes essential for neurotransmission.The retinoic acid receptor-related orphan receptors (RORs) are nuclear receptors that, as one of their physiologic functions, have been implicated in the transcriptional control of lipid metabolism. RORα (NR1F1) plays an essential role in development of the cerebellum and in regulation of the circadian rhythm [49], while RORγ (NR1F3) is best known for its role in regulation of T cell development [50]. Both RORα and RORγ are expressed in the liver and play roles in the regulation of glucose and lipid metabolism [51]. The group of T. Burris discovered in 2010 that the 7-substituted oxysterols bound to the ligand binding domains (LBDs) of RORα and RORγ with high affinity, altered the LBD conformation and reduced coactivator binding resulting in suppression of the constitutive transcriptional activity of these two receptors [52]. They recently reported that the two RORs also bind 24(S)-OHC (cerebrosterol), a brain-derived oxysterol, with high affinity, and that 24(S)-OHC functions as a RORα/γ inverse agonist suppressing the constitutive transcriptional activity of these receptors in cotransfection assays [53]. The authors also noted that 24(S),25-EPOX selectively suppressed the activity of RORγ. Interestingly, Wada et al. [54] found that RORα up-regulates expression of CYP7B1, which encodes oxysterol 7α-hydroxylase, an enzyme with a crucial role in bile acid synthesis and cholesterol metabolism. The authors also discovered that LXR target genes are induced in RORα-knock out mice and vice versa, suggesting a mutually suppressive function of these nuclear receptors. These findings indicate that RORα and RORγ serve as novel sensors for oxysterols and display an overlapping ligand preference and functional cross-talk with the LXR, suggesting the presence of an intriguingly complex sterol-controlled network of gene regulatory actions related with metabolic disease and atherosclerosis.The Hedgehog (Hh) signaling pathway plays a central role in the patterning of metazoan embryos, in post-embryonic development, as well as in the homeostasis of adult tissues and stem cell physiology [55,56]. In 2006, the first evidence suggesting that cholesterol or certain oxysterols are required for Sonic hedgehog pathway signal transduction and proliferation of medulloblastoma cells was published by Corcoran and Scott [57]. Dwyer et al. [58] soon demonstrated that the oxysterols 20(S)- and 22(S)-OHC exert osteoinductive effects through activation of the Hh signaling pathway. Consistently, Kim et al. [59] provided evidence that the inhibition of bone marrow stromal cell differentiation into adipocytes by 20(S)-OHC occurs through a Hh-dependent mechanism, and later similar results have been reported with structural analogues of 20(S)-OHC [60]. In addition to activating the Hh signaling pathway, oxysterol-induced osteogenic differentiation was reported to be mediated through a Wnt signaling-related mechanism [61]. Furthermore, 20(S)-OHC was shown to activate Notch target genes, but apparently via a non-canonical mechanism, which may involve the LXR [62]. These findings have established a novel role of oxysterols as regulators of mammalian development. The recent report by Nachtergaele et al. [63] addresses the so far unidentified molecular mechanisms mediating the oxysterol impact on Hedgehog signaling: They found that the most potent oxysterol, 20(S)-OHC, acts as an allosteric activator of the trans-membrane protein Smoothened, which mediates the signal upon induction by Hedgehog ligands.Estrogen receptors are nuclear receptors that, in addition to central functions in reproductive biology, mediate estrogen regulation of a number of other physiologic processes [64]. The cardiovascular protection observed in pre-menopausal females has been largely attributed to beneficial effects of estrogen on endothelial function and the lipid profile [65,66]. Umetani et al. [67] discovered that 27-OHC antagonizes the estrogen-dependent production of NO by vascular cells, resulting in reduced vasorelaxation of aorta, a potentially deleterious effect. Moreover, increasing 27-OHC levels repressed carotid artery re-endothelialization. Also cell type-specific pro-estrogenic actions of 27-OHC were reported [67,68], indicating that this oxysterol acts as an endogenous selective estrogen receptor modulator (SERM). Through its actions on both estrogen receptors and liver X receptors, 27-OHC decreases osteoblast differentiation and enhances osteoclastogenesis, resulting in increased bone resorbtion in mice [69,70]. Furthermore, the action of 27-OHC on estrogen receptors has been putatively linked with breast cancer: A decrease in the expression of E-cadherin and β-catenin, paralleling the loss of adherens junction complex, was observed in MCF7 breast cancer cells exposed 27-OHC, indicating an epithelial-mesenchymal transition characteristic of tumor development [71]. The findings suggest that 27-OHC may counteract the estrogen protection from vascular disease, impair the bone homeostasis, and possibly modify oncogenesis in estrogen receptor-dependent cancers, thus revealing novel, potentially harmful oxysterol functions. Oxysterols are incorporated into biological membranes, and different oxysterols have distinct impacts on membrane lipid packing and especially on the cholesterol-sphingolipid-enriched raft domains with key roles in a number of signal transduction events [72,73]. Modification of membrane biophysical/biochemical properties by oxysterols thus most likely plays an important role in cytotoxicity elicited by these compounds (reviewed by Olkkonen and Hynynen, [74]). However, much of the cytotoxicity attributable to oxysterols is due to their ability to induce apoptosis, an aspect that has been extensively reviewed by Lordan et al. [11]. The two major apoptotic pathways are the mitochondrial (intrinsic) and the death receptor (extrinsic) pathways. Both pathways finish at the execution phase initiated by the activation of effector caspases, which in turn activate endonucleases degrading nuclear material and proteases that break down nuclear and cytoskeletal proteins. Caspase-3, -6, and-7 act as effector caspases at this stage, cleaving a variety of substrates such as poly(ADP-ribose) polymerase (PARP), resulting in the biochemical and morphological manifestations characterizing apoptotic cells [75]. Oxysterols can induce apoptosis by both pathways, and caspase-3, considered the most important effector caspase, is involved in apoptotic processes induced by a variety of oxysterols in different cell types (reviewed by Lordan et al., [11]). However, the distinct oxysterols vary greatly in their ability to induce apoptosis, and also the pathways employed by distinct oxysterols are different. Therefore, there is most likely no universal mechanism responsible for oxysterol-induced apoptosis. When considering the cytotoxic activity of oxysterols, it is important to realize that most studies on the topic have been carried out by applying pure oxysterols on cultured cells. However, when oxysterols are administered as natural-like mixtures, as fatty acid esters, or as incorporated into lipoproteins, their cytotoxicity is strongly alleviated [76,77,78,79].Although there is only limited evidence to activation of the death receptor pathway by oxysterols, certain oxysterols are suggested to induce apoptosis via this route: Lee and Chau [80] demonstrated that 7β- and 25-OHC caused up-regulation of Fas and FasL (Fas ligand) and apoptosis of smooth muscle cells (SMCs) of vascular origin. This finding was supported by the study of Lordan et al. [81] showing that Fas inhibition reduced apoptosis of 7β-OHC-treated cells, and furthermore, that treatment with 7-KC predisposed human aortic SMC to Fas-mediated apoptotic cell death. Oxysterols have in certain cases connected also with tumor necrosis factor (TNF)-α induced apoptotic processes. Lee et al. [82] demonstrated that 7-KC treatment of SMCs resulted in TNF receptor-mediated cell death. It is to some extent unclear whether oxysterols in general have the capacity to induce TNF-α expression and secretion. There is evidence for such induction by 22-OHC and 7β-OHC in human monocytes and THP-1 macrophages [83,84]. However, there is also contradicting data from macrophages [85], and 7β-OHC or 7-KC induced apoptosis of human endothelial cells was not associated with enhanced TNF-α secretion [86].Alterations in mitochondrial transmembrane potential in response to various triggering factors result in the production of reactive oxygen species (ROS) or to mitochondrial membrane permeabilization. This permeabilization can be induced by the proapoptotic Bak/Bax proteins, the interaction of which with voltage-dependent anion channel/adenine nucleotide transporter results in the release of small molecules, such as cytochrome c, apoptosis-inducing factor (AIF), endonuclease G, and smac/DIABLO, activating both caspase-dependent and –independent cell death pathways [87]. Loss of mitochondrial membrane potential and release of cytochrome c upon oxysterol treatment has been documented in many studies (reviewed by Lordan et al., [11]). It is important to note that different apoptotic pathways are induced dependent on the specific oxysterol(s) and the cell type employed; For instance, Ryan et al. [88] showed that in U937 cells, inhibition of cytochrome c release inhibited apoptosis induced by 7β-OHC but not by β-EPOX. Ghelli et al. [89] showed that in ECV304 cells treated with 7-KC no cytochrome c release was associated with the observed loss of cell viability. More mechanistic insight has been gained from recent studies: Kim and Lee [90] provided evidence that the tyrosine kinase inhibitor AG126 reduced 7-KC-induced cell death by suppressing mitochondrial permeability change, apparently via inhibition of ROS generation and GSH (glutathione) depletion. On the other hand, Gao et al. [91] demonstrated that apoptosis of J774 macrophage induced by the oxysterol cholesterol secoaldehyde (3β-hydroxy-5-oxo-5,6-secocholestan-6-al) is mediated via the mitochondrial pathway but without involvement of ROS. Initiation of mitochondrial apoptotic pathway may result from activation of the pro-apoptotic Bak/Bax proteins and concomitant inactivation of the anti-apoptotic Bcl-2 family proteins such as Bcl-2 and Bcl-xL [92]. While some studies have documented oxysterol impacts on the balance of cellular Bax vs. Bcl-2/Bcl-xL levels [93,94,95], no such effect was seen in 7β-OHC-treated Caco-2 cells [96], suggesting involvement of a Bax/Bcl-2 independent apoptotic process. Lordan et al. [97] presented evidence that increase of cytosolic Ca2+ concentration may be an early trigger of 7β-OHC-induced apoptosis in U937 cells—however, it did not seem to play a role in β-EPOX-induced apoptosis. These observations illustrate well the variability in the pathways of cell death induced by different oxysterols in different model systems employed. A variety of protein kinases have been implicated in both the upstream induction phase of apoptosis and as targets of caspases during the actual execution of the apoptosis process, but also in other modes of cell death. The kinases implicated in apoptosis belong to mitogen activated protein kinase (MAPK), Akt/PKB, and the protein kinase C families (reviewed by Lordan et al., [11]). Anticoli et al. [98] reported that 7-KC and 5,6-secosterol modulate differently the stress-activated MAPKs in liver cells: Pathologic concentrations of both oxysterols induced necrosis of the cells after 48 h treatment. 5,6-secosterol, but not 7-KC, induced cell senescence at high concentrations, but caused sustained ERK1/2 (extracellular signal activated kinases) activation and cellular proliferation at low concentrations. 7-KC was reported to induce apoptosis of PC12 neuroendocrine cells via ROS-mediated activation of NF-κB and Akt/PKB pathways [99]. Consistently, Liu et al. [100] found that lanthanum chloride suppresses cholestane-3β,5α,6β-triol induced apoptosis of ECV-304 cells via inhibition of ERK and NF-κB activation. Further, Laynes et al. [101] reported that inhibitors of p38 MAPK, ERK1/2 and JNK (Jun N-terminal kinase) suppressed the cytotoxicity of cholesterol secoaldehyde (ChSeco) in H9c2 cardiomyoblasts. ChSeco is known to induce apoptosis by both intrinsic and extrinsic pathways, and the above study suggested the involvement of ROS such as hydrogen peroxide in the ChSeco cytotoxicity.A number of reports have demonstrated the capacity of exogenously administered oxysterols, mainly 7-KC, 7β-OHC and 25-OHC, to enhance expression of inflammatory mediators, such as IL-8 and MCP-1, by macrophages [84,85,102] and other cell types not belonging to professional immune cells [86,103,104,105]. Importantly, endogenous 25-OHC has been shown to be secreted by dendritic cells and macrophages in response to Toll-like receptor activation, the cholesterol 25-hydroxylase (CH25H) gene expression being regulated by type I interferons via signaling through the interferon-α receptor and the JAK/STAT1 pathway [106,107]. The secreted 25-OHC was shown to suppress IgA class switching in B-cells, demonstrating a crucial function of endogenous 25-OHC in immunity control [106]. On the same theme, two recent studies identified the orphan G-protein coupled receptor EBI2 (Epstein-Barr virus induced gene 2) as a cell surface oxysterol receptor that directs B-cell migration [108,109]. 7α,25-dihydroxycholesterol generated from 25-OHC by oxysterol 7α-hydroxylase CYP7B1 was identified as the most potent ligand of this receptor. The data shows the potency of endogenous oxysterols as immune regulators, and brings up the idea that oxysterols or their synthetic analogues could offer therapeutic benefits either as adjuvants of immune modulators in the treatment of inflammation or autoimmune diseases. Increasing evidence demonstrates the ability of the LXRs, presumably as liganded by endogenous cellular oxysterols, to down-regulate inflammatory signaling (reviewed by Calkin and Tontonoz, [24]). Joseph et al. [23] first showed that LXR activation attenuated the expression of IL-6, iNOS (inducible nitric oxide synthase) and COX-2 (cycloxygenase 2) in macrophages stimulated with E. coli or lipopolysaccharide. Moreover, the LXR were shown to down-regulate the expression of matrix metalloproteinase 9 (MMP-9), a protease abundant in the macrophage-rich regions of atherosclerotic lesions and apparently plays a role in extracellular matrix degradation and plaque destabilization [23,110]. The anti-inflammatory effects of the LXRs have been attributed to their ability to dampen the activity of NF-κB, which controls expression of all of the genes mentioned above, via a mechanism designated transrepression [111]. More recent studies have demonstrated the cross-talk of LXRs with multiple Toll-like receptors (TLR): TLR3/4 stimulation was shown to suppress LXR signaling [112], and the LXRs were reported to dampen atherogenic signaling via the TLR2/TLR4/MyD88 route [113]. Moreover, activation of the macrophage LXRs by apoptotic cells was shown to enhance apoptotic cell clearance with simultaneous down-regulation of inflammatory signals [114]. Importantly, Bensinger et al. [115] also showed that LXR signaling couples the cellular sterol metabolism, via regulation of ABCG1 (cholesterol transporter) expression, with T-cell proliferation in acquired immune response. These investigations have revealed that the LXRs control of number of immune and inflammatory pathways that play important roles in the development of atherosclerosis, and underscore that potential value of the LXRs as future therapeutic targets.Niemann-Pick C (NPC) disease is characterized by accumulation of free cholesterol and sphingolipids within late endocytic compartments of cells [116]. The disease is caused by mutations in either of two proteins designated NPC1 or NPC2. NPC1, mutations of which account for a majority of the disease, is a multi-spanning membrane protein localized in late endosomes, and has a consensus cholesterol binding motif consisting of five trans-membrane helices. The mechanism by which NPC1 facilitates egress of lipids from late endocytic compartments is poorly understood, but it seems to receive cholesterol from the NPC2 protein localizing in endosome lumen [117] and was recently suggested to donate it to the OSBP homologue ORP5 (see section 7; [118]). Interestingly, NPC1 was found to bind not only cholesterol but also 25-OHC, the binding of which was efficiently competed by 24(S)-OHC and 27-OHC [119]. The authors found that cholesterol and 25-OHC bind not to the trans-membrane helix motif but to a luminal loop of NPC1 [120]. However, in further experiments it seemed that the oxysterol binding may not be involved in the classical function of NPC1 in mediating cholesterol transport. StarD5, a cytoplasmic protein belonging to the family of proteins carrying a steroidogenic acute regulatory protein (StAR)-related lipid transfer (START) domain [121], was shown to bind both cholesterol and 25-OHC [122]. Although the physiologic function of StarD5 is poorly understood, it was shown to enhance steroidogenesis in an in vitro assay, evidencing for its ability to transfer cholesterol [123]. As for NPC1, the functional role of 25-OHC binding by StarD5 remains unclear.Families of proteins with sequence homology to the cytoplasmic oxysterol-binding protein, OSBP, are present throughout the eukaryotic kingdom. In mammals the family consists of 12 members. These proteins, designated OSBP-related (ORP) or OSBP-like (OSBPL) proteins, have been implicated in a variety of functions involving cellular lipid metabolism, vesicle transport, and cell signaling (reviewed by Vihervaara et al., [124]). However, the mechanisms of their action have remained poorly understood. Furthermore, many family members have, in addition to oxysterols, been found to bind cholesterol, which is far more abundant in cells. Hence, the identity of the true physiologic ligands of the ORP proteins is not fully solved. In the following we summarize the most important findings on ORP function.OSBP, the founder member of the ORP family, acts as a sterol sensor involved in the regulation of sphingomyelin synthesis via control of the localization and activity of PtdIns-4-kinase IIα and ceramide transporter, CERT, responsible for non-vesicular transport of ceramides from the endoplasmic reticulum (ER) to Golgi where sphingomyelin synthase is located [125,126]. Adenoviral overexpression of rabbit OSBP in mouse liver was shown to result in an increase of plasma very-low-density lipoprotein (VLDL) and liver tissue triglycerides (TG) [127]. Analysis of the underlying mechanism revealed up-regulation of SREBP-1c expression and increase of the active nuclear form of this lipogenic transcription factor in the liver of mice injected OSBP-expressing adenovirus. Silencing of OSBP by RNA interference in cultured hepatocytes attenuated the insulin induction of SREBP-1c and fatty acid synthetase (FAS), as well as TG synthesis. Furthermore, OSBP overexpression was shown to inhibit phosphorylation of the ERKs. In the light of the finding that changes in ERK activity impact the stability of nuclear SREBP-1c [128], this provides one tentative mechanistic explanation to the OSBP overexpression phenotype: Its impact of ERK could modify the stability of nuclear SREBP-1c and thereby hepatic lipogenesis and VLDL secretion. The findings demonstrate a role of OSBP as a sterol-dependent regulator sphingomyelin synthesis and hepatic TG metabolism, as well as a putative function in insulin-induced signaling cascades.Bowden and Ridgway [129] reported that silencing of OSBP by RNA interference resulted in increased cellular amount and cholesterol efflux activity of ABCA1, in the absence of effects on the ABCA1 mRNA level or LXR activity. OSBP knock-down was shown to increase the half-life of the ABCA1 protein, the effect being dependent on an intact OSBP sterol-binding domain. Thus, it seems that OSBP opposes the activity of LXR by destabilizing ABCA1.R. Anderson’s group (Wang et al. [130,131]) identified OSBP as a sterol-sensing scaffolding factor that regulates the dephosphorylation and hence the activity of the ERK, important components of the mitogen activated protein kinase (MAPK) signaling pathways. Romeo and Kazlauskas [132] found that up-regulation of profilin-1, an actin-binding protein implicated in endothelial dysfunction and atherosclerosis, by 7-KC is mediated by OSBP. The signaling route involves interaction of the OSBP-7-KC complex with the tyrosine kinase JAK-2, which phosphorylates Tyr394 on OSBP. This apparently leads to the activation of STAT3, which induces profilin. An important implication of these findings is that also other members of the ORP family could act as lipid sensors with scaffolding functions in cell signaling. Consistent with this idea, Lessman et al. [133] demonstrated that ORP9 contains a phosphoinositide-dependent kinase-2 (PDK-2) phosphorylation site, the phosphorylation of which is dependent on PKC-β or mTOR. ORP9 was shown to interact with these kinases to negatively regulate phosphorylation of the PKD-2 site in Akt/protein kinase B, a major controller of cell survival, cell cycle progression, and glucose metabolism. Zerbinatti et al. [134] demonstrated that OSBP overexpression in a neuroglioma cell line and in HEK293 cells down-regulated the processing of amyloid precursor protein (APP) to β-amyloid (Aβ); OSBP silencing had the opposite effect. OSBP overexpression resulted in the sequestration of APP-Notch2 heterodimers in the Golgi complex, an effect reversed by addition of the OSBP high-affinity ligand 25-OHC. This is consistent with the established findings that the distribution of OSBP itself between cytosol/ER and Golgi membranes is regulated by the cellular sterol status [135,136], and suggests that OSBP modulates the intracellular trafficking of APP. OSBP could thus play a pivotal role in controlling APP metabolism in a sterol-dependent fashion. Localization of OSBP in the Golgi complex is negatively regulated by protein kinase D (PKD)-mediated phosphorylation [137]. OSBP was shown to interact in the Golgi complex with the Hepatitis C virus (HCV) non-structural protein NS5A, and OSBP silencing inhibited secretion of HCV particles by infected cells [138]. Moreover, HCV release was shown to be suppressed by PKD, an effect mediated by phosphorylation of OSBP and its functional partner, CERT [139]. A further implication of OSBP-mediated lipid regulation of microbial infection is provided by the study of Auweter et al. [140] suggesting that OSBP enhances Salmonella replication. As a conclusion, there is increasing evidence for pleiotropic functions of OSBP as a modulator of signaling and intracellular transport events. These properties make OSBP a target that is highjacked by micro-organisms to facilitate their intracellular replication and progeny release. Related with the above report on OSBP, Yan et al. [141] demonstrated that silencing of ORP8 expression in THP-1 macrophages induces the transcription of ABCA1 and, consequently, cholesterol efflux to apolipoprotein A-I. This effect was reproduced using a luciferase reporter assay, in which ORP8 silencing synergized with a synthetic LXR agonist and was significantly suppressed when a mutant ABCA1 promoter devoid of a functional LXR response (DR4) element was used, providing the first solid piece of evidence for a functional interplay between the LXR and the ORPs. Importantly, ORP8 was found to be abundant in the macrophages of human coronary artery lesions, bringing up the possibility that ORP8 may play a role, possibly an adverse one, in the development of atherosclerotic lesions. Our recent work employing mouse macrophage models indicates that the modulation of ABCA1 expression upon ORP8 silencing or knock-out is weak and is most likely mediated by an indirect mechanism (unpublished observations).We recently found that ORP8 overexpression in mouse liver results in a reduction of plasma and liver tissue lipid levels, associated with down-regulation of the active, nuclear forms of SREBP- 1 and -2 [142]. Moreover, ORP8 was found to physically interact with the nuclear pore complex component NUP62, and a normal level of NUP62 in HuH7 hepatoma cells was found to be necessary for the ORP8-mediated reduction of nuclear SREBPs. These findings indicated that ORP8 has the capacity to modulate SREBP-dependent transcription, and it may affect the activity of several transcription factors systems, possibly via an indirect mechanism involving the nuclear pore complex and transport of transcription factors and other nuclear components in and out of the nucleus.Johansson et al. [143] employed ORP1L knock-down to show that the protein is required for the clustering of late endocytic compartments in the pericentriolar region. The protein was shown to form a complex with the late endosomal GTPase Rab7 and its second effector protein RILP. The tripartite complex apparently recruits dynein/dynactin motor to late endosome (LE) membranes to drive minus-end directed motility of the compartments. In a recent study we demonstrated how ligand interactions regulate ORP1L function [144]. By deleting four amino acids from the lid of ORP1L ORD we created a mutant deficient in binding oxysterols (ORP1L Δ560−563). This mutation did not result in translocation of the protein from the LE, but the distribution of LE changed dramatically. Whereas the wild-type (WT) protein induced LE clustering, the LE decorated by mutant ORP1L displayed a scattered phenotype. Similar LE scattering was recently achieved by cellular cholesterol depletion, and ORP1L was shown to be key mediator of the effect [145]. A double mutant deficient in sterol binding and containing a disrupted ER-targeting two phenylalanines in an acidic tract (FFAT) motif (ORP1L Δ560−563 mFFAT) induced LE clustering, demonstrating that destruction of the ER interaction of ORP1L reversed the scattered LE phenotype [144]. A model arising from these experiments and those of Rocha et al. [145] is that non-sterol bound ORP1L bridges between Rab7 on LE and VAPs in the ER, thus forming a membrane contact site (see Levine and Loewen, [146]), which is likely to restrict LE motility; Silencing of ORP1L increased the motility of LE and late endosomal tracks had a more peripheral distribution, probably due to impairment of dynein/dynactin recruitment, or to an altered balance between the minus- and plus-end directed motor complexes Yan et al. [147] showed that transgenic macrophages overexpressing ORP1L increased the size of atherosclerotic lesions in LDL receptor deficient mice. The transgenic macrophages displayed a defect in cholesterol efflux to spherical high-density lipoproteins (HDL) and reduced expression of ABCG1 and apolipoprotein E, as well as increased expression of phospholipid transfer protein (PLTP). All these genes are subject to transcriptional regulation by the LXRs. Furthermore, ORP1L overexpression in cultured macrophages was shown to attenuate the response of the ABCG1 mRNA to the LXR agonist 22(R)OHC, which is also a ligand of ORP1L. The work of Vihervaara et al. [144] showed that ORP1L silencing in mouse Raw264.7 macrophage inhibited the efflux of endocytosed lipoprotein cholesterol to apolipoprotein A-I, thus presumably affecting ABCA1-dependent cholesterol transport. One interpretation of these results is that ORP1L could modulate the LXR ligand interactions, thereby affecting the expression of LXR target genes and the development of atherosclerosis. However, it is also possible that other, more indirect mechanisms account for the observed phenotypic effect. Evidence for ORP function in actual sterol transport was reported by the groups of W. Prinz and A. Menon: Raychaudhuri et al. showed that the Saccharomyces cerevisiae ORP Osh4p is able to extract sterols from liposomes and to deliver them to acceptor membranes. Cells deficient of all seven Osh proteins displayed a marked, 80% reduction in the transfer of cholesterol or ergosterol from plasma membrane (PM) to ER [148]. Involvement of Osh proteins was also suggested in transfer of newly synthesized ergosterol from ER to PM; depletion of all 7 Osh proteins reduced the rate of this transport 5-fold [149]. However, the group of A. Menon recently reached the opposite conclusion: They provided evidence that the role of Osh proteins in ER to PM ergosterol transport is minor, and that Osh protein defect rather distorts the sterol organization in the plasma membrane [150]. Of the mammalians ORPs, OSBP, ORP9L, and the ORP5 ORD have been shown to transfer cholesterol between membranes in vitro [118,151]. Silencing of ORP5, which is anchored to ER membranes, was shown to cause cholesterol accumulation to the limiting membrane of late endosomes. The protein was demonstrated to associate with NPC1, consistent with a model in which ORP5 accepts cholesterol from NPC1 and routes it to the ER [118]. Jansen et al. [152] recently provided evidence that mammalian ORPs have the capacity to stimulate cholesterol transfer in live cells. They used an assay to monitor the transport of a Bodipy-labeled cholesterol derivative from the PM to lipid droplets (LD), and studied the contribution of ORP overexpression to the process. In this study, out of all human ORPs, overexpression of two ORPs of the short subtype, ORP1S and ORP2, gave the largest increase in sterol transport, and silencing both of them simultaneously decreased the transport rate [152]. The effect of the ORPs was further localized to the PM-ER transport step, whereas the rapid ER-LD sterol transport was unaffected by ORP manipulation. In addition, a functional sterol-binding pocket was shown to be necessary for the observed transport enhancement in the case of ORP2. The finding that the PM to ER sterol transfer was most affected by the short ORPs, ORP1S and ORP2, could reflect the ability of these proteins, devoid of PH domains, to associate and dissociate from membranes more rapidly than the long ORPs, perhaps allowing more efficient shuttling between donor and acceptor membranes. This study supports a role for ORPs in sterol transfer but like the studies on yeast Osh proteins, does not provide conclusive evidence for a sterol carrier function. One cannot exclude the possibility that the role of ORP sterol binding might be regulatory, and the observed impacts on sterol trafficking indirect, for instance via the control of membrane contact site formation or the function of other proteins at such sites (reviewed in Levine and Loewen, [146]). Alternatively, modulation of membrane lipid composition by ORPs could affect the ability of membranes to accommodate sterols and thereby alter sterol fluxes, as suggested by Georgiev et al. [150]. ORP3 and ORP7 were found to interact physically with R-Ras, a small GTPase that regulates cell adhesion and migration [153], implying a role of these ORPs in R-Ras signaling. Lehto et al. [154] reported that ORP3 controls cell adhesion and spreading, organization of the actin cytoskeleton, β1-integrin activity and macrophage phagocytic function, cellular processes also subject to regulation by R-Ras. Furthermore, ORP3 phosphorylation was suggested to be regulated by outside-in signals mediated by integrins and cadherins. ORP3 is expressed abundantly in leukocytes and in several epithelia, and its abnormally high expression is detected in certain forms of leukemia and solid tumors, suggesting that the protein may modify cell signaling and adhesion properties in a manner that facilitates malignant growth.Deletion of the S. cerevisiae ORP OSH4/KES1 leads to by-pass of the temperature-sensitivity of mutants in SEC14, a gene encoding a phosphatidylinositol transfer protein (PITP; Sec14p) essential for secretory vesicle biogenesis [155,156]. Osh4p thus acts as a negative regulator of Golgi secretory function, but the underlying mechanism has remained poorly understood. The group of C. McMaster showed that Osh4p reduces both the cellular content of phosphatidylinositol-4 phosphate (PI4P) and its availability for recognition by other proteins, which include components required for transport vesicle formation [157]. Interestingly, Alfaro et al. [158] recently found a direct role for Osh4p in exocytic vesicle transport: They showed that Osh4p associated with exocytic vesicles that move from the mother cell into the bud, where Osh4p facilitated vesicle docking by the exocyst tethering complex at sites of polarized growth on the plasma membrane. Osh4p formed complexes with the small GTPases Cdc42p, Rho1p and Sec4p, and the exocyst complex subunit Sec6p, which was also required for Osh4p association with vesicles. Although Osh4p directly affected polarized exocytosis the relationship of this function with that in sterol trafficking remained unclear. Importantly, de Saint-Jean et al. [159] recently made the ground-breaking finding that Osh4p, in addition to sterols, binds also PI4P employing the same binding pocket. It is able to exchange bound sterol for PI4P and transport the two lipids between membranes along opposite routes. The results suggest a model in which Osh4p transports sterol from the ER to late compartments and, in turn, PI4P backward from trans-Golgi and plasma membrane to the ER. This transport cycle would create a sterol gradient, promoting sterol enrichment in membranes of the late secretory pathway, and play a role in controlling the concentration and distribution of PI4P, with an essential function in secretory vesicle transport.Oxysterols were previously considered to represent harmful substances that accumulate in pathophysiologic states. However, work carried out during the past decade has revealed their physiologic functions as signaling molecules that maintain cellular and body lipid homeostasis and determine cell fate. The discovery of oxysterol interactions with new receptors involved in lipid metabolism, such as the LXRs, Insigs, RORs, the estrogen receptors, NPC1, StarD5 and the OSBP-related proteins, paves the way for future work to elucidate in detail the mechanisms of oxysterol action in lipid metabolism and in diseases such as atherosclerosis and Alzheimer’s disease. Discovery of oxysterols as regulators of Hedgehog signaling uncovered their role in embryonic development, and the recent finding on oxysterol liganding of EBI2 pinpoints a novel effector route for regulation of immunity. These observations have shed light on the diversity of oxysterol activities and opened new perspectives in oxysterol research: With developing oxysterol analytical methodologies and advanced approaches of in vivo genetic manipulation of various organisms and cells in culture, we can expect a plethora of novel connections of oxysterols and their cellular receptors regimes such as differentiation, development, and immunity. The currently known major functions of oxysterols are summarized in Figure 2. The major functions of cellular oxysterol receptors are summarized in Table 2.A schematic presentation summarizing the major functions of oxysterols. ROR, retinoic acid receptor-related orphan receptor; OSBP, oxysterol-binding protein; ORP, OSBP-related protein; LXR, liver X receptor; SREBP, sterol regulatory element binding protein; Insig, insulin-induced gene; ROS, reactive oxygen species; Bcl-2, B-cell lymphoma 2; IgA, immunoglobulin A; CNS, central nervous system. Cellular receptors for oxysterols.* The indicated references are review articles included to avoid excessive listing of literatureWork in the authors’ group is supported by the Academy of Finland (grant 121457), the Sigrid Juselius Foundation, the Finnish Foundation for Cardiovascular Research, the Novo Nordisk Foundation, the Liv och Hälsa Foundation, the Magnus Ehrnrooth Foundation, and the European Union FP7 (LipidomicNet, agreement no. 202272).

Med-MDPI/biomolecules/biomolecules-02-01-00104.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).The phosphorylated kinase-inducible activation domain (pKID) adopts a helix–loop–helix structure upon binding to its partner KIX, although it is unstructured in the unbound state. The N-terminal and C-terminal regions of pKID, which adopt helices in the complex, are called, respectively, αA and αB. We performed all-atom multicanonical molecular dynamics simulations of pKID with and without KIX in explicit solvents to generate conformational ensembles. Although the unbound pKID was disordered overall, αA and αB exhibited a nascent helix propensity; the propensity of αA was stronger than that of αB, which agrees with experimental results. In the bound state, the free-energy landscape of αB involved two low free-energy fractions: native-like and non-native fractions. This result suggests that αB folds according to the induced-fit mechanism. The αB-helix direction was well aligned as in the NMR complex structure, although the αA helix exhibited high flexibility. These results also agree quantitatively with experimental observations. We have detected that the αB helix can bind to another site of KIX, to which another protein MLL also binds with the adopting helix. Consequently, MLL can facilitate pKID binding to the pKID-binding site by blocking the MLL-binding site. This also supports experimentally obtained results. A traditional view related to protein function, is that a folded three-dimensional structure plays a fundamental role as a scaffold to hold the function. However, this view has been modified by discovery of intrinsically disordered proteins (IDPs), which are proteins (or protein regions) that lack a well-defined three-dimensional structure in the isolated state (i.e., unbound/free state), existing as an ensemble of interconverting conformations. Many IDPs interact with partner molecules, transferring to the folded state (i.e., bound state). The tertiary structures of IDPs in the bound state have been determined experimentally. This remarkable feature is called coupled folding and binding [1], which combines two major subjects––protein folding and molecular recognition––each of which has been studied individually in the protein science field.Actually, IDPs differ from ordered proteins in several respects: They are found ubiquitously in transcriptional regulators of eukaryote [2] and they frequently undergo posttranscriptional modification [3] such as phosphorylation. Furthermore, some severe diseases are related to IDPs [4]. Consequently, they have been recognized as potential drug targets [5,6]. A biophysically interesting point is that the intrinsic disorder in the unbound state offers advantages over folded proteins [7]. For instance, molecular association is enhanced by the intrinsic disorder [8].Experiments do not provide sufficient information related to early events in the coupled folding and binding, where the IDP and its partner molecule are separate or weakly interacting with one another. Consequently, theoretical [9] and computational [10,11,12,13] studies might be crucially important to reveal important aspects of their early events. One fascinating mechanism from these studies is a “fly casting” model [9], in which the disordered state allows IDP to capture a distant partner molecule because the disordered polypeptide has a greater interaction radius than a well ordered structure does. After capturing the partner, the disordered polypeptide is reeled to form the native complex. However, Huang et al. [12] have implied that the kinetic advantage derived from the greater interaction radius must be countervailed by its slow diffusion. They have suggested a picture in which the kinetic advantage may not be derived from the greater interaction radius but from a lowered free energy barrier. Then the IDP reaches the final bound form through fewer encounter complexes with its partner than an ordered protein does. Consequently, the association mechanism of IDP remains controversial. All of these binding schemes were derived from simplified protein models. We suggest that an all-atom computation can provide useful insight into this experimentally undetectable process.Most computational studies [11,12,13,14,15] have been conducted using simplified protein models such as a Gō-like coarse-grained model, where one amino-acid residue is expressed usually by one sphere. This model has been used widely to investigate protein folding and molecular recognition because of its low computational cost. A typical Gō-like model postulates that natively formed residue–residue contacts (native interactions) in the native structure are energetically favored, even at a transition state. The other contacts (non-native interactions) likely slow the folding rate. However, the non-native interactions might speed up the folding rate when they help the unfolded polypeptide to collapse, as occurs with competition with chain entropy [16]. If this scheme is correct, then non-native interactions might also positively support or facilitate the IDP–partner association. However, the protein models so far used are too simple to support a realistic discussion of the native and non-native factors. To elucidate these factors, an all-atom protein model is expected to be useful.The all-atom model involves all the interaction factors. However, it is usually difficult to construct a statistically significant conformational ensemble of protein because the conformational space (potential energy surface) is constructed with a huge number of degrees of freedom and the conformation is frequently trapped in energy local minima during a simulation. Consequently, sampling the ensemble requires unrealistically long computation time by conventional simulations. We have overcome this difficulty using an enhanced sampling method: multicanonical molecular dynamics (McMD) [17,18]. The advantage of McMD is that the energy barriers among the energy local minima are overcome, as explained later. Recently, we developed a more efficient sampling method, trivial trajectory parallelization of multicanonical molecular dynamics (TTP-McMD) [19,20]. Using TTP-McMD, we have conducted an all-atom folding simulation of a 57 amino-acid residue protein [21] and the coupled folding and binding of NRSF and mSin3 [22], in explicit solvent. The NRSF is IDP and folds into the helix when binding to the partner mSin3. The simulation reproduced a conformational ensemble at room temperature, which contained the native-like complex structure in the largest population cluster (i.e., the most thermodynamically stable cluster/the lowest free-energy cluster) as well as non-native complex structures in minor clusters. Therefore, the TTP-McMD is a useful computational technique to examine IDP systems.The cAMP-response element-binding protein (CREB) induces transcription via an interaction with its co-activator CREB binding protein (CBP). The transcription factor CREB contains a kinase-inducible domain (KID) to bind the kinase-induced domain interacting domain (KIX) of CBP [23,24]. The binding affinity of KID with KIX depends on phosphorylation [25,26]: the affinity increases as Ser133 of KID is phosphorylated. Both KID and the phosphorylated KID (pKID) are IDPs [27], and the tertiary structure of the pKID–KIX complex was determined using NMR at 315 K (PDB ID: 1kdx [27]). The deposited 17 NMR models show that pKID adopts a helix–loop–helix structure on the KIX surface. Therefore the binding of pKID to KIX is cooperated with folding [28]. Sugase et al. studied the kinetics of pKID binding with KIX by NMR [29]. This system is suitable for all-atom computations because the pKID sequence deposited in PDB is short (28 residues).We note some experimental features of this pKID–KIX system. First, the binding affinity of the C-terminal helix of pKID (called αB; residues 134–145 in the original PDB file) is one order stronger than that of the N-terminal helix (αA; residues 120–131), and formation of helix in αB is necessary for the affinity maintaining, although the helix formation of αA is not [25]. The NMR study [27] has shown that the orientation of αA relative to the KIX framework is disordered, although that of αB is well determined with contacting tightly with KIX. Contrarily, in the unbound state, αA has a higher helix propensity than αB [30]. These features should be confirmed through simulations.As described in this paper, we performed TTP-McMD simulations of pKID in the presence and absence of KIX. We denote the residues 120–131 as “αA residues” and the residues 134–145 as “αB residues” whether these residues form helices in simulation snapshots or not. Similarly, when the αA and αB residues are expressed as elements, they are denoted respectively as an “αA region” and “αB region.” We show that the obtained conformational ensemble from the simulation agrees with the experimental features described above, and that the αA and αB regions have different mechanisms of coupled folding and binding. We designated the simulation system of pKID in the absence of KIX as the “pKID system” and that in the presence of KIX as the “pKID–KIX system.” In the NMR experiment on the pKID–KIX complex [27], the pKID sequence was longer than that deposited in PDB (residues 119–146) because unstructured regions are not deposited. We used the deposited pKID region for the simulation, which is the minimum sequence of pKID binding with KIX (residues 586–666).We prepared the pKID system as follows. The coordinates of pKID were taken from the first model out of the 17 NMR ones. The pKID was immersed in a solvent sphere (called sphere 1; radius = 30 Å), which consists of water molecules with the density of 1 g/mL equilibrated at 300 K in advance. The mass center of pKID was set to the center of sphere 1, and water molecules overlapping with pKID were removed. To neutralize the net charge of the system and make consistency with the ionic strength of the NMR experiment, nine water molecules were selected randomly and replaced with five Cl− and four Na+ ions. Finally, the pKID system consisted of 3473 water molecules, pKID, and nine ions. Although pKID was taken from the NMR model at this stage, the structure was randomized completely using a high-temperature simulation, as described later.Next, we prepared the pKID–KIX system as follows. The coordinates of the two proteins were taken from the first NMR model again. They were immersed in a solvent sphere (called sphere 2; radius = 37 Å). The sphere 2 radius was sufficiently large to contain the entire KIX structure, as described later. The center of sphere 2 was set to the mass-center position of pKID of the NMR model. It is noteworthy that the center of sphere 2 is fixed in space (i.e., the center of sphere 2 is not fixed to pKID) when pKID is moving in the simulation: The translation and rotation of pKID are not restrained. Water molecules overlapping the proteins were removed. Eighteen randomly selected water molecules were replaced with nine Cl− and nine Na+ ions. Then, the pKID–KIX system came to consist of 6166 water molecules, pKID, KIX, and 18 ions. This complex was dissociated by a high-temperature simulation, as shown later.We used the AMBER-based hybrid force field for the proteins. This force field is the combination of AMBER force field parm94 (E94) [31] and parm96 (E96) [32] with a mixing rate ω as E = (1-ω)×E94 +ω×E96 . We set ω= 0.75 because our previous works indicated that ω= 0.75 reproduces the optimal secondary structure preference for some peptides [33,34]. The TIP3P model [35] was used for the water molecules. After energy minimization, we performed a high-temperature (700 K) canonical MD simulation for each of the pKID and pKID–KIX systems to generate the initial simulation conformations for the following TTP-McMD. Figure 1 shows that this temperature was sufficiently high to demolish the native conformation of pKID and to dissociate the native complex. Although pKID was able to move freely in the solvent spheres, the structure of KIX was weakly restrained in the simulation of the pKID–KIX system, as described later. We used the simulation program PRESTO ver. 3 [36]. The canonical MD and TTP-McMD simulations were done in the following condition: 1.0 fs time step, SHAKE algorithm [37] to constrain covalent bonds between heavy atoms and hydrogen atoms, and the cell-multipole expansion method [38] to compute the long-range electrostatic interactions. Throughout the simulation, we maintained the volume of sphere 1 and 2 by supplementing a harmonic potential on atoms flying out of the sphere to pull them into the sphere. Additionally, for the pKID–KIX system, the positions of the main-chain atoms (N, Cα, C, and O) of KIX residues (591–594, 597, 608–611, 617–621, 623–640, 646–648, 660–661) were restrained on those of the NMR model by a weak harmonic potential to maintain the KIX structure around the NMR model. No restraint was applied on pKID to diffuse it freely in sphere 2. Figure 1(a) and Figure 1(b) are the last snapshots of the high-temperature canonical MD simulations of the pKID and pKID-KIX systems, respectively, that are used for the initial structures in TTP-McMD simulations.Initial structures for TTP-McMD simulations of the pKID system (a) and the pKID–KIX system (b). Blue and red ribbons respectively represent KIX and pKID. Small molecules surrounding the proteins are solvent molecules. Arrows below the spheres indicate the sizes of solvent spheres. (c) First NMR model, where KIX is represented by blue ribbon and pKID by rainbow (blue N-terminal; red C-terminal). Red and magenta arrows indicate vectors  and , respectively, which are defined in Section 3.2. (d) pKID sequence, where αA and αB residues are highlighted and character P in the circle represents the phosphate group. Arrows below the sequence indicate the starting and ending residues to define the vectors  and . The structure images were created using Chimera viewer software [39].Before introducing TTP-McMD, we describe the conventional McMD, i.e., the single-run McMD. The single-run McMD is a canonical simulation at a temperature T (a constant-temperature method [40] was used as thermostat in this study) using the multicanonical energy Emc instead of the original potential energy E of the system, asWhere n(E) is the density of states of the system, R signifies the gas constant, Pc(E,T) the probability distribution function of the canonical ensemble at T. If the simulation is sufficiently long, then the single-run McMD provides a flat distribution along the axis of E because the probability distribution is formally given by the following equation.In that equation, Zmc is the partition function for McMD, which can be regarded simply as a factor to normalize the distribution function Pmc To derive Equation (2), we used a formulation of statistical mechanics: Pc(E,T)=n(E)exp(-E/RT)/Zc, where Zc is a factor to normalize Pc(E,T) i.e., the partition function of the system). This flat energy distribution guarantees that the conformation can overcome the energy barriers and visit low energy conformational regions during the simulation. We refer to the entire conformational ensemble obtained from McMD as a “multicanonical ensemble.” Then we can reconstruct a canonical energy distribution Pc at any target temperature Ttag from Pmc asTo derive Equation (3), we used Equation (2) in the following form: n(E)=ZmcPmc(E)exp(Emc/RT). The canonical conformational ensemble at Ttag is constructed by assigning the probability Pmc(E,Ttag) to all conformations in the multicanonical ensemble.McMD uses the probability density function Pc(E,T) in Equation (1), which is generally unknown a priori (before simulation). Then, we must construct it self-consistently through iterative simulation runs, during which Pc(E,T) converges to a precise function. TTP-McMD takes an advantage over the single-run McMD [19,20]. TTP-McMD is technically equivalent with performing independent multiple runs of single-run McMD starting from various initial conformations. The multiple trajectories generated are integrated simply into an ensemble. It is noteworthy that low-energy conformations (low-energy basins) distribute widely in the conformational space, which are spaced by high-energy barriers. Then, a single-run McMD takes a long flight while overcoming the barriers to reach the low-energy basins. On the other hand, the multicanonical algorithm tries to ensure the flat distribution along the energy axis (Equation (2)) so that the conformation is expected to exist evenly in both the low-energy and high-energy regions. This evenness might cause a difficulty of single-run McMD, i.e., no convergence. The connected multiple runs (TTP-McMD), which are spread in the conformational space, are equivalent to a long trajectory, each part of which flights and searches the low-energy basins. Consequently, TTP-McMD provides the convergence of conformational ensemble faster than the single-run McMD.The initial conformations for TTP-McMD were those shown in Figure 1(a) and (b), which were generated from the high-temperature canonical MD simulations, as described above. In this study, we performed 256 multiple runs for each system. To obtain the accurate Pc(E,T) in Equation (1), we performed the simulations iteratively. First, we performed a canonical simulation at 700 K and generated the probability Pc(E,700K) for each system. This probability distribution is restricted around the peak position (designated as Ehigh) of Pc(E,700K), and is accurate only in the narrow region. We designate this energy range as [E0,Ehigh] (E0<Ehigh), where E0 is a lower limit of the accurately sampled energy region. Second, we performed the first TTP-McMD at 700 K with the multicanonical energy Emc, where Pc(E,700K) obtained above was used for Pc(E,T) in Equation (1). In the simulation, we set artificial energy walls at E0 and Ehigh so that the conformation did not escape out of the range [E0,Ehigh]. This simulation produced a flat energy distribution Pmc(E) only in [E0,Ehigh]. Then, using Equation (3) we reconstructed the probability Pc(E,200K). Here, Emc defined by E + RT ln Pc(E,200K) is effective only for multicanonical runs at 200 K. The reason for this temperature reset (i.e., 700 K → 200 K) is explained later. We again extrapolated Pc(E,200K) to an energy range as [E1,Ehigh] (E1<E0), and performed multicanonical runs at 200 K to produce a flat energy distribution Pmc(E) in this extrapolated energy range. Multicanonical runs were performed iteratively until the sampled energy range reaches an energy (Elow) that corresponds to a temperature lower than a room temperature. After reaching Elow, we performed the final TTP-McMD to generate a flat energy distribution in [Elow,Ehigh]. We used 256 computing cores (intel Xeon X5365 3.0 GHz), each core executed one run of TTP-McMD. In the equilibration stage (i.e., the stage to estimate Emc before the final sampling stage), the simulations were done for 21 ns and 23 ns in each of 256 trajectories (total of 5.4 μs and 5.9 μs) of the pKID and pKID-KIX system, respectively. The final runs were done for 8.51 ns in each of 256 trajectories (total time of 2.18 μs) and stored the snapshots every 5 ps for subsequent conformational analyses. The computation times for the equilibration stage were 33 and 76 days for the pKID and pKID-KIX system, respectively. Those for the final sampling stage were 13 and 28 days for the pKID and pKID-KIX system, respectively.As described above, we reset the simulation temperature from 700 K to 200 K. The multicanonical energies at the two temperatures are, respectively, Emc(700K)=700×R ln n(E) and Emc(200K)=200×R ln n(E). Mathematically, the multicanonical ensembles from Emc(700K) and Emc(200K) are expected to be equivalent:.Consequently, the temperature reset is theoretically non-sense. However, the two simulations produce practically different ensembles. We use a polynomial function to approximate Pc(E,T) in Equation (1). The function form Pc(E,200K) is smoother than Pc(E,700K). Then, Pc(E,200K) is more suitable than Pc(E,700K) to define Emc.The conformational ensembles at 315 K of the pKID and pKID–KIX systems were investigated. In Section 3.1, we analyzed the pKID system and showed that pKID is intrinsically disordered in the unbound state. In Section 3.2, we investigated the pKID–KIX system and showed that the free energy landscape of pKID in the bound state is rugged. Furthermore, the orientation of αA of pKID relative to the KIX framework fluctuates more than αB does. In Section 3.3, we discuss the mechanism of the coupled folding and binding for this system. In the final section, differences between the unbound and bound states are discussed.The TTP-McMD simulation for the pKID system is explored evenly in an energy range [−35000.0 kcal/mol, −23480.0 kcal/mol] that corresponds to a temperature range [300 K, 700 K]. From the multicanonical ensemble, we constructed a conformational ensemble at 315 K (denoted as “315K-ensemble”) consisting of 9922 conformations. It is noteworthy that 315 K is the temperature of the NMR experiment [27]. The average of energy E at 315 K was −34272.0 kcal/mol. The secondary structure content at each residue position is shown in Figure 2. Apparently, the αA region preferred helix more than the αB region does. This result agrees with the NMR observation [30]. However, the helix content rate was small: The content for αA was below 50% and that for αB was about 20%, and the unbound pKID was fluctuating among the helix and non-helix conformations at room temperature. Furthermore, the relative positioning between the αA and αB regions fluctuated highly in the 315K-ensemble. Consequently, the simulation confirmed that pKID in the unbound state is intrinsically disordered.α-helix content rate per residue in the pKID system computed from conformations in the 315K-ensemble. DSSP program [41] was used to assign the α-helix to each residue. The x-axis represents the residue ordinal number of pKID in the original PDB file [27]. Arrows indicate the αA and αB regions.We investigated whether the conformation in the unbound state showed similarity with that in the native bound state (i.e., NMR structure). Picking up a sampled conformation from the 315 K-ensemble, the root-mean-square deviation (RMSD) was calculated for each of the αA and αB regions as follows: After superimposing Cα atoms of the αA region onto those in each of the 17 NMR models, 17 RMSD values were calculated and the smallest RMSD of the 17 values, denoted as RMSDαA, was selected. Similarly, the smallest RMSD, denoted as RMSDαB, was calculated for the αB region. The Cα-atomic RMSD for the entire pKID (residues 121–144) is denoted as RMSDall, where raw coordinates of pKID were used for the RMSD computation without superposition. Figure 3 presents the probability distributions of RMSDαA and RMSDΑB for all conformations in the 315K-ensemble. A non-negligible peak was found in the αA region at small RMSDαA: RMSDαA ≈ 1.0 Å. Contrarily, the αB region showed a small peak at RMSDαB≈ 1.0 Å. These results suggest that the αA and αB regions have different mechanisms of coupled folding and binding: αA might bind with KIX with a population-selection mechanism, where the structured fraction of αA is used for binding to KIX. In contrast, αB might bind to KIX with an induced-folding mechanism, where αB is bindable to KIX with various conformations and then the native complex is formed.Probability distributions of (a) RMSDαA and (b) RMSDαB in the pKID system at 315 K.The TTP-McMD simulation of the pKID–KIX system produced a flat energy distribution in an energy range [−64440.0 kcal/mol, −42500.0 kcal/mol], which corresponds to a temperature range [300 K, 700 K]. From the multicanonical ensemble, we derived the 315K-ensemble (8840 conformations). The free energy landscape for the αA and αB regions on binding to KIX was constructed as a function of two variables RαA and RαB defined as follows:  and , where  and  respectively represent the position vectors of the centroids of the αA and αB regions in a sampled conformation, and  and  those in the reference structure (i.e., the NMR model), respectively. Using these distances as the reaction coordinates, we calculated the potential of mean force (PMF) aswhere the probability P(RαA,RαB) was represented as the number of conformations counted in a fraction [i ± 0.5 Å, j ± 0.5 Å] (i,j  = 0,1,2,∙∙∙) of [RαA,RαB] over the total conformations (8840) in the 315-K ensemble, R is the gas constant, and T is the temperature (315 K). Figure 4(a) shows the free energy landscape PMF, which was complex and ragged. We found a low free-energy fraction circled by green circle around (RαA,RαB)=(4 Å, 3 Å). We designate this fraction as the “native-like low-free-energy fraction.” This native-like fraction comprised conformations with RMSDall<7.1 Å. The nearest-native structure, shown in Figure 4(b), had RMSDall of 5.65 Å, whose position in Figure 4(a) is: (RαA,RαB)= (4.43 Å, 2.94 Å). This structure forms a partially disordered αA helix and a well-ordered αB helix.(a) Free energy landscape (potential of mean force (PMF) defined by Equation (4)) constructed on the plane of RαA and RαB. The lowest PMF was set to zero. The green and red circles are described in the text. (b) Smallest-RMSDall (nearest-native) structure of pKID in the pKID–KIX system. KIX is represented by the blue ribbon, and pKID by rainbow (blue N-terminal and red C-terminal). Values for RαA and RαB shown near the structures indicate its position in panel a. (c) Structures taken from the red-circle region of panel a. (d) pKID structure bound at the MLL binding site of KIX via the αB region. (e) MLL (magenta) structure bound to KIX (blue) taken from PDB ID: 2agh.We also found a non-native broad fraction with low free energy in Figure 4(a) (red circle): [RαA,RαB]= [5–13 Å, 0–7 Å]. We designate this fraction “non-native low-free-energy fraction”, where pKID attached KIX with various binding poses as exemplified in Figure 4(c). Because the two fractions were distinguishable in Figure 4(a), a free energy barrier exists between them. The non-native low-free-energy fraction suggests that pKID can bind to KIX with a structural diversity, and the generated various encounters reach the final native-like fraction across the free energy barrier. This result suggests that the high flexibility of pKID might help the pKID–KIX association because pKID binds to KIX without adopting a well-ordered structure. The final native-complex formation is completed after forming the various non-native complexes. Later in this report, we describe our examination of why the non-native low-free-energy fraction was larger than the native-like low-free-energy fraction in the free-energy landscape and why the αA helix is partially disordered, even in the native-like low-free-energy fraction (Figure 4(b)).We found pKID bound to another site of KIX in 323 snapshots of the 315K-ensemble (Figure 4(d)). This site is a binding site for the activation domain of the mixed lineage leukemia (MLL) transcription factor [42]. The MLL–KIX complex structure (Figure 4(e)) shows that a segment of MLL adopts helix and binds to KIX, and the other parts are unstructured. In all of these snapshots, the αB region adopted helix to bind to the MLL binding site with the αA region unstructured. The orientation of the αB helix cylinder was approximately the same as that of the MLL segment in the MLL–KIX complex structure [43]. Consequently, the αB region corresponds to the MLL segment. In fact, both the MLL-binding site and the genuine αB-binding site on the KIX surface consist of hydrophobic amino acids. Furthermore, the hydrophobic residues in the αB region and the MLL segment have similarity; they contains ϕ-x-x-ϕ-ϕ motif (ϕ = hydrophobic residue and x = any residue), which is conserved in many KIX binding proteins (see Figure 9 of reference [43]). It is particularly interesting that in the presence of MLL, pKID binds to KIX with the two-fold higher affinity than pKID in the absence of MLL [44]. Our simulation results suggest that MLL might facilitate the pKID binding to the genuine binding site via blocking the MLL binding site.The orientations of the αA and αB regions with respect to the KIX framework were investigated, respectively, using inner products, IαA and IαB. The inner product IαA was defined aswhere vectors  and  respectively represent the unit vectors of vectors  and : see Figure 1(c). The vector  is pointing from the Cα-atomic position of the 124th residue to that of the 128th residue in a sampled conformation. The vector  is defined in the same way for the NMR structure. For the orientation of αB residues, IαB was defined as The unit vectors  and  were calculated similarly as  and  by replacing the residue numbers 124–134 and 128–141. When the inner products IαA and IαB are 1, the orientations are aligned as in the native bound form.The inner products were calculated for 2476 conformations whose distances [RαA,RαB] satisfy RαA≤13 Å and RαB≤7 Å, which involves the native-like (green circle in Figure 4(a)) and non-native (red circle) low-free-energy fractions. The histogram for IαB (Figure 5(b)) has the largest peak at 1, which indicates that the αB region attaches to KIX with the same orientation as that in the final bound form in both low-free-energy fractions. We discussed the factors stabilizing this orientation later in Section 3.4. Recall that the αB region less adopts helix than the αA region in the isolated state (Figure 2) and that the αB region seldom adopts native-complex form in the isolated state (Figure 3(b)). Consequently, it is likely that the αB region binds to KIX with the right orientation and then the helix is formed. In contrast, the αA region has a large variety in the orientation (Figure 5(a)). These results agree well with the experimental observation: The αB orientation ordered well and the αA orientation does less in the native complex form [27]. Furthermore, the simulation results are consistent with the experimental report that the αB region has a stronger binding affinity than the αA region to bind to KIX [25].Orientational probability distribution of (a) αA and (b) αB regions. Definitions for IαA and IαB are given in the text (Equations 5 and 6).The folding of αA and αB regions upon binding was investigated by measuring residue contacts between pKID and KIX. An intermolecular contact was determined as the distance between the centers of side chains of two residues: If this distance is below 6.5 Å, then we judge that the two residues are contacting. We calculated the contacts (“native contacts”) in the 17 NMR models, where a contact is assigned to a residue pair if the pair is contacting in at least 8 of the 17 models. We found 30 native contacts: 8 between αA and KIX residues, 18 between αB and KIX residues, and 4 between the other residues in pKID and KIX residues. We designate contacts other than the native contacts found in simulation snapshots as “non-native contacts”.We counted the quantities of the native and non-native contacts, respectively denoted as Nnc and Nnnc, for each conformation. Figure 6 presents the helix content rate in pKID at 315 K projected on the Nnc-Nnnc plane. Coupled folding and binding for αB is explained well from Figure 6(c): In complexes where the native contacts are less formed, various non-native contacts are formed. With increasing native contacts, the non-native contacts decrease. Finally the full native contacts are formed with a few (ca. 5) non-native contacts. This result suggests that coupled folding and binding of the αB region accords to the induced-fit mechanism, where the αB region varies the conformation in the encounter complex to reach the native complex. Because the helix content rate of αB is small in the unbound state (Figure 3(b)), the induced-fit mechanism has an advantage over the population-selection mechanism. In fact, the formation of the non-native contacts before the native contacts might facilitate the pKID–KIX association, as discussed earlier in the Introduction: the αB region can bind to KIX via non-native contacts without waiting until a helix is formed.Helix content of pKID at 315 K projected on plane of Nnc and Nnnc. Results are shown for the entire pKID (a), for the αA region (b), and for the αB region (c).However, the αA folding depends less on Nnc than the αB folding did. A few native contacts (ca. 4, which is 50% of the entire native contacts) were able to stabilize the binding of αA to KIX. This result derives from the high directional fluctuations of αA, as shown in Figure 5(a). This result agrees qualitatively with those of the NMR experiment [27], where the flexibility of αA is greater than αB. One can recognize that αA has directional diversity by viewing the NMR models (PDB ID: 1kdx). Furthermore, the fragile contacts between αA and KIX qualitatively support and agree with the experiment [25] showing that the binding affinity assigned to αA is weaker than that to αB. We were unable to specify which of the induced fit or population selection characterizes the αA binding better, although Figure 3(a) supports the population selection. The large flexibility of αA prevents us from answering which is likely. However, we emphasize that the dynamics [27] and the weak binding affinity [25] of αA are characterized by its large flexibility.We demonstrated the change of accessible surface area (ASA) on binding. We denote ASAs for the αA and αB regions respectively as ASAαA and ASAαB. The distributions of ASAαA and ASAαB are presented in Figure 7. The average of ASAαB was 1136 Å2 for the pKID system and 889 Å2 for the pKID–KIX system. Consequently, the reduction of ASA upon binding is clear for αB (Figure 7(b)). This reduction was caused mainly by hydrophobic contacts formed in the complex. However, for αA the reduction was small (Figure 7(a)): the average ASAαA was 1294 Å2 in the pKID system and 1150 Å2 in the pKID–KIX system. This small reduction occurs because the αA region has a large helix probability in the pKID system (Figure 2) and the αA-KIX interaction is weak (Figure 5(a)). These results are consistent with experimental observations [25], showing that αA and αB regions are characterized, respectively, by weak and strong binding affinities.The averages of ASAαA and ASAαB over the 17 NMR models (complex form) are, respectively, 1093±72 Å2 and 723±52 Å2. Consequently, the computed ASA is larger than the experimental value because the computed 315K-ensemble consisted of various complex conformations, as shown in the non-native low-free-energy fraction (Figure 2(c)) and even the conformations in the native-like low-free-energy fraction partially disordered as in Figure 4(b).Normalized probabilities of the accessible surface area (a) for αA (ASAαA) and (b) for αB residues (ASAαB). Dashed and solid lines respectively represent the probabilities for the pKID and pKID–KIX systems.Distribution of the Cα-atom of the phosphorylated SER133 (pS133) in the pKID–KIX system at 315 K. (a) Radial (rpS) distribution function defined as:  where  denotes the Cα-atom position of pS133 in the sampled snapshots and  one in the reference structure (NMR model 1). (b) Spatial distribution of the Cα-atomic position of pS133, where red and magenta contours respectively represent the iso-density regions with density larger than 0.0032 Å–3 and 0.0001 Å–3. The black sphere is the position in the NMR Model 1.A hydrophobic interaction is a non-directional interaction. However, the αB residues were able to bind KIX with the same orientation as in the native complex form (Figure 5(b)). This directional conservation is inferred from a spatial distribution of the phosphorylated SER (pS133), which is located in the middle of the αA and αB (Figure 1(d)). The pS133 interacted with a hydrophilic residue (Y658) and a positively charged residue (K662) of KIX. Figure 8(a) depicts that pS133 restrained the hydrophobic interaction of the αB helix with KIX as in the native complex form. Figure 8(b) demonstrates the spatial distribution of the Cα-atomic position of pS133, where the iso-density contours were computed for conformations with RαA ≤ 13 Å and RαB ≤ 7 Å. Although the high-density site was slightly deviated from that in the NMR structure, the anchor effect of pS133 is clearly shown. Consequently, phosphorylation plays an important role for specifying the αB-helix orientation (Figure 5(b)). We presume that this phosphorylation-induced restraint on the αB helix increases the binding affinity more than non-phosphorylated KID [25,26].We performed the TTP-McMD simulations for the pKID folding and binding with KIX in explicit solvents. Although the overall property of pKID in the unbound state was disordered, pKID has the nascent helix propensity in the αA and αB regions in the computed conformational ensemble. The propensity for αA was stronger than that for αB, which agrees with an experiment described in the literature [30].In the presence of KIX, the free-energy landscape at 315 K involved two low-free-energy fractions: The native-like low-free-energy fraction and non-native low-free-energy fraction. Because the αB region can bind to KIX with various non-native contacts (various encounter complexes), the αB region might provide fast association with KIX [8]. This landscape proposes an induced-fit mechanism for coupled folding and binding of the αB region: various encounter complexes are possible in the early stage, and the complex passes through the free-energy barrier to reach the native-like low-free-energy fraction. The well-oriented binding of the αB region was controlled by the phosphorylated SER133 located in the middle of the αA and αB regions. This control supports the higher binding affinity of pKID than KID, as observed experimentally [25,26]. In contrast to the αB region, the αA region exhibited high flexibility, which agrees qualitatively with that found in the NMR structure [27]. An earlier experiment [25] has demonstrated that the helix formation of αA is not important for binding to KIX. Consequently, the simulation supports the high binding affinity of αB and the low binding affinity of αA.It is particularly interesting that the αB region bound to another shallow hydrophobic concave of KIX than the genuine pKID-binding site, where αB adopted helix. This hydrophobic concavity is the MLL binding site of KID, to which a segment of MLL binds with adopting helix [42]. It has been pointed out in an earlier report that the hydrophobic residue pattern of the αB and MLL segments have a similar hydrophobic amino-acid residue pattern [43]. In the presence of MLL, pKID binds KIX with the two-fold higher affinity than pKID alone [44]. We presume that MLL might facilitate the pKID binding to the genuine binding site by blocking the MLL binding site.The current study demonstrated the importance of hydrophobic interactions between pKID and KIX. Because the multicanonical simulation is an efficient sampling method, the multicanonical trajectory can overcome high potential energy barriers in the conformational space. At the cost of this high performance, the trajectory does not provide information of time series. A conventional MD simulation may provide another important aspect, such as electrostatic interaction, on the complex formation. Finally, it is noteworthy that the non-native low-free-energy fraction (red circle in Figure 4(a)) is larger than the native-like low-free-energy fraction (green circle) in the free-energy landscape and that the αA helix is partially disordered, even in the native-like low-free-energy fraction (Figure 4(b)). These points disagree with the NMR experimentally obtained results [27]. Presumably, this disagreement results from the truncation of pKID, which are unstructured in the NMR experiment. It is important to remember that the computed pKID segment is the only part deposited to PDB. The unstructured region might stabilize the αA helix more, which might result in an increase of the native-like low-free-energy fraction.K.U., M.T. and J.H. were supported by a Grant-in-Aid for Scientific Research on Innovative Areas (21113006) from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) Japan. J.H. and H.N. were supported by grants from the New Energy and Industrial Technology Development Organization (NEDO) Japan.

Med-MDPI/biomolecules/biomolecules-02-01-00122.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Protein tyrosine phosphatase interacting protein 51 (PTPIP51), also known as regulator of microtubule dynamics protein 3, was identified as an in vitro and in vivo interaction partner of CGI-99 and Nuf-2. PTPIP51 mRNA is expressed in all stages of the cell cycle; it is highly expressed six hours post-nocodazole treatment and minimally expressed one hour post-nocodazole treatment. Recent investigations located PTPIP51 protein at the equatorial plate. This study reports the localization of the PTPIP51/CGI-99 and the PTPIP51/Nuf-2 complex at the equatorial region during mitosis. Moreover, Duolink proximity ligation assays revealed an association of PTPIP51 with the microtubular cytoskeleton and the spindle apparatus. High amounts of phosphorylated PTPIP51 associated with the spindle poles was seen by confocal microscopy. In parallel a strong interaction of PTPIP51 with the epidermal growth factor receptor phosphorylating PTPIP51 at the tyrosine 176 residue was seen. In the M/G1 transition a high level of interaction between PTPIP51 and PTP1B was registered, thus restoring the interaction of PTPIP51 and Raf-1, depleted in mitotic cells. Summarizing these new facts, we conclude that PTPIP51 is necessary for normal mitotic processes, impacting on chromosomal division and control of the MAPK pathway activity.Protein tyrosine phosphatase interacting protein 51 (PTPIP51) is a multifunctional protein with implication in processes like proliferation, differentiation and apoptosis [1,2]. To fulfill these opposite functions the PTPIP51 protein possesses different domains necessary for protein-protein interaction and subcellular localization [1,3,4].Data for the involvement of PTPIP51 in mitogenic processes are limited. Mitosis is characterized by four stages, named prophase, anaphase, metaphase and telophase, controlled by various checkpoints [5]. Moreover, distinct signaling pathways are required for entering mitosis. The mitogen activated protein kinase pathway was identified to play a critical role in the G2/M transition [6]. During mitosis itself, Raf-1 is tightly regulated by the raf kinase inhibitory protein (RKIP), which titrates the signal of the MAPK pathway ensuring normal chromosomal segregation. RKIP-depleted cells display chromosomal abnormalities due to incorrect spindle assembly [7]. PTPIP51 is linked to the MAP pathway on the level of Raf-1 [3]. As described recently, these interactions are regulated by the tyrosine phosphorylation status of PTPIP51. Hyperphosphorylation of PTPIP51, as induced by pervanadate inhibits the interaction with 14-3-3β and Raf-1, respectively [8]. Highly tyrosine phosphorylated PTPIP51 also was observed in blasts of acute myeloid leukemia (AML). In addition, due to the lack of the phosphatase PTP1B, PTPIP51 cannot be dephophorylated in cells of AML [9]. These leukemic cells feature autonomous proliferation with a high mitotic index [10]. Thus, phosphorylated PTPIP51 might be involved in proliferation processes of normal as well as cancer cells. The phosphorylation of PTPIP51 might be regulated by the presence or mutation of its interaction partners.PTPIP51 is alternatively named regulator of microtubule dynamics protein 3 (RMD-3). This protein family of microtubule associated proteins comprises three members, named RMD-1, RMD-2 and RMD-3 [1]. RMD-1 was identified to play a key role in mitogenic processes, especially in the correct segregation of the chromosomes during mitosis [11]. Notably, in mitotic cells PTPIP51 is associated with the mitotic spindle apparatus. Error-free chromosome segregation is only achieved if PTPIP51 provides stable attachment of the spindle apparatus to the kinetochores [11]. Yet, the exact mechanism of spindle assembly to the kinetochores remains elusive [12]. As has been revealed by co-immunoprecipitation experiments, PTPIP51 interacts with two mitotic proteins, namely CGI-99 and Nuf-2 [1,2].The human Nuf2 (hNuf2) protein is associated with the outer kinetochore plate and is essential for correct spindle attachment to the kinetochore plate [13]. In addition, the centrosome is also essential for the formation of the mitotic spindle apparatus. Human ninein (hNinein) is one component of the pericentriolar matrix stabilizing the linkage between the microtubules and the centrioles [14]. CGI-99 is an interaction partner of hNinein, a protein associated with the centrosome. Of note, CGI-99 could inhibit the phosphorylation of hNinein mediated by glycogen synthase kinase 3 β, thus influencing the nucleation of the microtubules [15].These findings suggest PTPIP51 may play a pivotal role in mitotic processes, both in chromosomal segregation and cell signal titration. Therefore, we investigated the functional implication of PTPIP51 in cell cycle progression in a human keratinocyte cell line during a synchronized cell cycle. The association of microtubule formation was investigated by double-immunostainings and duolink proximity ligation assays. To gain further insights in phosphorylation events, samples were probed with a peptide specific antibody raised against the tyrosine residue 176 of PTPIP51. PTPIP51 interaction with the epidermal growth factor receptor and c-Src, the phosphatase PTP1B and Raf-1 was investigated by doulink proximity ligation assays. The acquired data suggest an involvement of PTPIP51 in correct chromosome segregation by the spindle apparatus, as well as for uncoupling Raf-1 from the MAPK pathway ensuring correct chromosomal segregation during cell cycle progression.For all experiments of this study HaCaT cells were used kindly provided by Dr. Teschemacher (Department of Pharmacology, Justus Liebig University, Giessen, Germany) with the permission of Dr. Fusenig (DKFZ, Heidelberg, Germany, MTA number L-4598). Cells were kept at 37 °C in humidified 5% CO2 atmosphere and were cultured in RPMI 1640 medium (PAA, Paching, Austria, Cat.# E15-840) supplemented with 10% fetal calf serum (FCS), 100 U/mL penicillin, 100 μg streptomycin. To arrest cells in prophase they were treated with 0.1 μg/mL nocodazole for 12 hours.The total RNA from nocodazol treated cells was prepared using RNeasy kit (Qiagen) according to the manufacturer’s protocol. The isolated mRNA was reverse-transcribed using M-MuLV reverse transcriptase (Fermentas) and poly-dT primer according to standard protocol. The subsequent PCRs were done using Tag-DNA-polymerase (Fermentas). The reactions were optimized for concentration of primers and number of cycles to achieve an almost equal signal from PTPIP51 and actin (for normalization) band in the control samples (untreated cells). The best results were obtained using 22 pmol PTPIP51 and 8 pmol actin primer per 100 µL reaction mix and using 32 to 37 cycles (denaturation: 94 °C 30 sec; annealing: 58 °C 30 sec; elongation 72 °C 1 min 30 sec). The PCR products were separated on 1% agarose gel and stained with ethidium-bromide. The intensities PTPIP51 and actin PCR product were quantified with AIDA software.The PTPIP51 antibody (P51ab) was produced as described below and is consecutively named as P51ab.The cDNA sequence encoding amino acids (aa) 131–470 was inserted into the BamHI and HindIII sites of the plasmid pQE30 and expressed as a His6-tagged protein in the protease-deficient Escherichia coli expression strain AD202 [araD139DE(argFlac) 169 ompT1000:kan flhD5301 fruA25 relA1 rps150(strR) rbsR22 deoC1]. The protein was purified to electrophoretic homogeneity by chromatography on an Ni-agarose column [16]. Immunization of rabbits was performed with 0.5 mg of the purified protein in 0.5 mL RIBI adjuvant, followed by booster injections with 0.5 and 0.3 mg on days 14 and 21, respectively. The antiserum was collected on day 28. Monospecific antibodies were prepared following the method described by Olmsted [17]. Briefly, 2 mg of purified antigen was blotted on nitrocellulose after SDS electrophoresis. The protein band was marked with Ponceau solution and cut out. After blocking of the membrane strip with 1% low-fat milk powder in phosphate-buffered saline, the membrane was incubated with the antiserum for 1 hour, followed by extensive washing with Tris-EDTA-buffered saline. The antibodies were eluted with 0.2 M glycine (pH 2.0) for 2 minutes, followed by immediate neutralization with 1 M triethanolamine.The specificity of the PTPIP51 antibody was tested by ELISA and by immunoblotting of the isolated purified recombinant protein staining bands with 52 kDa, 34 kDa, and 30 kDa. Immunoblotting of homogenates from porcine spleen tissue revealed bands of 48 kDa, 40 kDa, and 29 kDa [18]. The antibody binds to the EGFP fusion PTPIP51 protein expressed in HEK293 [19]. Preabsorbing the PTPIP51 antibody against its antigen completely abolished the immune reaction in all tested samples [20,21,22].For analysis of the tyrosine phosphorylation state of PTPIP51, an antibody (BioLux, Stuttgart, Germany) to the tyrosine 176 phosphorylated sequence DAESEGGYTTANAE was used (P51ab-PTyr). Identity and purity of the synthetized peptide was approved by ESI-MS and UV-analysis. Guinea pigs were immunized with the KLH-conjugated peptides. The specificity of each antibody was tested by ELISA and Western blot. To verify the use of these peptide specific antibodies for immunostaining, preabsorption experiments were performed.Specificity of the PTPIP51 immunoreactivity for both antibodies P51ab and P51ab-PTyr was controlled by preabsorbing both antibodies with the corresponding purified antigen (P51ab: recombinant PTPIP51 full length protein; P51ab-PTyr: phophorylated antigenic peptide described in Section 2.4) at a concentration of 20 µg/mL for 18 hours at 4 °C prior to the immunostaining. As positive control, a normal incubation mixture including the same concentration of PTPIP51 antibody was used. Figure 1 displays in the left panel an immunoblot done with the preabsorbed P51ab-antibody. In Figure 1 right panel an immunoblot done with preabsorbed P51ab-PTyr antibody is shown.Samples of HaCat cell lysate were separated on a 10% SDS-PAGE gel. Transfer on an Immobilon P membrane (Millipore) was performed according to Towbin et al. [23]. The membrane was blocked with 10% fat-free milk powder in PBS. Incubation with polyclonal rabbit anti-PTPIP51 (P51ab) or polyclonal guinea pig anti-pTyr176-PTPIP51 (P51ab-PTyr) was done overnight at room temperature. Either alkaline phosphatase-conjugated anti-rabbit or alkaline phosphatase-conjugated anti-guinea pig immunoglobulins were applied for 1 h at room temperature diluted in 0.5% fat-free milk powder. The reaction was visualized with the SigmaFast BCIP/NBT substrate. A prestained molecular weight marker (Biorad, Cat# 161-0374) was used for calibration.Control experiments for immunoblotting. First panel: immunoblot done with the preabsorbed P51ab-antibody. Second panel: immunoblot done with preabsorbed P51ab-PTyr antibody. Third panel: Negative control with the omission of the P51ab antibody. Fourth panel: Negative control with the omission of the P51ab-PTyr antibody.The Axioplan 2 fluorescence microscope equipped with Plan-Apochromat objectives (Carl Zeiss, Jena, Germany) was used for photo documentation. For visualization of the secondary antibody Cy3-donkey-anti-rabbit an excitation filter with a spectrum of 530–560 nm and an emission filter with a spectrum 572–647 nm were used. Alexa Fluor 488 goat anti-mouse IgG was visualized by an excitation filter with a range of 460–500 nm and an emission filter with a range of 512–542 nm.Confocal images of cells were obtained with a Leica confocal laser scanning microscope (CLSM, TCS SP2, Leica, Bensheim, Germany). Confocal images of Cy3 fluorescence were acquired using Plan-Apochromat ×63/1.4 oil objective, 548 nm excitation wavelengths (helium-neon laser) and a 560–585 nm bandpass filter. The pinhole diameter was set to yield optical sections of 1 Airy unit. For the detection of Alexa 488, we used a Plan-Apochromat ×63/1.4 oil objective, the 488 nm excitation wavelength of an argon laser, and a 505–530 nm band-pass filter. The pinhole diameter was set to yield optical sections of 1 Airy unit. Confocal images of To-Pro-3 (Molecular probes, Cat.# T3605) (nuclear staining) fluorescence were acquired using Plan-Apochromat × 63/1.4 oil objective, 633 nm excitation wavelengths (helium-neon laser), the 650–670 nm bandpass filter. The pinhole diameter was set to yield optical sections of 1 Airy unit. Acquired DIC and confocal images were analyzed and combined using the LCS software (Leica Confocal Software).Acquired images were subsequently processed by ImageJ (v1.43m; Rasband, W.S., ImageJ, U.S. National Institutes of Health, Bethesda, Maryland, USA, 1997–2011) [24] using an iterative deconvolution plug-in by Bob Dougherty [25]. Options were set for all confocal acquired images as follows: 8 numbers of iteration and 2.0 pixels of LP filter diameter. Point spread function was calculated for each channel separately by the ImageJ plug-in created by Bob Dougherty [26].Intensity correlation analysis (ICA) was carried out using ImageJ (v1.43m; Rasband, W.S., ImageJ, U.S. National Institutes of Health, Bethesda, Maryland, USA, 1997–2011) [24] and an appropriate plug-in for ICA included in the plug-in package of the Wright cell imaging facility [27,28].In situ interactions were detected by the proximity ligation assay kit Duolink II (Olink Bioscience, Uppsala, Sweden; PLA probe anti-rabbit minus, Cat.#92005–0100; PLA probe anti-mouse plus, Cat.#92001–0100; PLA probe anti-goat plus, Cat.#92003–0100; Detection Kit Orange, Cat.#92007–0100). The DPLA probe anti-rabbit minus binds to the PTPIP51 antibody, whereas the PLA probe anti-mouse plus or PLA probe anti-goat plus binds to the antibody of the probable interaction partner (see Table 1), respectively. The Duolink proximity ligation assay secondary antibodies only generate a signal when the two DPLA probes have bound, which only takes place if both proteins are closer than 40 nm, indicating their interaction [29]. PFA-fixed HaCat cells were pre-incubated with blocking agent for 1 h. After washing in PBS for 10 min, primary PTPIP51 antibody (1:1000) was applied to the samples. Primary antibodies of the interacting partners (Table 1) were used for proving the interaction by co-incubation with the PTPIP51 antibody. Incubation was done overnight in a pre-heated humidity chamber. Slides were washed three times in PBS for 10 min. Duolink II PLA probes detecting rabbit, goat or mouse antibodies were diluted in the blocking agent in a concentration of 1:5 and applied to the slides followed by incubation for 1 h in a pre-heated humidity chamber at 37 °C. Unbound DPLA probes were removed by washing two times in PBS for 5 min. The samples were incubated with the ligation solution consisting of Duolink II Ligation stock (1:5) and Duolink Ligase (1:40) diluted in high purity water for 30 min at 37 °C. After ligation the Duolink Amplification and Detection stock, diluted 1:5 with addition of polymerase (1:80), was applied to the slides for 100 min. Dapi was used to identify the nuclei. After the final washing steps the slides were dried and cover slips were applied.List of antibodies used in this study.DPLAs were controlled by performing the DPLA with either omission of both primary antibodies, omission of one primary antibody, with P51ab and an antibody detecting a non-expressed cytoplasmic protein (vimentin) or with P51ab and an antibody to a non-PTPIP51 interacting protein (actin) (Figure 2).Control experiments of theDuolink II Proximity Ligation Assay (DPLA) specifity. Negative controls were done for both rabbit and mouse PLA probe and rabbit and goat PLA probe (omission of all primary antibodies). The positive control was done using two antibodies against MEK and Erk, respectively. PTPIP51 interaction with the cytoskeleton was tested by using an actin and vimentin antibody. All antibodies used for DPLA were tested in single staining experiments using the appropriate mixture of PLA probes. Immunocytochemistry (ICC) negative controls were done for the Cy3 donkey anti-rabbit/Alexa Fluor 488 goat anti mouse IgG (ICC negative control rb+ms) mixture and Cy3 donkey anti-guinea pig/Alexa Fluor 488 goat anti mouse IgG (ICC negative control gp+ms) mixture (omission of the primary antibodies for PTPIP51 and Tubulin). Bar: 20 µm.Samples of HaCat cells were analyzed at specific periods after cell cycle synchronization with nocodazole. PTPIP51 mRNA was traced in all samples. The amount of PTPIP51 is stable throughout the observation period of 2–4 hours. At 5 hours after nocodazole treatment an increase in PTPIP51 mRNA was observed. Moreover, after 6 hours post-nocodazole PTPIP51 mRNA amount was increased (Figure 3).Quantitative reverse transcriptase polymerase chain reaction (qRT-PCR) of HaCat cell lysate harvested at distinct periods after nocodazole treatment. (A) Gel of the RT-PCR. Upper bands display β-actin and PTPIP51 mRNA is shown in the lower bands. (B) Relative values of PTPIP51 mRNA at different time points post-synchronization. The value for the control was set to 1. Samples were normalized to the control.Immunoblots of synchronized cells revealed bands at molecular weights of 30 kDa, 38 kDa, 45 kDa 52 kDa, 60 kDa, 65 kDa, 90 kDa, 110 kDa after 2 and 4 hours post-nocodazole treatment.Immunoblots using the peptide specific antibody against the phospho-tyrosine 176 residue of PTPIP51 revealed specific bands at molecular weights of 30 kDa, 38 kDa, 45 kDa, 52 kDa and 60 kDa, 90 kDa, 110 kDa 2 hours after restart of cell cycle. After 4 hours post-nocodazole treatment bands of 30 kDa, 37 kDa, 45 kDa, 52 kDa, 90 kDa, 110 kDa were observed. Yet, a band of 60 kDa was lacking (Figure 4).Immunoblot of nocodazole treated HaCat cells. Cell lysate was analyzed at 2 h and 4 h post-nocodazole. PTPIP51 antibody: P51ab.As seen in Figure 5A mitotic HaCat cells identified by Ki-67 staining displayed PTPIP51 protein, which were marked by Ki-67 staining. Investigating the four mitotic stages PTPIP51 showed a cell cycle stage dependent co-localization with tubulin. In prophase cells no co-localization for PTPIP51 and tubulin was observed (Figure 5B). In contrast, cells of the metaphase, anaphase and telophase displayed co-localization of PTPIP51 protein and tubulin. PTPIP51 overlapped with tubulin at the spindle pole (Figure 5B). The calculated co-localization by ICA, basing on the comparison of fluorescence intensities (see Materials and Methods), is displayed in Figure 6. The co-localization is indicated in yellow to orange and parts with non-co-localization are shown in blue.PTPIP51 protein expression during the cell cycle. (A) Confocal laser scanning microscopic picture of the PTPIP51 protein in proliferating HaCat cells. Cells were treated with nocodazole for cell cycle synchronization. Cells were analyzed in metaphase. PTPIP51 antibody: P51ab, Ki-67 antibody for identification of metaphase. (B) Confocal laser scanning microscopy of PTPIP51 and tubulin in the four mitotic stages. PTPIP51 antibody: P51ab. Nuclei marked in blue using To-Pro3. (C) Confocal laser scanning microscopic picture of the tyrosine 176 phosphorylated PTPIP51 protein and tubulin in the four mitotic stages. PTPIP51 antibody: P51ab-PTyr. Nuclei marked in blue using To-Pro3.Intensity correlation analysis (ICA) of PTPIP51 and tubulin in mitotic cells. Region of interest (ROI) was set to the tubulin staining of each cell of Figure 5B. The co-localization of PTPIP51 and tubulin is displayed in orange. Sites of non-co-localization are marked in blue.Using the phospho-tyrosine specific antibody for co-localization studies revealed that PTPIP51 protein was mostly phosphorylated at its tyrosine 176 residue in all mitotic stages (Figure 5C). In prophase cells tyrosine 176 phosphorylated PTPIP51 was not co-localized with tubulin, substantiating the data seen in immunocytochemical experiments, where no co-localization of PTPIP51 and tubulin was observed. Yet, at the spindle pole tyrosine 176 phosphorylated PTPIP51 protein displayed a co-localization with tubulin corresponding to the observation in immunocytochemical experiments. The computed data of the ICA are displayed in Figure 7. Co-localization is displayed in yellow to orange and non-co-localized parts are shown in blue.Intensity correlation analysis (ICA) of tyrosine 176 phosphorylated PTPIP51 and tubulin in mitotic cells. Region of interest (ROI) was set to the tubulin staining of each cell of Figure 5C. The co-localization of pTyr-PTPIP51 and tubulin is displayed in orange. Sites of non-co-localization are marked in blue.In interphase cells tyrosine 176 phosphorylated PTPIP51 protein also exhibited a partial co-localization with the tubulin cytoskeleton (Figure 8A). The computed data of the ICA is displayed in Figure 8A (ICA). Co-localization is displayed in yellow to orange and non-co-localized parts are shown in blue.A direct interaction of PTPIP51 with tubulin was substantiated by the duolink proximity ligation assay in mitosis as well as in the interphase (Figure 8B). Tyrosine 176 phosphorylated PTPIP51 in HaCat interphase cells and PTPIP51/tubulin DuoLink proximity ligation assay (DPLA). (A) Confocal laser scanning microscopy of tyrosine 176 phosphorylated PTPIP51 and tubulin in interphase cell. PTPIP51 antibody: P51ab-PTyr. Nucleus marked in blue using To-Pro3. Intensity correlation analyses were carried out using the full image of pTyr-PTPIP51/tubulin as input. The co-localization of PTPIP51 and tubulin is displayed in orange. Sites of non-co-localization are marked in blue. (B) DPLA of PTPIP51 and tubulin in mitosis and DPLA of PTPIP51 and tubulin in interphase. PTPIP51 antibody: P51ab. Nuclei are marked by Dapi. Bars: 10 µm.Mitotic as well as interphase cells revealed an interaction with the two mitotic proteins CGI-99 and Nuf-2 (Figure 9). In interphase PTPIP51/CGI-99 interaction was spread throughout the whole cytoplasm (Figure 9A). In mitotic cells of the metaphase PTPIP51/CGI-99 interaction was partially restricted to the region near the equatorial plate (Figure 9B). Interphase cells also displayed an interaction of PTPIP51 and Nuf-2 as seen by DPLA (Figure 9C). The interactions as indicated by the DPLA spots were dispersed throughout the whole cytoplasm and displayed a partial nuclear localization (Figure 9C). Yet, in metaphase cells the PTPIP51/Nuf-2 interaction was traced to the chromosomes of the equatorial plate (Figure 9D).DPLA of PTPIP51 with CGI-99 and Nuf-2 in interphase and during mitosis. (A) DPLA of PTPIP51 and CGI-99 in interphase. (B) DPLA of PTPIP51 and CGI-99 in a mitotic cell of the metaphase. (C) DPLA of PTPIP51 and Nuf-2 during interphase. (D) DPLA of PTPIP51 and Nuf-2 in a mitotic cell of the metaphase. PTPIP51 antibody: P51ab. Nuclei are marked by Dapi. Bar: 10µm.In mitosis as well as in interphase PTPIP51 interacted with the intracellular part of the epidermal growth factor receptor (EGFR), with c-Src, with Raf-1 and with PTP1B as seen by Duolink proximity ligation assay (Figure 10). The magnitude of these interactions strongly varied during the cell cycle (Figure 10). The interaction of PTPIP51 with EGFR was highest in mitotic cells (arrows) compared to interphase cells as seen in Figure 10A. In contrast, PTPIP51 and c-Src interaction was lowest in mitotic cells (Figure 10B, arrow). In interphase cells the number of DPLA spots indicating the PTPIP51 and c-Src interaction were measured in higher amounts (Figure 10B). Such a cell cycle dependent interaction pattern also was seen in cells tested for PTPIP51 and PTP1B interaction. PTPIP51 and PTP1B interaction was highest in post-mitotic cells undergoing cell division (Figure 10C, arrows). The inset in Figure 10C shows nuclei of dividing cells. Moreover, PTPIP51 and Raf-1 interaction levels also varied during cell cycle progression (Figure 10D). Highest interaction levels were found in interphase cells (Figure 10D), whereas mitotic cells displayed single or no DPLA spots for PTPIP51 and Raf-1 interaction (Figure 10D, arrow).DPLA of the interaction partners of PTPIP51: EGFR, c-Src, PTP1B and Raf-1. (A) DPLA of PTPIP51 and EGFR. Mitotic cells are marked by arrows. (B) DPLA of PTPIP51 and c-Src. Arrow: mitotic cell. (C) DPLA of PTPIP51 and PTP1B. Dividing cell is marked by the arrow. The inset shows the nuclei of the dividing cell marked by the arrow. (D) DPLA of PTPIP51 and Raf-1. Dividing cell is marked by the arrow. PTPIP51 antibody: P51ab. Nuclei are marked by Dapi. Bar: 10 µm.Keratinocytes grown under normal conditions were submitted to cell cycle synchronization by nocodazole. The synchronized cells were grown under standardized conditions. After restart of the cell cycle, cells were harvested at defined time points. PTPIP51 protein was highest 2 hours after the restart of the cell cycle where mitotic cells were most numerous (data not shown), whereas the mRNA expression reached a maximum 6 hours post synchronization.Synchronized HaCat cells expressed different molecular forms of unphosphorlyated-PTPIP51 and phosphorylated-PTPIP51. The exact function of the phosphorylation at tyrosine residue 176 of PTPIP51 is still unknown. The absence of the 60 kDa phosphorylated PTPIP51 form 4 hours post synchronization probably hints to an involvement in mitotic processes. Two studies from our laboratory pointed to a regulation of the PTPIP51/Raf-1 interaction by tyrosine phosphorylation [8,9]. In cells of acute myeloid leukemia PTPIP51 is unable to interact with Raf-1 while phosphorylated at its tyrosine 176 residue [9]. In general, this indicates a mechanism for conducting PTPIP51 to specific cellular compartments and specialized cellular functions. This is corroborated by the present data. In mitotic cells PTPIP51 is almost completely phosphorylated at its tyrosine 176 residue. In interphase cells tyrosine 176-phosphorylated PTPIP51 is found in a more dispersed manner. Oishi and co-workers [11] found ectopically expressed PTPIP51 in association with the spindle pole of the forming spindle apparatus in mitotic cells. Such a co-localization is substantiated by the present study. In addition, the co-localization of the tyrosine-phosphorylated form with tubulin at the spindle pole was observed. Moreover, the direct interaction with tubulin was corroborated. These facts point to a pivotal role of tyrosine phosphorylated PTPIP51 in the formation of the spindle apparatus. The spindle apparatus enucleates from the doubled centrosome. Here, ninein is a key player for correct enucleation and anchorage of the minus ends of the microtubules [30]. Ninein was found to interact with CGI-99 and thereby the phosphorylation of ninein by glycogen synthase kinase 3 β (GSK3β) is inhibited [15]. Interestingly, PTPIP51, CGI-99, ninein and GSK3β are located at the centrosome [15,31]. The complex of these four proteins probably regulates the polymerization of the microtubules. The PTPIP51/CGI-99 complex possibly influences the interaction of GSK3β and ninein and subsequently the association of ninein to the centrosome. Thus, ninein is unable to influence anchorage and enucleation of the microtubules. Yet, more data are needed to elucidate the full regulatory network formed by PTPIP51, CGI-99, ninein and GSK3β. Furthermore, ninein is released from the centrosome in speckles with restriction to the microtubular cytoskeleton. Ninein can be transported to and from the centrosome and may play a role in cell polarization of epithelial cells [31]. In this study PTPIP51 was associated with tubulin and CGI-99 in interphase cells. The interaction of PTPIP51 and CGI-99 could be traced throughout the whole cytoplasm. Thus, the interaction of PTPIP51/CGI-99 may also influence microtubular association and cell polarity by the modulation of ninein.Furthermore, PTPIP51 interacts with the kinetochore protein Nuf-2 mediating correct recruitment of the microtubules to the kinetochore during mitosis [13]. Nuf-2 is part of the Ndc80 complex next to Hec1, Spc24 and Spc25 [32]. Spc24 and Spc25 anchor the complex into the kinetochore. Using these two proteins as a base, the N-terminal domains of Nuf2 and Hec1 interact with the plus ends of the spindle microtubules [33,34]. All four members of the NDC80 complex are predicted to inherit coiled-coil domains. The NDC is divided into two subcomplexes: the Hec1-Nuf-2 complex and the Spc24-Spc25 complex. These two subcomplexes are stabilized by the coiled-coil domains [35]. PTPIP51 is also found to have a coiled-coil domain [1]. Thus, PTPIP51 may be a member of the outer kinetochore NDC80 complex modulating anchoring and function of the spindle apparatus microtubules. In concordance, PTPIP51/tubulin interaction was also observed at the dividing chromosomes. This goes along with the observations of Kittler et al. [36] who reported defect cell division after knockdown of PTPIP51 gene. Summarizing these facts, we postulate that PTPIP51 plays a pivotal role in the formation of the spindle apparatus with impact on the minus and plus ends of the microtubules by its interaction with CGI-99 and Nuf-2.As discussed earlier, the recruitment of PTPIP51 to the spindle apparatus is probably induced by its tyrosine 176 phosphorylation through c-Src [2]. Moreover, Brobeil and co-workers showed that Lyn, a member of the Src family kinases, also interacts with PTPIP51 in cells of acute myeloid leukemia [9]. Strikingly, during mitosis cdc2 phosphorylates c-Src at specific serine and threonine residues. In parallel, tyrosine kinase activity is increased approximately two-fold and accessibility of its SH2 domain for binding relevant phosphotyrosine-containing ligands increases by about 15-fold [37]. In the present study cells undergoing mitosis showed reduced levels of PTPIP51/c-Src interaction. In contrast, PTPIP51/EGFR interaction was increased in mitotic cells mirroring the main phosphorylation event of PTPIP51 at its tyrosine 176 residue during mitosis. Dangi and Shapira [38] showed that the downstream pathways of the EGFR signaling are uncoupled through phosphorylation mediated by the cdc2 kinase. During mitosis transcription and translation have to be inhibited due to their high energy demand. Extracellular regulated kinase (Erk) is the downstream molecule of the EGFR signaling coupled by the Ras-Raf-MEK cascade [38]. Cdc2 uncouples this cascade and there is no Erk activation during mitosis [38]. Moreover, PTPIP51 is known to activate Erk on Raf-1 level by its interaction with 14-3-3 protein and formation of a PTPIP51/14-3-3/Raf-1 complex [3]. This interaction is reduced if PTPIP51 is tyrosine phosphorylated [8]. Thus, the phosphorylation of PTPIP51 at its tyrosine 176 residue by the EGFR receptor with subsequent inhibition of Raf-1 activation may be a further control mechanism for correct mitosis and inhibition of Erk activity. This is in concordance with the acquired data on PTPIP51/Raf-1 interaction in post-mitotic dividing cells, where no or only single interactions of PTPIP51 with Raf-1 were observed. Interestingly, in post-mitotic dividing cells PTPIP51/PTP1B interaction is strongly increased. PTPIP51/PTP1B interaction is thought to be an enzyme-substrate complex [2]. Thus, PTP1B dephosphorylates PTPIP51 at its tyrosine 176 residue to function as a modulator of the mitogen activated protein kinase (MAPK) pathway in interphase cells.In conclusion, PTPIP51 is a newly recognized protein with functional implication in spindle apparatus formation in mitotic cells. The recruitment of PTPPI51 is mediated through distinct tyrosine phosphorylation at the tyrosine 176 residue mediated by the EGFR. Tyrosine 176 phosphorylated PTPIP51 also ensures the uncoupling of the MAPK pathway to assure the energy supply for mitosis.We thank Dietmar Schreiner, Biozentrum, University of Basel, Switzerland for the determination of PTPIP51 mRNA.The authors declare no conflict of interest.

Med-MDPI/biomolecules/biomolecules-02-01-00143.txt

@@ -0,0 +1,12 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Mortalin is a highly conserved heat-shock chaperone usually found in multiple subcellular locations. It has several binding partners and has been implicated in various functions ranging from stress response, control of cell proliferation, and inhibition/prevention of apoptosis. The activity of this protein involves different structural and functional mechanisms, and minor alterations in its expression level may lead to serious biological consequences, including neurodegeneration. In this article we review the most current data associated with mortalin’s binding partners and how these protein-protein interactions may be implicated in apoptosis and neurodegeneration. A complete understanding of the molecular pathways in which mortalin is involved is important for the development of therapeutic strategies for cancer and neurodegenerative diseases. Mortalin is a 74 kDa mitochondrial-resident protein also known as p66mot-1 [1], mitochondrial stress-70 protein (mtHsp70) [2], peptide-binding protein 74 (PBP74) [3], and GRP75 [4]. Despite not being a heat-activated protein, based on sequence similarity, mortalin has been classified as another member of the heat shock protein 70 (Hsp70) family of chaperone proteins [5]. Mortalin, which is encoded by the nuclear gene HSPA9B (GeneID: 3313) [6,7,8], contains an N-terminal 46-amino-acid-long signal peptide that undergoes calcium-dependent autophosphorylation [9]. Genomic analysis revealed the presence of 2.8 kb human mortalin transcribed from an 18 kb region on chromosome 5q31.1 consisting of 17 exons with boundaries almost identical to its murine counterpart [10], and the first intron interrupted in the N-terminal leader sequence, a pattern similar to that of cytochrome-c (cyt-c), another mitochondrial protein [11].Mortalin is translated in the cytoplasm and is transported into mitochondria [12]. The crystal structure of mortalin has not yet been elucidated; therefore using amino acid sequence comparison and molecular modeling we developed a potential 3D structure (Figure 1). This 3D structural representation suggests that mortalin has two functional domains: an ATPase, N-terminal nucleotide-binding domain (NBD) and the C-terminal substrate-binding domain (SBD) [13]. The biochemical activities of each domain are essential for both general and specialized chaperone functions [14].Despite being predominantly localized in the mitochondria [1,15,16], mortalin has also been found in other sub-cellular compartments, including the endoplasmic reticulum [17], cytoplasmic vesicles [18], and the cytosol [2,17,19]. Mortalin activity and function are determined by its localization in the cell and by its binding partners (Table 1, and Figure 2). Several post-translational modifications (PTMs) have been found in mortalin, including phosphorylation, oxidation, and ubiquitination [19]. We found that mortalin is likely to be differentially phosphorylated in brain samples from Alzheimer’s disease patients [20], and that it is oxidized in the brains of hAPOE targeted replacement mice [21]. Further confirmation of mortalin phosphorylation, identification of the specific phosphorylation sites, and elucidation of the biological effects of differential phosphorylation on mortalin function are still in progress.Molecular modeling of Mortalin. Representation of the 3D structure of mortalincreated by homology modeling with the program PyMOL (The PyMOL Molecular Graphics System [22]) and energy-minimized with Hyperchem 8.0 (Hypercube, Inc. Gainsville, FL. USA). The (N-terminal binding domain, NBD; amino acid residues 1–443) includes the N-terminal region (blue) and includes the ATP binding motif (amino acid residues 61–443; indicated in green); the substrate binding domain (SBD; amino acid residues 444–679) is shown in yellow and includes the peptide binding domain (PBD; amino acid residues 444–581, indicated in red) [2,5,12]. p53 binds mortalin somewhere in the peptide-binding domain of mortalin (black arrow) [23].Mortalin is a stress response protein induced by metabolic stress, glucose deprivation [24,25], the calcium ionophore A23187 [26], thyroid hormone treatment and hyperthyroidism [27], ionizing radiation [28] and some cytotoxins [19]. Increasing levels of mortalin expression are associated with cellular protection, as they permit cells to survive lethal conditions [29,30,31]. Mortalin has also anti-apoptotic [15] and pro-proliferative activities [32]. Mortalin accelerates the immortalization of normal human cells in cooperation with telomerase [33], and influences the function, dynamics, morphology, and homeostasis of mitochondria [15].Depending on its localization and its binding partners, the following functions have been associated with mortalin: control of cell proliferation [34], intracellular trafficking [35,36], guidance of other proteins to their final localization [34], antigen processing [3,37], regulation of cell response to stress conditions [25,26,27,38], regulation of cell response to variation in glucose levels [25], receptor internalization and muscle activity [39], in vivo nephrotoxicity and cell fate determination [40], inactivation of the tumor suppressor protein p53 [34,41,42], and inhibition of apoptosis (programmed cell death) [32]. All of these functions and the corresponding binding partners are summarized in Table 1 and are represented in Figure 2.Multiple functions and multiple localizations of mortalin. Mortalin is involved in mitochondrial, nuclear, plasma membrane and endoplasmic reticulum processes. The distribution of mortalin is highly dependent on cellular conditions. Mortalin interacts with the following proteins: 1. mitochondrial pre-proteins interact with mortalin and Hsp60 upon entering the mitochondrial matrix compartment; following these interactions, the mortalin/Hsp60 complex acts as a mitochondrial import motor. This coupling process allows proteins to refold, assemble, sort, and perform their corresponding functions; 2. mortalin interacts with VDAC1 and modulates its channel properties; 3. p66Shc localizes into the mitochondria and forms a complex with mortalin that modulates the mitochondrial pathway of apoptosis; 4. binding of mortalin to MVD1 (that inhibits p21(ras)-induced growth arrest) may represent another pathway to immortalization and may be a part of mechanisms of cell proliferation; 5. mortalin associates with the IL-1R (interleukin-1 receptor) protein leading to receptor internalization and downstream signaling cascades; 6. mortalin binds p53 thereby inactivating p53 translocation to the nucleus and inhibiting its activity as an apoptosis inducer; and 7. mortalin promotes intracellular trafficking of FGF-1.Proteins that bind or regulate mortalin and corresponding functions.Mortalin is involved in multiple basic mitochondrial processes, including energy metabolism, free-radical generation [31], and maintenance of mitochondrial protein integrity [19,59]. In addition, mortalin plays a role in mitochondrial biogenesis, translocation of cytosolic protein precursors, and their partitioning within the matrix and across the two mitochondrial membranes [23,60,61] (Figure 2). Mortalin is the only ATPase component of the pre-protein mitochondrial import machinery [62,63] where it binds the translocase of the mitochondrial inner membrane (TIM) to form an ATP-dependent motor [15,64,65]. In the mitochondrial matrix, mortalin floats freely as it participates in protein folding in association with Hsp60 [5].In eukaryotic cells the majority of mitochondrial precursor proteins (pre-proteins) are synthesized in the cytosol, recognized by receptor proteins on the mitochondrial surface, and translocated across the mitochondrial membranes [64,65] via specific transport machines. These machines include the Translocase of the Outer Membrane (TOM, [66]) and the Translocase of the Inner Membrane (TIM, [15,65,67]). These molecular machines are used to bring the translocated proteins to their final destination in the mitochondria, including the intramembrane space (IMS), the inner membrane, and the matrix [68,69,70]. Many mitochondrial precursor proteins are chaperoned into the mitochondrial matrix by mortalin, using an ATP-dependent mechanism and the assistance of co-chaperones [69,71]. Some specific details of the mortalin-associated translocation mechanism are shown in Figure 3 and include:
+        
+          
+            Tim44, a matrix protein, associates simultaneously the Tim23 complex (the translocation channel in the inner membrane) and mortalin [72,73].
+          
+          
+            Tim14 (Pam18 or DNAJC19), a J-domain protein, stimulates mortalin’s ATPase activity [74].
+          
+          
+            Tim16 (Pam16 or Magmas), which controls the activity of Tim14 and Mge1 (hMge1), stimulates the release of adenosine diphospate (ADP) [14,73].
+          
+        Tim44, a matrix protein, associates simultaneously the Tim23 complex (the translocation channel in the inner membrane) and mortalin [72,73].Tim14 (Pam18 or DNAJC19), a J-domain protein, stimulates mortalin’s ATPase activity [74].Tim16 (Pam16 or Magmas), which controls the activity of Tim14 and Mge1 (hMge1), stimulates the release of adenosine diphospate (ADP) [14,73].The matured protein is then transferred by mortalin to the Hsp60 protein. Hsp60 allows proteins to refold back, assemble, sort and finally perform their duties as components in the bioenergetics network [5]. This coupling is essential for maintaining the mitochondrial proteome integrity [19].Mitochondria are vulnerable to oxidative damage, including oxidative stress (OS) in which free radicals modify proteins, lipids, and nucleic acids [75]. Under normal conditions, the mitochondrial electron transport results in production of reactive oxygen species (ROS) [31] that, in excess, can result in cellular membrane damage and cellular dysfunction [76]. ROS are a causal step in apoptosis and a key element in some neurodegenerative diseases [15]. The importance of mortalin in ROS-associated neurodegeneration stems from the fact that mortalin inhibits ROS accumulation in the mitochondria [24,45,77]. Glucose deprivation causes a rapid increase in ROS accumulation, which is reduced by mortalin over-expression, suggesting that mortalin has a cytoprotective effect and could decrease the ROS accumulation maintaining cell viability [24].In multicellular organisms, cells that are no longer needed are destroyed by a regulated process known as apoptosis [78,79,80]. Apoptosis is important for embryo development, tissue homeostasis, and regulation of the immune system as well as for the development of the nervous system [79,81]. Apoptosis may play a role in neurodegeneration and aging [80,82].There are two apoptotic pathways in mammals; i.e., the extrinsic and the intrinsic pathways [83]. Both of these pathways involve the activation of caspases—proteolytic proteins that cleave their target polypeptides at specific locations without degradation of the target protein [84,85], thus creating gain-of-function or loss-of-function events that generate the apoptotic phenotype [85].The intrinsic, or Bcl-2-regulated, mitochondrial pathway is triggered in response to several cellular stressors. Bcl-2 homology 3 (BH3)-only members either directly activate the pro-apoptotic Bcl-2 family members Bax and Bak, or antagonize anti-apoptotic Bcl-2 family members [86,87]. Bax and Bak are thought to homo-oligomerize and form pores in the outer mitochondrial membrane thereby increasing mitochondrial outer-membrane permeabilization (MOMP), considered as the ‘point of no return’ in apoptosis signaling. The MOMP allows the efflux of multiple pro-apoptotic proteins from the mitochondrial intermembrane space including cyt-c [85]; as a result, second mitochondrial activators of caspases (Smac or DIABLO) can pass from the intermembrane space into the cytoplasm. In the cytosol, cyt-c interacts with an adaptor protein, the apoptotic protease-activating factor-1 (Apaf-1), a crucial step of the intrinsic pathway [88]. The interaction between cyt-c and Apaf-1 induces Apaf-1 conformational changes driven by dATP hydrolysis [89]. The resulting complex recruits and binds pro-caspase-9 to the caspase recruitment domain (CARD) of Apaf-1 [84]. Caspase-9 is activated in a dATP/ATP-dependent process and the resulting complex cleaves, and activates caspases 3 and 7 [84,89,90,91]. These proteins mediate the molecular signals leading to cell death through the selective proteolysis of key protein substrates.p53 is a key tumor-suppressor protein that abolishes genetically-unstable cells by inducing cell cycle arrest or apoptosis through transcriptional regulation or by direct interaction with apoptotic proteins [92]. p53 can be inactivated by post-translational modifications, mutations [93], or as a result of sequestration by binding proteins [94,95,96]. p53 has been implicated in transcriptional activation of several proteins (such as Ras, p21, Bax, BH3-only proteins Noxa and PUMA (p53-up-regulated modulator of apoptosis), PIG3, Killer/DR5, CD95 (Fas), p53AIP1 and Perp) or repression of genes involved in apoptosis [97]. Some studies have reported functional interactions between p53 and mortalin in the cytoplasm [23,52,53,92,98,99], leading to the inhibition of the transcriptional activation of p53 and control of centrosome duplication functions [92,99]. Specifically, p53 presents two binding sites for mortalin, one in the C-terminal domain and the other in the p53-tetramerization (TET) domain [23]; any of these domains is sufficient for a mortalin-p53 binding interaction. This interaction occurs through the PBD of mortalin [23,99] (Figure 1).A recent study indicates that the mortalin-p53 interaction causes inactivation of p53-mediated apoptosis depending on the cellular stress levels [92]. Specifically, stress-associated induction of mortalin expression protects the cells against the initial insult, improves cell recovery, and improves resistance to subsequent stress signals. Unstressed or mildly stressed cells do not display mortalin-p53 interaction [92] (Figure 3). Mortalin may prevent the entry of p53 to the nucleus by physical entrapment that leads to proteasomal degradation [52]. During the late G1 phase, mortalin localizes in the centrosome and represses the p53-dependent suppression of centrosome duplication [98]. p53 can induce Bax activation, leading to changes in the mitochondrial membrane permeabilization; however, in the absence of mortalin, there is nuclear accumulation of p53, concomitant with increased levels of Bax, suggesting that the low levels (or absence) of mortalin cause activation of the p53-Bax apoptosis pathway [92]. Mortalin may also be associated with cell immortalization via binding to diphosphomevalonate decarboxylase (MVD1; previously known as mevalonate pyrophosphate decarboxylase or MPD), an inhibitory protein of p21 (Ras). Furthermore, co-expression of the human telomerase reverse transcriptase (hTERT) with mortalin can avoid cell death [33].Another important protein in the mortalin/p53/OS-associated molecular mechanisms is the 66 kDa isoform of the SHC-transforming protein 1 or p66Shc, a protein that mediates OS-induced apoptotic mechanisms [53]. p66Shc is a downstream target of p53 that predominantly exists in the cytoplasm and is translocated into the mitochondria in a process mediated by mortalin and prolyl isomerase 1 (Pin1) [100]. In mitochondria, following pro-apoptotic stimulation, p66Shc oxidizes cyt-c, producing H2O2, which promotes the opening of the mitochondrial permeability transition pore triggering apoptosis [53].The literature review presented here suggests that mortalin participates in apoptosis by regulating proteins that are implicated in cellular stress response mechanisms. Under low levels of stress, mortalin acts as an anti-apoptotic protein by inactivating p53 [32,92], and interfering with the ability of cyt-c and Apaf-1 to trigger the recruitment of procaspase-9; on the other hand, under stress conditions, mortalin alters mitochondrial functions while cytoplasmic p53 can induce apoptotic signals (Figure 3). These opposing functions point out to a mortalin protein that may represent a sensitive marker of stressed cells and apoptotic function associated with p53 activity.Role of mortalin in oxidative stress-induced apoptosis. Mortalin has different functions under cellular stress or under non-stressed conditions. 1. Exposure of cells to stress induces the phosphorylation of p53 and its interaction with mortalin. Mortalin tries to protect the cells against oxidative damage; however, if the cells cannot recover, p53 induces the transcriptional activation of Bax, and BH3-only proteins including Noxa and PUMA, resulting in apoptosis; 2. increased levels of cellular oxidative stress can alter mortalin’s function; 3. In non-stressed (normal) or mildly-stressed conditions the phosphorylation levels of p53 are low, and mortalin does not interact with p53.Aging is a biological process characterized by a general and progressive deterioration in metabolic processes affecting tissues that exhibit a high rate of oxygen consumption, such as the brain [19]. Aging and neurodegeneration also affect the proteome. Oxidative protein damage results in protein aggregation, changes to secondary and tertiary structures, and loss of catalytic functions that may activate cell death-associated signal transduction pathways. Unfolded proteins have a strong tendency to form neurotoxic insoluble protein aggregates resulting in the impairment of the ubiquitin-proteasome degradation system, and suppression of the heat shock and OS response mechanisms [101]. The abnormal accumulation of unfolded and/or aggregated polypeptides usually leads to the loss of specific neuronal populations resulting in the onset and progression of several neurodegenerative diseases [34,77]. In general, the coupling of stress with impairment of the chaperone system can cause premature aging [19].Potential interaction between mortalin and ApoE in Alzheimer’s disease. (a) Astrocytes from hApoE ε2/2-TR, hApoE ε3/3-TR, and hApoE ε4/4-TR mice were solubilized and immunoprecipitated with mortalin (Mot-Ab) or ApoE (ApoE-Ab) antibodies (Left panel). The ApoE-Ab immunoprecipitants were challenged with the Mot-Ab and only the hApoE ε4/4-TR astrocytes displayed interaction between ApoE and mortalin ((a) Right panel, indicated with a white arrow on the right panel). (b) Human brain tissues from hippocampus, were solubilized as described [20], followed by immunoprecipitation with an ApoE antibody. The immunoprecipitated proteins were separated in a 10% SDS-PAGE gel, transferred to a PVDF membrane, and immunoblotted with a Mot-Ab. Quantitation of the mortalin-apoE bands indicates that the binding is genotype- and disease-dependent (c). The complementary experiment, in which mortalin is immunoprecipitated with the Mot-Ab, and the IP is immunoblotted against ApoE-Ab (d,e) shows almost identical results. Proteins were identified by MALDI-TOF/TOF mass spectrometry (white arrow). “M” indicates proteins that were immunoprecipitated with the mortalin antibody and identified by mass spectrometry; correspondingly, “E” represents proteins that were immunoprecipitated with the apoE antibody.The level of oxidized proteins in a cell reflects the balance between the rates of protein oxidation (generation of ROS) and protein degradation (degradation of oxidatively-damaged proteins) [76,102]. Some studies have shown that there is an association between aging and oxidative damage of stress chaperones [21], like mortalin, in neurodegenerative diseases, including Alzheimer’s disease [11,21,102,103] and Parkinson’s disease [15,104]. Our studies have demonstrated that mortalin is oxidized in the brain tissues of an animal model of Alzheimer’s disease [21]. Another potential role of mortalin in neurodegeneration stems from the participation of mortalin in calcium channel regulation [58], a critical process for neuronal health.Apolipoprotein E (ApoE) is important in the regulation of cholesterol and metabolism of triglycerides. There are three common ApoE isoforms: ε2, ε3 and ε4. The APOE4 allele is associated with an increased risk of Alzheimer’s disease [105,106,107,108]. Studies of ApoE4 transgenic and ApoE-deficient mice have confirmed an association between reduced ApoE activity, oxidative damage, and age-dependent neuronal alterations. Using proteomics, Osorio et al. performed a study of human ApoE4-Targeted Replacement mice (hApoE4-TR) compared to hApoE3-TR as control [20]. It was found that different mortalin isoforms are present in hApoE4-TR and hApoE3-TR mice brains, as well as between Alzheimer’s disease patients and age- and gender-matched controls [20]. In addition, using immunoprecipitation with ApoE- and with mortalin-antibodies, we have found that mortalin binds ApoE in hApoE-TR mice, as well as in human brains of Alzheimer’s disease patients (Figure 4). This binding, whose functional effect is under investigation, is different between diseased and non-diseased brains, and between APOE ε3/3 and APOE ε4/4 genotypes (Figure 4).OS, and mitochondrial and proteosomal dysfunction have been implicated in the pathogenesis of Parkinson’s disease [15,109,110,111]. Parkinson’s disease is a progressive disorder characterized by dopaminergic neurodegeneration in the Substantia Nigra pars compacta (SNpc) and by the appearance of proteinaceous cytoplasmic inclusions (Lewy bodies) in the remaining nigral neurons [15,112]. A reduced expression level of mortalin has been observed in the affected brain regions of Parkinson’s disease patients [15,113] and in a cellular model of Parkinson’s disease [47]. Specifically, in dopaminergic neurons, manipulations of the level of mortalin resulted in changes to the sensitivity to Parkinson’s disease phenotypes via different pathways related to OS, mitochondrial and proteasomal function [47], correlating with reduced mitochondrial membrane potential and increased production of ROS [45].Like in other neurodegenerative diseases, ROS is a key element in the pathophysiology of Parkinson’s disease [114]. Parkin, an E3 ubiquitin–protein ligase that mediates polyubiquitination of VDAC [115], is associated with mitochondrial dynamics [116], is involved in the regulation of mitochondria morphology, and is related with autosomal-recessive Parkinson’s disease. Parkin may also play a role in sporadic cases of Parkinson’s disease. There is evidence indicating that mortalin and Parkin provide a protective effect against oxidative damage, and that mortalin is involved in Parkinson’s disease-related abnormal mitochondrial morphology. Under OS, mortalin knockdown stimulates disintegration of mitochondrial connectivity and low-grade branching of mitochondria [15].Dj-1 is an oncogene that protects cells against OS and cell death, and mutations in Dj-1 are associated with familial forms of Parkinson’s disease [117]. Dj-1 is associated with chaperones including Hsp70, CHIP and mortalin and undergoes OS-mediated translocation into mitochondria [118]. Mortalin has been identified as one of the five major proteins (mortalin, nucleolin, Grp94, calnexin and clathrin) that bind α-synuclein and Dj-1, two critical proteins in Parkinson’s disease pathogenesis [34,47].Mortalin-null cells exposed to OS show disintegration of mitochondrial connectivity, suggesting that mortalin is implicated in the control of the mitochondrial dynamics and morphology [15]. It has also been reported that Tid-1, a chaperone protein involved in the regulation of cell survival, interacts with mortalin on an isoform-specific basis, and can mediate the reactivation of protein aggregates. It was suggested that mortalin can serve as a scavenger of toxic protein conformers in human mitochondria, making it an attractive target for therapies against protein conformational diseases [14]. Apoptosis allows the elimination of non-viable cells without affecting the neighboring cells. Unlike the rapid turnover of cells in proliferative tissues, neurons show only slight regeneration and normally stay alive for the entire life of the organism [119,120]. OS and metabolic stress are able to activate the chaperone system and can initiate neuronal apoptosis (Figure 5), and under OS there is inhibition of the electron transport chain and production of ROS in neurons [121]. Qu et al. demonstrated that overexpression of mortalin in neuroblastoma cells can reduce OS [30]. Mortalin increases the stress response capacity of the cells resulting in increased cell viability and extended longevity of an organism.Mortalin responds to ROS accumulation under stress conditions while regulating other housekeeping functions, including control of cell proliferation, intracellular trafficking, or anti-apoptotic activity; this down-regulation of housekeeping functions may result in uncontrolled cell proliferation. For example, OS induces mortalin translocation into the mitochondria [104], leaving other proteins, including Apaf-1 and p53, unchecked, thus potentiating the disproportionate activation of apoptotic biochemical cascades.From our observations that the mitochondrial proteome is affected by mortalin expression levels, we would expect that, under normal conditions, mortalin behaves as an anti-apoptotic protein that inactivates p53, resulting in cyt-c or Apaf-1 not being released. On the other hand, in the presence of oxidative stress, mortalin is responsible for mitochondrial homeostasis, allowing cytoplasmic p53 to induce apoptosis. These observations point to mortalin being a sensitive marker of stressed cells and the apoptotic function associated with p53 activity. Mortalin behaves as a regulatory protein that can alter cell function by associating with vital cellular proteins, including p53, Dj-1, FGF-1, and Hsp60. Regulating the functions of these proteins most likely affects signals involved in neurodegenerative diseases and apoptosis as discussed above. Bearing in mind the multiple functions of mortalin in cell control, is not surprising that over-expression of mortalin is able to promote cancer and may trigger features associated with neurodegenerative diseases, including Parkinson’s and Alzheimer’s diseases.A schematic model showing the mechanism of action of mortalin. (a) The normal function of mortalin. Under normal conditions, mortalin associates with certain chaperones, including Dj-1 (1 in the figure); following this interaction, the complex travels to the mitochondria, where mortalin is detached and enters the mitochondria (2). A magnified image (3) shows that mortalin crosses the outer membrane (TOM) (4) and the inner membrane (TIM) where it performs multiple functions, including chaperoning of the precursor proteins into the mitochondrial matrix (5). During, this process, mortalin binds simultaneously to TIM 44 (a peripheral membrane protein) and to the Tim23 complex (5). Next, the mature protein is transferred by mortalin to Hsp60 (6). Under normal conditions, mitochondrial mortalin forms stable complexes with p66Shc and Apaf-1 (7), which are released under cellular stress. (b) Under OS, mortalin levels are increased in the mitochondria, inhibiting ROS accumulation, and acting as a cytoprotective protein while maintaining cell viability (8). The magnified portion of the image shows that under stress conditions however, thelevels of mortalin needed to control other activities are reduced and the cells can suffer an imbalance; for example, the interaction between mortalin and p66Shc can be disrupted, and p66Shc oxidizes cyt-c and promotes the opening of the mitochondrial permeability transition pore triggering apoptosis (9); alternatively, Apaf-1 can induce apoptotic pathways (10). As a consequence, apoptosis increases in the absence of free mortalin, as a result of a rich OS environment.Cellular homeostasis is maintained by a strict regulation of the balance between ROS production, cell growth, and apoptosis. Many pathological states, including cancer and neurological diseases, are often associated with OS and disregulation of apoptotic pathways. Changes in mortalin expression are associated with cellular protection as they permit cells to survive lethal conditions modulating the cell’s lifespan [29,30,31]. Mortalin also has anti-apoptotic and pro-proliferative activities that influence the functions, dynamics, morphology and homeostasis of mitochondria.Proteomic, molecular and biochemical data suggest that cell death, degeneration, and immortalization are not controlled by a single mechanism; they are regulated by a complex networks of proteins interconnected via multiple molecular pathways. The mitochondria and mitochondrial proteins play a fundamental role and are an indispensable part of this regulatory network. Several process that are related to the mitochondria, such as apoptosis, have been extensively studied for many years, but some of the processes, such as protein import, complex assembly, and the molecular mechanisms by which mortalin influences apoptosis have not yet been sufficiently elucidated. This review summarizes the present knowledge on mortalin and its relationship to apoptosis and neurodegenerative diseases, and mortalin’s role in OS and mitochondrial function. Although several cellular proteins are known to interact with mortalin, mortalin appears to be a regulatory protein that maintains the integrity of the cell via multiple molecular processes that are still under investigation. Further research on the function and dynamics of mortalin could provide valuable information about the complex balance between longevity, neurodegeneration, and apoptosis.We would like to thank Carol Parker of the UVic Proteome Centre for careful review of the manuscript. This research was sponsored with funding from the Department of Neurobiology of Duke University (OA), the Systems Proteomics Center of the University of North Carolina at Chapel Hill (OA), and two grants from the CIDI - Universidad Pontificia Bolivariana, Medellin, Colombia (OA).

Med-MDPI/biomolecules/biomolecules-02-01-00165.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).The endoplasmic reticulum (ER) is a major intracellular calcium storage pool and a multifunctional organelle that accomplishes several calcium-dependent functions involved in many homeostatic and signaling mechanisms. Calcium is accumulated in the ER by Sarco/Endoplasmic Reticulum Calcium ATPase (SERCA)-type calcium pumps. SERCA activity can determine ER calcium content available for intra-ER functions and for calcium release into the cytosol, and can shape the spatiotemporal characteristics of calcium signals. SERCA function therefore constitutes an important nodal point in the regulation of cellular calcium homeostasis and signaling, and can exert important effects on cell growth, differentiation and survival. In several cell types such as cells of hematopoietic origin, mammary, gastric and colonic epithelium, SERCA2 and SERCA3-type calcium pumps are simultaneously expressed, and SERCA3 expression levels undergo significant changes during cell differentiation, activation or immortalization. In addition, SERCA3 expression is decreased or lost in several tumor types when compared to the corresponding normal tissue. These observations indicate that ER calcium homeostasis is remodeled during cell differentiation, and may present defects due to decreased SERCA3 expression in tumors. Modulation of the state of differentiation of the ER reflected by SERCA3 expression constitutes an interesting new aspect of cell differentiation and tumor biology.Calcium is actively accumulated into the endoplasmic reticulum (ER) from the cytosol by Sarco/Endoplasmic Reticulum Calcium ATPase (SERCA)-type calcium pumps. By using the energy of ATP hydrolysis, these enzymes, located in the ER membrane, generate a strong calcium ion concentration gradient between the ER lumen (high micromolar [1,2]) and the cytosol (low nanomolar [3]). ER calcium storage is essential for the initiation of calcium-dependent cell activation. The activation of many normal plasma membrane receptors (EGF, FGF, PDGF, chemokine, bioactive peptide receptors, etc.), but also oncogenic mutant receptor activity [4,5] leads, in parallel with the activation of other signaling pathways, to the activation of phospholipase C enzymes and the hydrolytic cleavage of membrane phosphatidylinositol-4,5-bisphosphate into diacylglycerol and inositol-1,4,5-trisphosphate (IP3) [3]. Binding of IP3 to IP3-receptor calcium channels (IP3R) leads to IP3R opening and to calcium release from the ER into the cytosol. ER calcium depletion upon IP3R opening then leads to calcium influx into the cytosol from the extracellular space through Orai-type plasma membrane calcium channels [6], and TRP-type channels, as well as non-capacitative calcium influx can also contribute to calcium entry into the cell. Opening of Orai-type channels is induced by STIM-1, an integral ER membrane protein which is in turn activated by the dissociation of calcium from its ER-luminal calcium binding region when ER calcium decreases during IP3-induced calcium release [7,8]. Calcium release from the ER combined with capacitative calcium influx from the extracellular space leads to markedly increased cytosolic calcium levels and the activation of key calcium-dependent enzymes such as protein kinase-C isoforms, calcineurin, calpains, calmodulin dependent kinases and other calmodulin binding proteins involved in cell activation [9,10].The amplitude of calcium release from the ER depends on the magnitude of the calcium concentration gradient between the ER and the cytosol. Moreover, the variations of intra-ER calcium levels are known to modulate the opening of IP3R calcium channels by IP3 [3,11], and are essential for STIM activation [12,13]. Therefore, the precise regulation of the intra-ER calcium concentration constitutes an important mechanism to adjust the sensitivity of a cell to calcium mobilizing stimuli. Because calcium is accumulated in the ER exclusively by SERCA enzymes, SERCA-dependent calcium transport constitutes a key nodal point in the control of cell activation. In addition, because during cell activation SERCA enzymes rapidly re-accumulate part of the calcium released into the cytosol, SERCA activity exerts an important influence on the amplitude, the shape, as well as the frequency of cellular calcium peaks and oscillations [14,15,16,17,18,19,20] and therefore on cell activation [21,22].In addition, calcium accumulated by SERCA enzymes is required also for intra-ER functions such as chaperoning of newly synthesized proteins transiting through the organelle [23,24,25,26]. Several ER resident chaperones such as calreticulin or calnexin bind and require calcium for activity [23,25,27]. Therefore, defects of ER luminal calcium homeostasis can lead to defects of protein maturation and to the accumulation of misfolded proteins in the ER, that activates various adaptive ER stress responses or, if overwhelming, leads to cell death [28,29,30,31,32].Availability of calcium ions in the ER lumen for (1) second-messenger-induced calcium release; (2) the control of capacitative calcium influx; and (3) intra-ER chaperone activities are all critically dependent on proper SERCA function. SERCA-type enzymes occupy therefore a critical position in cellular calcium homeostasis and signaling, and subtle changes in SERCA expression and activity can have far-reaching consequences for the dynamics of calcium signaling and the behavior or the survival of the cell. The position that SERCA-dependent calcium transport occupies in cell signaling makes it also a promising pharmacological target. For example, SERCA inhibition for the targeted therapy of prostate cancer is currently being evaluated using peptide conjugates of thapsigargin, a highly potent and selective SERCA inhibitor. The peptide conjugates are hydrophilic and therefore remain extracellular and thus inactive when administered intravenously. However, hydrolysis of the conjugate by PSA, a prostate-specific peptidase will lead to the release of free thapsigargin in the vicinity of the cells, its diffusion into the cell, SERCA inhibition and induction of cell death [33,34].Three SERCA genes are known in the human genome (ATP2A1, ATPA2 and ATP2A3), that by alternative splicing can generate several protein isoforms that differ in their C-terminal regions [35,36,37,38,39]. The expression of SERCA isoenzymes is tissue dependent and developmentally regulated. Whereas SERCA1a and 1b are expressed in adult and neonatal fast twitch skeletal muscle, respectively, SERCA2a is expressed in cardiomyocytes, and SERCA2b is abundant in smooth muscle cells. A minor isoform, SERCA2c has also been detected in various tissues [40]. Although abundantly expressed in smooth muscle, SERCA2b has also been detected in almost all non-muscle cell types as well, indicating that SERCA2b is a ubiquitous isoform involved in calcium uptake in the ER in most cells. The third member of the SERCA family, SERCA3 bears approximately 80% homology with other SERCA isoforms [37,41]. The ATP2A3 gene can give rise to six known SERCA3 isoforms that arise by alternative splicing in the 3’ region of the transcripts [39,42,43]. Comparative analysis of the localization and of the biochemical characteristics of various SERCA isoforms revealed significant differences. Transport activity is stimulated by calcium in a concentration-dependent manner [44,45,46]. When the calcium concentration dependency of calcium transport by various SERCA isoforms was compared, it has been shown that the apparent calcium affinity (KCa2+, as defined by the calcium concentration that leads to half-maximal induction of transport) of all SERCA3 isoforms is weaker (approximately 1.2 μM) than that of other isoforms, and in particular of SERCA2b (0.2 μM) [35,47,48,49,50,51]. Based on this observation SERCA2b is thought to be a more “stringent” calcium pump than SERCA3: whereas SERCA2b-dependent calcium sequestration is fully active already above the 0.2 μM cytosolic calcium concentration range, fully active calcium sequestration by SERCA3 would be observed only above 1.2 μM calcium, and SERCA3 would pump calcium very weakly at around 0.2 μM [35,37,39]. A new level of complexity has been discovered when it was shown that in several cell types SERCA2b and SERCA3 enzymes are expressed simultaneously [52,53]. In cells of hematopoietic origin (lymphoid, myeloid, megakaryocytic cells, cell lines, as well as platelets), insulin-secreting pancreatic β-cells, gastric and colonic epithelium, as well as Purkinje neurons, SERCA2 and SERCA3 enzymes can be found in various amounts simultaneously. SERCA3 is expressed also in vascular endothelial cells, and expression levels vary according to the proliferative state and the anatomic location of the cells [54]. Several excellent reviews are available about SERCA structure [42,50,55,56,57,58,59], function [35,37,38,39,57,60,61], knock-out animal models [62,63] and genetic diseases [50,64,65,66,67], as well as about the role of calcium signaling in cancer [68,69]. With the aim of attracting attention to the remodeling of ER calcium homeostasis in cancer, we will briefly summarize here available data on the modulation of the expression of SERCA enzymes in several in vitro models of cancer cell differentiation, and on the patterns of SERCA3 protein expression in various human tumors and corresponding normal tissue in situ.Acute promyelocytic leukemia (APL) is a myeloid malignancy in which cells blocked at the promyelocyte stage of myeloid differentiation accumulate. In most cases leukemic cells carry the t(15;17)(q24;q21) chromosomal translocation that leads to the expression of the PML/RARα (Promyelocytic Leukemia/Retinoic Acid Receptor-alpha) fusion oncoprotein. At physiological (nanomolar) all-trans-retinoic acid concentrations PML/RARα acts as a dominant negative inhibitor of gene expression that, by binding to target gene promoters and the recruitment of nuclear co-repressors inhibits granulocytic differentiation. At higher, pharmacologic concentrations (near micromolar), binding of all-trans retinoic acid to PML/RARα relieves this inhibition, leading to the dissociation of co-repressors, the recruitment of transcriptional co-activators and expression of target genes, followed by the proteolytic degradation of the PML/RARα oncoprotein [70,71,72]. Treatment by all-trans retinoic acid (ATRA) leads to growth arrest and to the terminal neutrophil granulocytic differentiation of APL cells in vitro, as well as in vivo, and constitutes the first example of molecularly targeted therapy of leukemia. When combined with cytotoxic treatments aimed at the elimination of the leukemia initiating cells, ATRA is highly successful for the treatment of APL [73]. When the neutrophil granulocytic differentiation of cell lines or freshly isolated APL cells is induced by ATRA, significant changes of SERCA expression occur [74]. Similarly to all cell lines of hematopoietic origin tested so far, untreated cells express SERCA2, as well as SERCA3. However, during differentiation SERCA3 expression is selectively induced, whereas that of SERCA2 is decreased, or is not modified significantly [74]. The induction of SERCA3 expression could be observed also during cell differentiation induced by an RARα-specific synthetic agonist, and ATRA-induced differentiation, as well as SERCA3 induction was inhibited by an RARα-selective antagonist [74]. Importantly, SERCA3 expression was induced during the differentiation of the cells induced by cAMP analogues as well, and SERCA3 expression was not modified by ATRA in cells refractory to the differentiation-inducing effect of the drug [74]. Taken together, these observations show that the induction of SERCA3 expression is an integral part of the neutrophil granulocytic differentiation program of APL cells. The functional consequences of the modulation of SERCA expression on calcium transport activity were investigated in HL-60 cells. ATRA-induced neutrophil granulocytic differentiation of HL-60 cells leads to increased SERCA3 expression, whereas SERCA2 expression is at the same time decreased [74]. When ATP-dependent 45Ca2+ accumulation into microsomal membrane preparations obtained from control and differentiated HL-60 cells was compared, calcium accumulation into the SERCA3-associated compartment was markedly increased, whereas calcium accumulation into the SERCA2-associated pool was decreased [74]. This indicates that the modulation of SERCA expression leads to the functional remodeling of calcium transport and a shift of calcium uptake into a SERCA3-associated storage pool.Induction of SERCA3 expression could also be observed during the megakaryocytic differentiation of various human erythro-megakaryoblastic leukemia cell lines induced by protein kinase C activating phorbol esters [75]. Platelets, that correspond to the ultimate stage of megakaryocyte differentiation contain very high amount of SERCA3 protein [53,76]. Induction of SERCA3 expression during in vitro differentiation of megakaryocytic cell lines, expression of SERCA3 in mature normal human megakaryocytes and circulating platelets indicate that induction of SERCA3 expression is part of the differentiation program of this lineage. SERCA2 and SERCA3 signal intensities on Western blots with isoform-specific [52,74], as well as pan-SERCA antibodies [53] lay within the same order of magnitude, and 32P-labeled phosphoenzyme levels for SERCA2 and SERCA3 are also roughly comparable in platelet membranes [53]. In addition, 45Ca2+-transport measurements on platelet- as well as HL-60 cell-derived microsomal membrane preparations suggest that SERCA2 and SERCA3 contribute to ER calcium uptake to comparable extents [74,77]. These observations indicate that the contribution of SERCA2 and SERCA3 to total SERCA function lies within the same order of magnitude. In other words, SERCA3 is not a minor or marginally expressed isoform when compared to SERCA2 in differentiated cells.Normal colonic epithelium is a rapidly renewing tissue in which asymmetric division of epithelial stem cells located in the region of the crypt base is followed by the proliferation and the differentiation of upward migrating epithelial cells, which thereafter undergo apoptosis in the surface epithelium [78]. Tumorigenesis in the colonic epithelium is regarded as a multistep process whereby the accumulation of mutations that inactivate tumor suppressor genes and activate oncogenes leads to the stepwise acquisition of neoplastic phenotypes of increasing malignant potential [79,80]. This is best illustrated by the adenoma to adenocarcinoma sequence: mutations in the APC/β-catenin/TCF4 pathway induce the formation of low grade benign tumors (adenomas) that upon the acquisition of further mutations (k-Ras, SMAD-4, p53 and others) increase in grade and then become malignant (adenocarcinomas) [79,81]. Premalignant, as well as malignant lesions in the colon can be graded according to histological differentiation, and the low grade to high grade adenoma to in situ and invasive well/moderately/poorly differentiated adenocarcinoma sequence corresponds to the sequential loss of phenotypic differentiation and increased malignant potential. Small lesions called hyperplastic polyps of Morson, which, in contrast to adenomas, are devoid of significant potential to develop into carcinoma can also arise in the colon [82,83,84].When SERCA3 expression is investigated in the colon by immunohistochemistry, strong SERCA3 expression can be observed in the epithelial cells, and staining increases from the region of the crypt base where colonic epithelial stem cells are located towards the surface epithelium [85]. This indicates that SERCA3 is abundantly expressed in the differentiating normal colonic epithelium, and a similar SERCA3 staining pattern can be observed also in hyperplastic polyps. On the other hand, SERCA3 expression is progressively decreased along the adenoma/adenocarcinoma sequence: in contrast to normal epithelium that strongly expresses SERCA3, staining is globally decreased and heterogeneous in adenomas, with a more marked decrease observed in high grade lesions, is very low in well differentiated adenocarcinomas, and barely detectable or absent in moderately and poorly differentiated carcinomas.Colon carcinoma cell lines can be induced to undergo differentiation in vitro by treatment with short chain fatty acid-type histone deacetylase inhibitors such as butyrate or valerate, butyrate releasing prodrugs or ω-aryl-substituted short chain fatty acid analogues such as phenylbutyrate [86]. Short chain fatty acid-induced differentiation is physiologically relevant. Short chain fatty acids present in the colon lumen due to the fermentation of dietary fibers by the colonic flora are thought to induce differentiation of the normal epithelium and of microscopic precancerous lesions thereby contributing to the cancer-preventive effects of dietary fiber consumption [87,88]. In addition, the Caco-2 colon adenocarcinoma cell line spontaneously undergoes differentiation when cultured in post-confluent conditions. This can be detected by morphological (formation of a polarised epithelial monolayer with brush border membrane and tight junctions), functional (transcellular solute transport, transepithelial electric resistance), as well as biochemical criteria (induction of the expression of markers such as dipeptidyl peptidase 4, carcinoembryonic antigen, sucrase-isomaltase or the isoform switch of the ZO-1 tight junction protein) [89]. Induction of colon and gastric carcinoma cell lines by various differentiation-inducing treatments including short chain fatty acids and analogues, as well as the spontaneous differentiation of Caco-2 cells is associated with the selective induction of the expression of SERCA3 protein [86]. In addition, the inhibition of the APC/β-catenin/TCF4 pathway in colon cancer cells by the forced expression of a transfected, dominant negative TCF4 variant also leads to increased SERCA3 expression [85].The effect of butyrate treatment on cellular calcium homeostasis was investigated in the Kato-III gastric carcinoma cell line, in which treatment leads to a marked induction of SERCA3 expression, whereas SERCA2 levels are at the same time decreased. As shown by Fura-2 calcium fluorimetry, butyrate treatment is associated with increased resting cytosolic calcium levels and decreased thapsigargin-sensitive intra-ER calcium storage [86].Taken together, these observations indicate that SERCA3 expression is lost during the multi-step process of colon carcinogenesis, that decreased SERCA3 expression is an early marker of colon tumorigenesis, and that SERCA3 expression is induced during colon and gastric cancer cell differentiation, a process during which the calcium homeostasis of the cell is modified.Breast tumorigenesis is a rather complex process in which several parallel molecular oncogenic mechanisms operate in a somewhat combinatorial manner [90,91,92,93]. This leads to the formation of several types of preneoplastic lesions with various types and degrees of dysplasia, and of various cancer types such as ductal and lobular carcinoma [94]. Most breast carcinomas are thought to arise in the terminal ductal lobular units that consist of acinar secretory cells, myoepithelial cells and the cells of the intralobular duct [94]. The classification of breast neoplasia can be performed based on histological, immunophenotypic and hormonal criteria, as well as by the detection of genetic mutations associated with different tumor types. Classification is pertinent for prognosis and response to various types of treatment [92]. In addition, whole genome and transcriptome analyses are currently used for the molecular classification of breast cancer [90,93,95]. However, because of the interconnectedness of several, not sufficiently known oncogenic mechanisms [92], different classification methods are not always concordant, and the behaviour and the response to treatment of individual tumors assigned to the same currently used categories may differ significantly.In order to investigate the role of endoplasmic reticulum calcium biology in breast tumorigenesis, SERCA3 expression was investigated by immunohistochemistry in normal breast, in various preneoplastic lesions and in invasive ductal and lobular breast carcinoma [96]. Whereas normal breast acinar epithelial cells displayed a strong SERCA3 staining, SERCA3 expression was markedly decreased already in very early benign lesions such as adenosis and lobular hyperplasia without atypia, and remained low in lobular carcinoma. This indicates that SERCA3 expression becomes anomalous already at the earliest morphologically detectable stages of lobular dysplasia and remains thereafter low at further stages of lobular tumorigenesis [96] (Figure 1).In invasive ductal carcinomas SERCA3 expression was globally decreased when compared to normal ducts, and, although variable, was inversely correlated with the Elston-Ellis grade, with the loss of steroid hormone receptor expression, as well as with triple negative (estrogen-, progesterone-receptor and HER-2 negative) status. These observations, combined with the analysis of tumor groups stratified simultaneously for markers such as hormone receptor expression and proliferative index or nuclear grade, showed that SERCA3 expression is inversely correlated with tumor differentiation and the degree of aggressiveness/malignancy of ductal carcinoma of the breast [96].SERCA3 expression in normal breast acini and in invasive lobular breast carcinoma. SERCA3 expression was detected by immunohistochemistry with the avidin-biotin-peroxydase method and 3,3’-diaminobenzidine chromogenic substrate. In normal breast (A) strong SERCA3 staining (brown) is observed in the acinar cells of lobules (lower left), and staining of normal ducts is weaker (upper right). When compared to normal acini, SERCA3 expression is markedly decreased in invasive lobular carcinoma (B). Tissue was counterstained with hematoxylin (blue).Calcium signaling plays an important role in T cell activation. Activation of the T-cell receptor complex leads to the hydrolysis of membrane phosphatidyl-inositol-4,5-bisphosphate by phospholipase Cγ into diacylglycerol (DAG) and inositol-1,4,5-trisphosphate (IP3). This leads to calcium mobilization from the ER, protein kinase C and calcineurin activation, and the activation of NF-κB and NF-AT-type transcription factors that orchestrate T cell activation, the acquisition of a blastic phenotype and lead to intense IL-2-dependent cell proliferation [97].The effects of IP3 and DAG can be mimicked, respectively, by a calcium ionophore (ionomycin) and a phorbol ester (PMA) in vitro. Treatment of the Jurkat (clone E6-1) human T cell line, a widely used model of T lymphocyte activation, by PMA and ionomycin leads to cell activation, as detected by the induction of the expression of the α chain of the IL-2 receptor and IL-2 secretion. When Jurkat E6-1 cells are treated with PMA plus ionomycin, cell activation is accompanied by a strong and selective down-regulation of SERCA3 expression, whereas SERCA2 levels are, at the same time, slightly increased [98]. Interestingly, SERCA3 down-regulation, as well as IL-2 secretion could be induced only by a combined treatment by PMA and ionomycin, whereas treatments by either drug alone were without effect [98]. In addition, T cell activation, as well as the down-regulation of SERCA3 could be inhibited by cyclosporine-A or FK-506 (tacrolimus) [98], clinically used immunosuppressive drugs [99] that by inhibiting calcineurin-induced NF-AT dephosphorylation block T cell activation.Epstein-Barr virus (EBV), a human gammaherpesvirus can immortalize human B lymphocytes by establishing a state of latent infection in which the virus is transmitted during mitosis to daughter cells in an episomal form [100,101]. During immortalization, resting B lymphocytes acquire a proliferating, activated lymphoblastic phenotype induced by the expression of a limited set of viral genes including EBNA2 (Epstein-Barr virus nuclear antigen-2) and LMP-1 (latent membrane protein-1). EBV-induced immortalization is involved in the formation of several lymphoid malignancies including a large fraction of Burkitt’s lymphomas, Hodgkin lymphoma, T/NK lymphomas, lymphomas of immunocompromised individuals, as well as of nasopharyngeal carcinoma and a subset of gastric carcinoma [102,103]. EBNA2, a major viral transactivator and activator of the Notch transcriptional regulatory pathway modulates the expression of several cellular, as well as viral genes including LMP-1. Expression of LMP-1 (considered as a truncated, constitutively active viral analogue of receptors belonging to the TNFα receptor family, that functionally resembles CD40, a key receptor in normal B cell activation) leads to the activation of several signaling pathways leading to NF-κB, AP-1, MAPK and Akt activation [104,105,106,107,108]. The reprogramming of signaling pathways involved in the control of survival and of the state of activation of resting B cells by EBV infection leads to the emergence of autonomously proliferating immortalized lymphoblastoid cell lines.The investigation of EBV-related effects on the cellular level is greatly facilitated by the availability of pairs of EBV-negative and corresponding latently EBV-infected cell lines. EBV-negative Burkitt’s lymphoma cell lines were infected by EBV in vitro, and latently infected cell lines were established [109,110]. Compared to the parental EBV-negative cells, latent EBV infection of the cells leads to significantly decreased SERCA3 expression, whereas SERCA2 levels are at the same time increased. Importantly, the modulation of SERCA expression by EBV was observed only in cell lines infected with a fully immortalizing EBV strain (B95.8 virus), whereas infection with the non-immortalizing P3HR-1 virus strain (in which LMP-1 expression is deficient due to a deletion in the EBNA2 sequence and consequent loss of trans-activation of LMP-1 expression by EBNA2 [111]), SERCA expression was not modified [112]. Investigation of the effect of individual viral proteins on SERCA expression using inducible expression vectors stably transfected into EBV-negative cells has shown, that whereas EBNA2 expression was without effect, SERCA3 expression was selectively down-regulated in cells expressing LMP-1 in the absence of any other EBV product [112]. In latently infected cells, as well as upon LMP-1 expression, increased calcium storage in a thapsigargin-sensitive, presumably SERCA2-associated intracellular calcium pool was observed [112].SERCA3 expression was also investigated by immunohistochemistry in normal lymph nodes. A strong labeling was obtained in the mantle zone of lymphoid follicles where resting B lymphocytes reside, whereas in germinal centers where activated and proliferating centroblast and centrocytes are located, SERCA3 staining was considerably weaker [112]. Observations on the effect of EBV infection and normal B lymphocyte activation taken together indicate that the down-regulation of SERCA3 expression induced by LMP-1 during EBV-induced immortalization mimics a phenomenon taking place during antigen-dependent activation of normal B lymphocytes in germinal centers. In addition, because SERCA3 down-regulation occurs also during the activation of T lymphocytes as shown in Jurkat cells, it is tempting to propose that the selective down-regulation of SERCA3 expression is a general phenomenon involved in lymphocyte activation in the T, as well as the B lineage. Several lines of evidence show that SERCA3 expression undergoes significant quantitative modifications when the state of differentiation or activation of various cell types changes. In several independent model systems of differentiation such as retinoic acid-induced differentiation of acute promyelocytic leukemia cells, phorbol ester-induced differentiation of megakaryoblastic cell lines, or short chain fatty acid-induced, as well as spontaneous differentiation of colon carcinoma cells, differentiation, detected by a multitude of established markers, is accompanied by a marked induction of SERCA3 protein expression, whereas the expression of the simultaneously expressed SERCA2 isoenzyme is much less modified, or is in fact often decreased. In addition, fully differentiated normal cells that correspond to the final step of these differentiation programs (such as platelets or normal colonic surface epithelial cells) express SERCA3 abundantly. Moreover, when investigated in neoplastic tissue in situ, a loss of SERCA3 expression is observed when compared to the corresponding normal, differentiated cell type, and the loss of SERCA3 expression is proportional to the degree of histologically observable loss of cell differentiation. This phenomenon has been observed when benign, precancerous and malignant lesions were studied comparatively in the colonic epithelium, as well as in breast epithelial lesions of various degrees of dysplasia or malignancy. Importantly, SERCA3 expression has been shown to decrease already at very early steps of dysplasia in colon adenomas, as well as in lobular breast lesions, and remains low, or becomes undetectable at more advanced stages of tumorigenesis and malignant transformation. Down-regulation of SERCA3 expression could also be observed during the acquisition of an activated phenotype during T, as well as B lymphocyte activation and B cell immortalization, processes associated with proliferation and the acquisition of a blastic phenotype.These observations show that when a cell undergoes phenotypic changes such as differentiation, activation or transformation, intracellular calcium sequestration by SERCA-dependent calcium pumping is modified in several cell types, and SERCA3 expression is a pertinent phenotypic marker of this process.The modification of the SERCA2 to SERCA3 molar ratio can have significant functional consequences on ER calcium homeostasis, handling and availability for signaling functions. The calcium concentration dependence of calcium transport (as defined by the apparent calcium affinity of transport, KCa2+) of SERCA2b (KCa2+ ≈ 0.2 μM) and SERCA3 (KCa2+ ≈ 1.2 μM) is distinct. As pointed out earlier [37,39], this calcium concentration range corresponds approximately to the concentration range in which cytosolic calcium levels vary between the resting and activated state. In a very simplified manner this means that whereas SERCA2b-dependent calcium sequestration in the ER is almost fully active already at resting cytosolic calcium levels, SERCA3-dependent calcium sequestration becomes active at cytosolic calcium levels encountered only during calcium-dependent cell activation. Therefore, whereas SERCA2b-dependent calcium accumulation into the ER would be expected to be constitutively active, SERCA3-dependent calcium transport may become important only during increased cytosolic calcium levels observed during activation, and thus SERCA3 would only blunt higher cytosolic calcium peaks, and would become active only at a later phase of the calcium peak. SERCA3 may also be associated with ER regions around which cytosolic calcium levels can reach significantly higher concentrations locally, such as regions in the immediate proximity of open calcium channels [113,114,115]. Interestingly, earlier work on IP3-induced release of calcium accumulated in platelet microsomal vesicle preparations in the presence of the PLIM430 SERCA3-specific inhibitory antibody [116,117] had shown that calcium accumulation into the IP3-mobilizable sub-compartment of platelet intracellular calcium stores is performed preferentially by SERCA3 [77]. The association of SERCA3, a lower affinity calcium pump with an ER sub-compartment involved in second messenger-induced calcium release probably permits the cell to mount larger second-messenger-induced calcium release responses upon calcium release, which otherwise would be blunted by SERCA2b. It can also be hypothesized that by limiting futile release/reuptake cycles, the association of SERCA3 with IP3-sensible calcium pools constitutes an energy-efficient mechanism that permits larger calcium release signals before re-sequestration is initiated. This notion is compatible with the observed association of SERCA3 expression with various differentiated cell phenotypes: one may hypothesize that the association of an IP3-sensitive ER sub-compartment with a less “stringent” calcium uptake mechanism is typical of differentiated cell types that respond to various external stimuli by calcium mobilization, whereas ER calcium homeostasis in undifferentiated cells behaves more autonomously. Interestingly, the working hypothesis of SERCA3 being associated with intracellular calcium pools specialized in signaling is compatible also with the observed down-regulation of SERCA3 expression during lymphocyte activation, a process during which cellular calcium signaling is chronically activated [118]. One may hypothesize that SERCA3 down-regulation in this configuration leads to the constitutive depletion of an IP3-sensitive intracellular calcium pool coupled to a chronically activated state of store-operated calcium influx mechanism and sustained calcium-induced activation.If the distribution of SERCA2b and SERCA3 is heterogeneous within the ER, this may have other interesting consequences for intra-luminal calcium homeostasis too. It can be hypothesized that the association of high and low calcium affinity SERCA pumps such as SERCA2b and SERCA3, respectively, with distinct sub-compartments of the contiguous ER membrane network, and consequent differential calcium uptake in these sub-compartments may lead to the formation of intra-luminal longitudinal calcium gradients and calcium ion migration, even in a resting cell. Such gradients and vectorial calcium fluxes, for example from a SERCA2b-associated region towards a SERCA3-associated one may contribute to the organization of structurally and functionally distinct intra-ER spaces.The modulation of SERCA expression is not a simple passive consequence of cell differentiation. Complete SERCA inhibition induces cell death due to ER stress responses. On the other hand, the partial inhibition of SERCA-dependent calcium sequestration by highly specific inhibitors such as thapsigargin, a sesquiterpene lactone that inhibits various SERCA isoenzymes with high affinity (below nanomolar), cyclopiazonic acid or 2,5-di-tert-butyl-1,4-benzohydroquinone has been shown to enhance or potentiate ATRA-induced differentiation of acute promyelocytic leukemia cells [119], and combined treatments with SERCA inhibitors and ATRA have been shown to induce cell differentiation in several cell lines that are resistant to differentiation induction by ATRA alone [119]. In addition, SERCA inhibition enhances the expression of carcinoembryonic antigen (CEA), a differentiation marker, during post-confluent differentiation of Caco-2 colon carcinoma cells [85], confers cytokine independency to TF-1 erythroleukemia cells [120], and induces HIV expression in latently infected T cells [121]. Moreover, chronic SERCA inhibition in vivo displays tumor-promoting activity. Although in most of these settings it is not possible to clearly assign the observed effect specifically to a given SERCA isoform, these observations show that SERCA function and mechanisms that control several types of cell activation and differentiation are functionally interconnected, and a cross-talk exists between the control of ER calcium sequestration and the regulation of cell differentiation in several cell types. Changes of ER calcium pumping may therefore exert important effects on cell activation and differentiation.Cellular calcium homeostasis is maintained by the concerted action of many calcium handling proteins in the cell leading to a steady state with very different calcium levels in various cellular compartments. Calcium pumping and release occur simultaneously in a cell. Therefore, the calcium concentration, as well as its changes are determined by the concerted action of the entire set of the calcium homeostatic “toolkit” (pumps, channels, calcium binding proteins and their regulators) present in the cell [122], and individual components of this toolkit can display partial functional redundancy. Importantly, the activity of calcium pumps and channels is critically regulated by calcium itself. Several negative, as well as positive feedback mechanisms, cumulative effects and delayed regulations have been described in this context that are modulated by calcium [11,123,124,125,126]. An in-depth understanding of the functional involvement of the remodeling of ER calcium sequestration due to the modulation of SERCA expression will require a more profound understanding of the complex interactions of calcium handling proteins and of the dynamic behavior of this signaling matrix. The consequences of the modulation of the expression and activity of various SERCA isoforms will depend on the given cell signaling context into which these are integrated in a cell. Although heterozygous knock-out of the SERCA2 gene leads to squamous tumorigenesis in mice with long incubation times [62], the corresponding human condition, Darier disease [127,128] does not appear to predispose to tumor formation, and SERCA3 knock-out mice don’t display a neoplastic phenotype. On the other hand, SERCA inhibitors such as thapsigargin [129] or 2,5-di-tert-butyl-1,4-benzohydroquinone [130] are known tumor promoters in vivo, and mutations in SERCA2, as well as SERCA3 sequences have been found in several human tumor types [131,132,133]. These observations, when combined with data on SERCA expression in cancers, indicate that ER calcium homeostasis is involved in the establishment of several types of the malignant phenotype. Phenotypic dedifferentiation is a hallmark of cancer, and, as shown in several tumor types, the loss of SERCA3 expression is part of this process. The accumulated data, when taken together, suggest that the loss of SERCA3 expression reflects the loss of a signaling function or ER sub-compartment present in differentiated cells. Observations made using SERCA inhibitors have shown that partial down-regulation of ER calcium sequestration may lead to differentiation, or enhance the differentiation-inducing effect of other stimuli [85,119]. Thus, it may be hypothesized, that the expression of SERCA3, a low calcium-affinity pump isoform constitutes a physiological mechanism, by which the cell, in analogy with pharmacological SERCA inhibition, renders the corresponding intracellular calcium pool poised to release more calcium into the cytosol, induce stronger capacitative calcium influx, and therefore activate calcium-dependent effector mechanisms involved in cell differentiation more efficiently.The detailed understanding of the mechanisms that connect ER calcium signaling to tumorigenesis requires further work. Computer modeling and systems biology-type approaches applied to experimental data will be undoubtedly very informative in this context [15,16,134,135]. Finally, it is interesting to note that the remodeling of cellular calcium homeostasis by the selective modulation of the expression of specific calcium transporter isoforms may not be limited only to SERCA3 and the ER. Indeed, colon cancer cell differentiation has recently been shown to lead to the selective induction of the expression of the PMCA4b plasma-membrane-type calcium pump isoenzyme as well [136,137]. By transporting calcium ions into the extracellular space through the plasma membrane, PMCA-type calcium pumps decrease cytosolic calcium levels and thus contribute to the control of cell activation. The modulation of PMCA expression during differentiation indicates that the remodeling of cellular calcium homeostasis during differentiation, as well as its defects in cancer may in fact involve an entire, yet unknown set of specific components of the calcium homeostatic toolkit.Accumulating evidence shows that the remodeling of ER calcium pump expression is part of the differentiation program of several cell types. Differentiation is associated with the selective induction of the expression of SERCA3, a lower calcium affinity calcium pump, which is more permissive for second-messenger-induced calcium release than the simultaneously expressed SERCA2b isoenzyme. The modulation of the expression of SERCA isoenzymes constitutes a new mechanism to fine tune ER calcium uptake according to cell phenotype, function and signaling requirements, and may be involved in the structural organization of the organelle. SERCA3 expression is selectively decreased or lost in many tumors, and this probably reflects the loss of a calcium-dependent function characteristic of fully differentiated normal cells. SERCA3 loss is proportional with histological atypia, and can be observed already in premalignant lesions, indicating that ER calcium homeostasis becomes abnormal already at early steps of the process of tumorigenesis. Anomalies of the cross-talk between SERCA function and the control of cell differentiation constitutes a previously unknown aspect of tumor biology that is potentially amenable to pharmacologic intervention, for example by targeted delivery of SERCA inhibitors to tumors.Research in the authors’ laboratories was supported by Inserm, the Association pour la Recherche sur le Cancer, Fondation de France, the Fondation pour la Recherche Médicale, the Ligue Nationale pour la Recherche contre le Cancer, the Agence Nationale de Recherche sur le Sida, the Association Laurette Fugain and Aprifel, France, by OTKA, NKTH/KPI, Hungary and by the Ministère des Affaires Etrangères, France. The support of Balázs Sarkadi, Sylviane Lévy-Tolédano, Randall A Byrn, Jerôme A Groopman, Frank Wuytack, Irène Joab, Remi Fagard, Ágnes Enyedi, Katalin Pászty, Jacqueline Mikol, and Françoise Gray is gratefully acknowledged. We are especially indebted to Neville Crawford for the PLIM430 hybridoma, and we thank Patrice Castagnet for excellent technical help. This work is dedicated to the memory of Andreï Tarkovski.

Med-MDPI/biomolecules/biomolecules-02-02-00187.txt

@@ -0,0 +1,9 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).The cell membrane is a highly selective barrier. This limits the cellular uptake of molecules including DNA, oligonucleotides, peptides and proteins used as therapeutic agents. Different approaches have been employed to increase the membrane permeability and intracellular delivery of these therapeutic molecules. One such approach is the use of Cell Penetrating Peptides (CPPs). CPPs represent a new and innovative concept, which bypasses the problem of bioavailability of drugs. The success of CPPs lies in their ability to unlock intracellular and even intranuclear targets for the delivery of agents ranging from peptides to antibodies and drug-loaded nanoparticles. This review highlights the development of cell penetrating peptides for cell-specific delivery strategies involving biomolecules that can be triggered spatially and temporally within a cell transport pathway by change in physiological conditions. The review also discusses conjugations of therapeutic agents to CPPs for enhanced intracellular delivery and bioavailability that are at the clinical stage of development.Cell penetrating peptides (CPPs) are short peptide sequences that are able to transport molecules across the cell membrane. The CPPs, also known as protein transduction domains (PTDs) are made up of three to 30 peptide residues [1]. They are employed to enhance extracellular and intracellular internalization of various biomolecules including plasmid DNA, siRNA, oligonucleotide, peptide-nucleic acid (PNA), peptides, proteins and liposomes [2].The notion of protein transduction domains (PTD) was proposed based on the observation that some proteins, mainly transcription factors, could shuttle within the cell and from one cell to another [3]. The first observation was made in 1988 when it was shown that the transcription transactivating (Tat) protein of HIV-1 could enter cells and translocate into the nucleus. In 1991, Drosophila antennapedia homeodomain was illustrated to be internalized by neuronal cells. This was the origin of the discovery in 1994 of the first PTD or CPP, a 16-merpeptide derived from the third helix of the homeodomain of Antennapedia termed Penetratin (RQIKIYFQNRRMKWKK) [4]. In 1997, the first non-covalent CPP for delivery of nucleic acids MPG was designed which was closely followed by development of Pep-1 for non-covalent cellular delivery of proteins and peptides. This was followed by the identification of the minimal peptide sequence of TAT (47YGRKKRRQRRR57) required for cellular uptake in 1998 [4].Proofs-of-concept of the CPPs in vivo application for the delivery of small peptides and large proteins and for delivery of PNAs using the chimeric peptide Transportan, derived form the N-terminal fragment of the neuropeptide galanin, linked to mastoparan, a wasp venom peptide was a great breakthrough. Other CPPs that could trigger the movement of a therapeutic agent across the cell membrane into the cytoplasm are continually being designed [5,6].There is no unified classification of CPPs, however they could be divided into two broad classes.Based on the interaction between cell penetrating peptides and the therapeutic agent two main sub-classes of CPPs can be distinguished. The first sub-class includes CPPs that are chemically linked to the therapeutic agents while the second sub-class encompasses those that form stable non-covalent complexes with the therapeutic agents [4].Apart from these two main sub classes in the recent past, another method for non-covalent coupling of CPPs for a nucleic acid cargo (i.e., antisense oligomer, siRNA) has been evidenced. This entails sequence complementary hybridization, for instance the binding of the oligonucleotide NFkappaB decoy to the nonamer PNA sequence that resulted in a stable complex that was efficiently translocated across the plasma membrane [7].This sub-class of CPPs forms a covalent conjugate with the therapeutic agent by chemical cross-linking or by cloning followed by expression of a CPP fusion protein [8]. Such interactions have been realized in several CPPs including peptides derived from Tat, penetratin, polyarginine peptide Arg8 sequence, Transportan, VP22 protein from Herpes Simplex Virus (HSV), antimicrobial peptides Buforin I and SynB as well as polyproline sweet arrow peptide [9,10,11]. This is further ascertained from the report of Schwarze and group confirming that several proteins do traverse biological membranes through protein transduction. The small sections of these proteins (10–16 residues long) were noted to be responsible for linking these domains covalently to the compounds, peptides, antisense peptide nucleic acids or 40-nm iron beads, or as in-frame fusions with full-length proteins [12].A variant of this sub-class has been reported in the recent past; new generations of CPPs which combine different transduction motifs or transduction domains in tandem with protein or oligonucleotide-binding domains [13]. Different explanations have been proposed for this stable but cleavable conjugates that involve mainly disulfide or thio-esters linkages [14]. These are CPPs that form non-covalent complexes with biomolecules thereby improving their delivery into cells. They occur mostly as short amphipathic peptide carriers consisting of a hydrophilic (polar) domain and a hydrophobic (non-polar) domain. The amphipathic character arises from either the primary structure or the secondary structure [5]. Primary amphipathic peptides have a sequential assembly of hydrophobic and hydrophilic residues. Secondary amphipathic peptides are formed by the conformational state that allows the positioning of hydrophobic and hydrophilic residues on opposite sides of the molecule [15]. Pep-1 and MPG are primary amphipathic peptides, which form stable complexes with oligonucleotide or protein/peptide through non-covalent electrostatic and hydrophobic interactions [16].Several studies have reported this kind of non-covalent interaction between CPPs and the therapeutic agent. Amphipathic peptide, Pep-1, was used for the delivery of small peptides and proteins while MPG has been shown to efficiently deliver siRNA into cultured cell lines [17].This classification of CPPs is based on their ionic properties. In regard to the ionic charge induced by the CPP, this class could be subdivided into cationic or amphipathic CPPs.The cationic CPPs essentially contain clusters of polyarginine in their primary sequence. TAT, the basic domain of HIV-1 TAT protein is a typical example of the cationic CPPs. It contains clusters of arginine and lysine residues. Other peptides derived from protein transduction domains include the Penetratin, and VP22 HSV-1 structural protein [18,19].These are the MAP and Transportan groups of cell penetrating peptides. The amphipathic class comprises peptides with a high degree of amphipathicity. The high degree of amphipathicity is usually contributed by lysine residues when present as a building block. Other examples include the MPG, Pep-1, MAP，SAP Proline-rich motifs and PPTG1 as a result of hydrophiliity / hydrophobicity of the CPP’s building blocks [20].Therapeutic agents are usually delivered intracellularly to exert their therapeutic action inside the cytoplasm or individual organelles such as the nuclei for gene therapy, delivery of deficient lysosomal enzymes in lysosomes for disease therapy, and proapoptotic anticancer drugs in mitochondria for cancer therapy. The cell membrane prevents proteins, peptides, and drug carriers from entering cells unless an active transport mechanism is involved [21]. Cell penetrating peptides therefore are used to promote the delivery of associated drugs and drug carriers into cells. Despite the fact that the exact mechanisms underlying the internalization of CPPs across the cellular membrane are not fully understood, a large number of different therapeutic agents have been efficiently delivered by CPPs. These range from small molecules to proteins and even liposomes and magnetic particles [11]. Typical applications of CPPs in the delivery of biopharmaceutical agents include:This involves delivering therapeutic genes into the nucleus of target cells to achieve expression of a deficient or incorrectly expressed gene product. Difficulties in developing safe and efficient gene delivery vectors that could sustain gene expression for long periods, poor permeability of the plasma membrane of eukaryotic cells to DNA results in low concentration of DNA and other oligonucleotides at their targets [22]. To overcome this, peptide carriers such as polylysine and polyarginine that have membrane-destabilizing properties and could bind with DNA mainly through electrostatic interaction have been developed to facilitate gene transfer into cultured cells and cells in living organisms [23]. For example amphipathic peptides with pH-dependent fusogenic and endosomolytic activities such as the fusion peptide of HA2 subunit of influenza hemaglutinin, or synthetic analogs GALA, KALA, JTS1, and histidine-rich peptides have been shown to increase transfection efficiency when associated with poly-L-lysine/DNA, condensing peptide/DNA, cationic lipids, poly-ethyleneimine or polyamidoamine cascade polymers [15].The ability to condense DNA and to favor endosomal escape by CPPs like PPTG1, and the prevention of endosomal uptake by MPG has led to their use for gene delivery in cultured cells. PPTG1 has been reported as an important component for in in-vivo gene expression following intravenous injection [24].Cationic liposomes, nanoparticles, cationic polymers, and CPPs have also been used as non-viral vectors in preference to the viral vectors. This is because viral vectors, despite their sustained levels of transduction and in some cases efficient and stable integration of exogenous DNA into a wide range of host genomes [25], have been noted to have some disadvantages such as immunogenicity, toxicity, difficulty of large-scale production, hence the preference of non-viral vectors for plasmid DNA delivery [22].CPPs easily conjugate covalently or non-covalently with siRNAs. The siRNAs covalently linked to Transportan and Penetratin have been associated with a silencing response. Non-covalent complexes or aggregates formed with siRNA usually have a net positive charge when there is a surplus of positive charges over negative charges. The covalent linkage of CPPs to siRNAs results in the formation of small, monomeric CPP/siRNA conjugates of known stoichiometry with high reproducibility [26]. CPPs could therefore be used for delivery of siRNAs either by covalent or non-covalent approaches [27]. However non-covalent strategies are more efficient for siRNA delivery, for example MPG peptide has been extensively reported to improve siRNA delivery into a large panel of cell lines including adherent cell lines, cells in suspension, cancer and challenging primary cell lines biological response [28,29]. It has also been applied for in vivo delivery of siRNA targeting OCT-4 into mouse blastocytes. It has equally been used for the delivery of siRNA targeting an essential cell cycle protein, cyclin B1. Intravenous injection of MPG/cyclin B1 siRNA particles has been shown to efficiently block tumor growth [30].TAT peptide associated with an RNA-binding motif has been reported to block in vivo epidermal growth factor (EGF) factor. The CPP complex, cholesterol-Arg9 was also shown to enhance siRNA delivery in vivo against vascular endothelial growth factors. A small peptide derived from rabies virus glycoprotein linked to polyarginine R9 has been reported recently to deliver siRNA in the CNS [31].A new cell-penetrating peptide, PepFect14 (PF14), which efficiently delivers splice-correcting oligonucleotides (SCOs) to different cell models including HeLa pLuc705 and mdx mouse myotubes; a cell culture model of Duchenne’s muscular dystrophy (DMD) was illustrated in a recent study as a new chemically modified CPP, PF14. Starting from stearyl-TP10, ornithines were utilized as the main source of positive charges instead of lysines. The superior efficiency of poly-L-ornithines was related to the higher affinity for DNA and the ability to make more stable complexes at lower charge ratios [32].These studies show that CPPs are certainly among the most promising candidates in the development of siRNA-based therapeutics.Antisense technology is based on the use of sequence specific oligonucleotides (ONs) that can hybridize with complementary mRNA strands to cause translational arrest or mRNA degradation by activation of the cellular enzymes of the RNaseH family and consequently block gene expression [22]. The ONs with therapeutic potential include aptamers, transcription factor-binding decoy ONs, ribozymes, triplex-forming ONs, immunostimulatory CpG motifs, antisense ONs, and antagomirs.CPPs have been used for the delivery of ONs by using either a covalent linkage or a non-covalent linkage with the therapeutic agent. They have been used to mediate the delivery of PNAs and PMOs through covalent linkage [33]. The formation of efficient non-covalent complexes comprising CPPs and both charged and uncharged steric block oligonucleotides like 2'-O-methyl, LNA, PNA and charged PNA derivatives, has also been described [34].PNA–CPP conjugates were first demonstrated to block the expression of the galanin receptor mRNA in human Bowes cells by a 21-mer PNA coupled to Penetratin or Transportan by Langel and group in 1998 [22]. In a different study, a model amphipathic peptide (MAP) conjugated to a PNA complementary to the nociceptin/orphanin FQ receptor mRNA was shown to improve cellular uptake and steric block effect in both CHO cells and neonatal rat cardiomyocytes [35].The efficacy of PNA and PMO coupled to CPPs was only significant in the presence of endosomolytic agents such as chloroquine and calcium ions [36,37,38]. Since most of the existing endosomolytic agents are too toxic to be considered for in vivo applications, CPP-based strategies that are efficient in the absence of PNA and PMO such as co-treatment with endosome-disrupting peptides are currently being explored [39,40].Certain proteins are currently being used as therapeutic agents for the treatment of various diseases. CPPs are usually coupled to proteins through covalent bonds or through fusion constructs except Pep-1, a CPP that forms non-covalent complexes [16]. There is evidence that CPPs are able to facilitate the delivery of proteins into a wide variety of cells both in vitro and in vivo; Dowdy and group revealed that the delivery of these therapeutic proteins into tissues and across the blood-brain barrier was severely limited by the size and biochemical properties of the proteins. An intraperitoneal injection of the 120-kilodalton beta-galactosidase protein, fused to the protein transduction domain from the human immunodeficiency virus TAT protein, resulted in delivery of the biologically active fusion protein to all tissues in mice, including the brain [41]. CPPs therefore constitute a powerful tool that could be used to facilitate the delivery of protein-based therapeutics in pathological conditions, such as cancer, inflammatory diseases, oxidative stress-related disorders, diabetes and brain injury [36,42]. However, protein stability relies on weak non-covalent interactions between secondary, tertiary and quaternary structures, which must be preserved throughout the delivery process [43]. This makes proteins vulnerable therapeutic agents with short in vivo half-lives and poor bioavailability. To improve the efficiency of delivery of proteins into cells, different types of lipid- and polymer-based vectors including liposomes, microparticles and nanoparticles have been used for protein delivery, but with relatively poor efficiency [44].Liposomes, nanoparticles and other different types of pharmaceutical nanocarriers have been used to increase the stability of drugs, modulate their pharmacokinetics and biodistribution, improve their efficacy while decreasing side-effects [21]. However, the intracellular delivery of these large molecules remains a challenge because of their three-dimensional structure, spatial occupation and hydrophilic/hydrophobic nature.CPPs have been used to functionalize these vectors with a view to increasing the cellular uptake of the encapsulated therapeutic agents [45]. CPPs have been shown to offer the opportunity to deliver therapeutic molecules that are even 200 times larger than the current bioavailability size restriction. In one such study Torchilin prepared targeted long-circulating PEGylated liposomes and PEG-phosphatidylethanolamine (PEG-PE)-based micelles possessing several functionalities [21]. These liposomes and micelles were modified with TATp moieties attached to the surface of the nanocarriers by using TATp-short PEG-PE derivatives. This made them degradable by inserting the pH-sensitive hydrazine bond between PEG and PE (PEG-Hz-PE) whose cleavability by acidic hydrolysis makes them acquire the ability to be effectively internalized and released. These can be considered as an important step in the development of tumor-specific stimuli-sensitive delivery systems.Dextran-coated superparamagnetic iron oxide particles (CLIO) coupled with TATp, were demonstrated to provide efficient labeling of cells, and could serve as a tool for magnetic resonance imaging (MRI). The uptake of the TATp-CLIO nanoparticles by cells was shown to be about 100-fold higher than that of the non-modified iron oxide particle [21]. Other studies reveal that modifying the surface of nanoparticles with CPPs, enhances the cell permeability of nanoparticulate-based therapeutics [46]. This is further proved by the CPP-mediated delivery of bioactive compounds into model organisms for cancer, cardiomyopathy, stroke, muscular dystrophy and viral infections [47,48]. For example PsorBanR, a cyclosporin–poly-arginine conjugate formulated as a topical treatment for psoriasis, and KAI-9803, a PKC (protein kinase C) δ peptide inhibitor–TAT conjugate for the treatment of acute MI (myocardial infarction) [49].A family of cell-penetrating peptides named Vectocell® peptides [also termed DPVs (Diatos peptide vectors)] originating from human heparin binding proteins and/or anti-DNA antibodies with both enhanced and safe cell penetration characteristics have been identified. These new peptidic sequences are reported to deliver small and large active molecules inside cells that would otherwise have limited or no bioavailability [45].CPPs have been reported to act as active pharmaceutical ingredients when used alone. A TAT peptide containing a cysteine residue at its C-terminal (TAT-C) was shown to be able to inhibit infection by irreversibly inactivating virions exposed to the tat-C prior to cell infection, blocking entry of cell-adsorbed viruses, or inducing a state of resistance to infection in cells pretreated with TAT-C [50]. The mechanism of the antiviral activity is yet to be elaborated. This has set the stage for exploring the therapeutic activity of CPPs as active pharmaceutical ingredientsTAT and penetratin, the first CPPs to be described, paved the way to the discovery of other naturally occurring CPPs such as the herpes virus tegument protein VP22 and the cell wall protein-derived peptide inv3 from Mycobacterium tuberculosis [51], Chimaeric CPPs such as Transportan (a chimera of the neuropeptide galanin and the wasp venom toxin mastoparan) and totally synthetic CPPs such as the model amphipathic peptide (MAP) or arginine oligomers [52]. All these have been designed and are routinely used.Advancements have resulted in the combination of a tumor homing peptide with a cell-penetrating peptide. This could yield a chimeric peptide with tumor cell specificity that could carry therapeutic molecules into the cells. In a previous study involving the use of linear breast tumor homing peptide, CREKA, in conjunction with a cell-penetrating peptide, pVEC, it was demonstrated that CREKA–pVEC is a suitable vehicle for targeted intracellular delivery of a DNA alkylating agent, chlorambucil. The chlorambucil–peptide conjugate was shown in vitro to kill cancer cells faster than the anticancer drug alone [53].Several reviews have elaborated on examples of novel drug delivery systems involving CPPs. All these systems are geared towards targeting a specific cell or tissue. Many of the strategies described have proved successful in in vivo experiments. This has led to preclinical and clinical studies of CPP-based delivery strategies [54]. Such studies include:
+      
+        a.)
+          PsorBan® a cyclosporine-poly-arginine conjugate for the topical treatment of psoriasis was the first CPP mediated therapeutic agent which entered phase II trials in 2003 (CellGate, Inc.)., Delcasertib as KAI-9803 was recently tested by Kai Pharmaceutical as a TAT-protein kinase C inhibitor peptide modulator of protein kinase C for acute myocardial infarction and cerebral ischemia, and orally administrated cyclosporine A (CsA), effective against a broad range of inflammatory skin diseases including psoriasis, are examples of ongoing preclinical studies on effective delivery strategies [55]. Conjugation of a CPP, heptaarginine with CsA through a linker designed to release the active compound at the pH of the tissue has been shown to enhance its topical absorption, inhibiting cutaneous inflammation [56].
+        
+        b.)
+          Avi Biopharma is working on the clinical development of CPPs for the in vivo steric block splicing correction using 6-aminohexanoic acid spaced oligoarginine [(RAhx-R)4]. It consists of a Morpholino oligo conjugated with the CPP [(RXR)4-XB-CPP]. The goal of this conjugate is to prevent eventual blockage of a transplanted vein after cardiovascular bypass surgery [57]. Several other companies including Traversa Inc., and Panomics Inc. are also evaluating CPPs in preclinical and clinical trials in addition to other molecules conjugated to CPPs which are being optimized [58].
+        
+      PsorBan® a cyclosporine-poly-arginine conjugate for the topical treatment of psoriasis was the first CPP mediated therapeutic agent which entered phase II trials in 2003 (CellGate, Inc.)., Delcasertib as KAI-9803 was recently tested by Kai Pharmaceutical as a TAT-protein kinase C inhibitor peptide modulator of protein kinase C for acute myocardial infarction and cerebral ischemia, and orally administrated cyclosporine A (CsA), effective against a broad range of inflammatory skin diseases including psoriasis, are examples of ongoing preclinical studies on effective delivery strategies [55]. Conjugation of a CPP, heptaarginine with CsA through a linker designed to release the active compound at the pH of the tissue has been shown to enhance its topical absorption, inhibiting cutaneous inflammation [56].Avi Biopharma is working on the clinical development of CPPs for the in vivo steric block splicing correction using 6-aminohexanoic acid spaced oligoarginine [(RAhx-R)4]. It consists of a Morpholino oligo conjugated with the CPP [(RXR)4-XB-CPP]. The goal of this conjugate is to prevent eventual blockage of a transplanted vein after cardiovascular bypass surgery [57]. Several other companies including Traversa Inc., and Panomics Inc. are also evaluating CPPs in preclinical and clinical trials in addition to other molecules conjugated to CPPs which are being optimized [58].Phosphorodiamidate Morpholino Oligomers (PMOs) are antisense DNA oligonucleotide analogues with a backbone composed of morpholine rings joined by uncharged phosphodiamidate linkages in place of the sugar and anionic phosphodiester linkage of DNA. They are water-soluble, nuclease-resistant, and have an uncharged backbone, which can interact weakly with serum and cellular proteins thereby reducing toxicity. They therefore offer great promise in clinical applications [59]. As PMOs do not efficiently enter cells on their own, investigators have been routinely linking them to CPPs to promote their uptake into virus-infected cells and enhance their antisense efficacy.PMO-technology to inhibit viral infections has been exploited as extensively reviewed by Stein [60] and recently discussed by Moulton and Jiang [61] where a very potent CPP-PMO conjugate strongly inhibited HSV-1 replication in cell cultures by reducing viral protein expression. It was reported to have been able to suppress the replication of several HSV-1 strains, including an acyclovir-resistant strain.In addition to improving PMO cellular uptake, the CPP moiety has also been reported to intensify PMO antisense activity against Ebola virus (EBOV) 10- to 100-fold in cell-free translation assays. These results are consistent with other findings. The authors proposed that the arginine-rich peptides enhance RNA-PMO binding affinity, thereby increasing specific antisense activity [62].Due to poor cellular uptake of the uncharged peptide-nucleic acids (PNAs), Pandey and co-workers covalently conjugated PNA at its N terminal to various cell-penetrating peptides, including Transportan, TAT and penetrain, via disulfide bridges. Upon coupling to carrier peptides, the PNA therapeutic agents were rapidly taken up by different types of human cells in culture. When added to culture medium, these anti-HIV-1 PNA–CPP conjugates effectively inhibited HIV-1 replication, tat-dependent trans-activation and viral production by infected cells. No significant decrease was observed with the unconjugated PNA, emphasizing the great promise held by these conjugates as antiviral agents [63].Preliminary toxicity, immunological and pharmacokinetic studies in mice for the anti-HIV-1 PNATAR-Penetratin conjugate have just started. Investigating the tissue distribution and clearance of 125I-labeled PNATAR, PNATAR-penetrain and PNATAR-TAT; Ganguly’s group reported the distribution of the conjugates throughout the mouse major internal organs when administered by oral route as well as their slow release and clearance from different organs [64]. The unconjugated PNATAR was noted to display a similar tissue distribution and clearance profile although the extent of its uptake was lower than its CPP conjugate. This calls for further investigation.The use of CPPs to deliver siRNAs into cells has received rather less attention especially for antiviral siRNAs due to the fact that siRNAs is less amenable to CPP delivery as a result of charge interactions between the peptide and the siRNA which results in inefficient endosomal escape of the conjugates. However a cell-permeable CPP-siRNA conjugate targeting hepatitis C virus 5’ untranslated region has been formulated [63]. The cellular uptake and antiviral effect of this CPP-siRNA conjugate were reported to be as effective as transfection with lipofectamine in Huh-7 cell culture.In addition to targeting HIV-1, CPP-mediated protein delivery has also been used to inhibit human papillomavirus type 18 (HPV-18) in cell culture. Delivery of artificial zinc-finger proteins (AZPs) into cultured cells by expressing AZPs in fusion with the 9R-peptide has been reported. These AZP-9R conjugates strongly reduced HPV-18 replication to 3% at 250 nM while non-conjugated AZP showed only 12% reduction. PTD4 conjugates were also tested but proved to be less effective than AZPs fused to 9R [65].PNA-CPP complexes have also been investigated for virucidal activity microbicides [63]. The study reported that pre-incubation of HIV-1 virions with these molecules rendered them noninfectious and blocked further cell infection. It is suggested that the PNA-CPP may have altered or disrupted the viral envelope through the interaction of the CPP moiety with the viral lipid bilayer, thereby inhibiting host cell infection [52].However CPPs have their limitations as drug delivery carriers. The in vitro uptake studies revealed high cellular uptake values, but no specificity toward any of the cell lines. Their biodistribution in PC-3 tumor-bearing nude mice showed a high transient accumulation in well-perfused organs and a rapid clearance from the blood [55]. Data obtained in the study reveal that CPPs readily penetrate into most organs and show rapid clearance from the circulation. In view of this, it is suggested that CPPs are suitable as drug carriers for in vivo application provided that their cell specificity and pharmacokinetic properties are also considered in design and development of CPP-based drug delivery systems.Apart from functional group modification of amino acids in the peptide chain, modification of cell penetrating peptides have also been shown to involve substitution of amino acids to achieve variability of the peptide properties such as hydrophobicity or cationic nature. This strategy has been noted to result in increased intracellular internalization by certain CPPs. Kaeko et al. conducting a comprehensive search for novel CPPs using an in vitro virus library of peptides consisting of 15 amino acids [19] and reported improved intracellular translocation efficiency at low concentrations due to the influence of cationic amino acids. Since the amino acid Arginine has a stronger affinity to the cell surface, a desirable modification in this case involved substituting another amino acid such as Lysine in the peptide chain with Arginine. Such substitutions were reported [66] to have remarkably improved intracellular translocation even at low concentration. Substitution (of amino acids) with Histidine residue is also a promising option as this could provide endosomal disruption by the “proton sponge effect” of Histidine residues in the acidic endosomal compartment as previously reported for LAH4; an amphiphatic peptide rich in Alanine and Leucine. It bears Lysines at its ends to condense DNA and Histidines inside the sequence which favours endosomal escape. LAH4/DNA polyplexes were reported to be able to transfect cells in the absence of serum. From a series of LAH4 derivatives, high transfection efficiencies were obtained only with peptides containing four to five Histidine residues in the central region of the peptide sequence [67].Increasing amino acids like Tryptophan which has the tendency to be buried in the cell membrane [68] improves CPP hydrophobic interaction. It has been confirmed by Kaeko et al. that in low peptide concentrations, two Trp residues were not enough to initiate effective translocation, but three Trp enhanced cellular uptake. However, when the number of Tryptophan residues increased to four the intracellular translocation activity decreased [19]. This may be due to a decrease in solubility of the peptide as a result of the increased number of Tryptophan residues. Therefore, the increase of the number of Trp residues may be ideal but to a certain limit and such modifications should be carried out with strict consideration of the impact of amino acid substitution to the physicochemical parameters, especially changes in dissolution properties.Functional group modification of the amino acids in the peptide usually involves the formation of masking groups or linkages to highly reactive sites. In either case, the peptide bonds formed should be labile and easily broken for regeneration of the initial cell penetrating peptide by simple variation of physiological conditions.Helicity of the peptide chain can be exploited for modification. For example through hydrocarbon stapling the overall cell penetrating peptide’s alpha helicity could further be stabilized. This is derived from the fact that the α-helix, as a major structural motif of proteins, frequently mediates intracellular protein-protein interactions that govern many biological pathways [69].Fei and co workers reported a modification strategy to decrease the non-specific binding and uptake of a CPP, model amphipathic peptide (MAP, for use as a potential targeted drug carrier [68]. This was illustrated using (S)-a-(2’-pentenyl)alanine containing olefin-bearing tethers to generate an all hydrocarbon “staple” by ruthenium-catalyzed olefin metathesis. The (S)-a-(2’-pentenyl)alanine peptides were made to flank three (substitution positions l and l + 4) or six (l and l + 7) amino acids within the peptide, so that reactive olefinic residues would reside on the same face of the a-helix. The modified hydrocarbon-stapled peptides were helical, relatively protease-resistant, and cell-permeable peptides that bound with increased affinity for the target [70]. Such hydrocarbon stapling could provide a useful strategy for therapeutic modulation of protein-protein interactions.In another study the modification of MAP with citraconic anhydride (CA) blocks, the epsilon-amino groups of the Lysine residues formed acid-labile amide linkages. This was confirmed by analysis for internalization in HeLa cells which revealed that 80.8 ± 2.2% of the amino groups were modified in the CA-MAP conjugate. In the study a MAP conjugate was designed to contain acid-labile modifications to mask the cationic charge, and therefore decreasing the non-specific binding and uptake [71,72]. After binding and internalization, the acidic pH in the endosomes was realized to facilitate the cleavage of the acid-labile attachments and hence release of MAP. The combination of CA-MAP with a targeting molecule (e.g., folic acid) offers significant improvement in the use of CPPs in targeted drug delivery.Over the years of research focus has been on how to optimize drug delivery systems to increase both cell specificity and drug delivery efficiency. The use of cell penetrating peptides demonstrates the possibility that these drug systems can be improved. Efficient targeting by utilizing multi-functional polymer vehicles with capabilities for endosome disruption and nuclear penetration if appropriately considered can yield cell-penetrating peptides with high cellular specificity. CPP modification presents viable options that such delivery systems can further be improved. This could result in biomolecules that could be triggered spatially and temporally within a cell transport pathway hence truly improving drug therapy. There is therefore a need to develop assays to compare the different existing methods in order to study and optimize modified CPP-mediated delivery.Sincere gratitude is expressed to all who supported the preparation of this work in all capacities, and to the National Natural Science Foundation of China (No. 30901867 and 30973649).

Med-MDPI/biomolecules/biomolecules-02-02-00203.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Many viral proteins have been shown to be sumoylated with corresponding regulatory effects on their protein function, indicating that this host cell modification process is widely exploited by viral pathogens to control viral activity. In addition to using sumoylation to regulate their own proteins, several viral pathogens have been shown to modulate overall host sumoylation levels. Given the large number of cellular targets for SUMO addition and the breadth of critical cellular processes that are regulated via sumoylation, viral modulation of overall sumoylation presumably alters the cellular environment to ensure that it is favorable for viral reproduction and/or persistence. Like some viruses, certain bacterial plant pathogens also target the sumoylation system, usually decreasing sumoylation to disrupt host anti-pathogen responses. The recent demonstration that Listeria monocytogenes also disrupts host sumoylation, and that this is required for efficient infection, extends the plant pathogen observations to a human pathogen and suggests that pathogen modulation of host sumoylation may be more widespread than previously appreciated. This review will focus on recent aspects of how pathogens modulate the host sumoylation system and how this benefits the pathogen.Pathogens, especially intracellular ones like viruses, have evolved numerous mechanisms to counteract host defenses and to reshape the cellular environment into a permissive milieu that allows pathogen persistence and/or replication. Because of their limited genetic repertoires, viruses are highly efficient and typically encode proteins with multiple functions and which target critical cellular regulatory pathways to evoke maximal effects on the host cell. Well studied examples of viral manipulation of host systems include viral transcription factors that also regulate cellular gene expression [1], viral proteins that inactivate p53 to circumvent cell cycle arrest and apoptosis (reviewed in [2]), and viral proteins that drive cells into the proliferative phase (reviewed in [3]). An additional established theme is viral interaction with the host ubiquitin-proteasome system to regulate cellular activities through targeted degradation or protection of host proteins (reviewed in [4]). In many cases viral proteins co-opt cellular ubiquitin E3 ligases to re-direct substrate specificity or alter ligase activity. Alternatively, some viral proteins have intrinsic ubiquitin E3 ligase activity that directly promotes ubiquitination of selected host substrates. Both mechanisms lead to increased degradation of specific host proteins and thus diminish their available functional activity. In contrast, several well documented examples exist of viral proteins that have deubiquitinating activity, though in most cases their authentic substrates have not yet been identified. Presumably each of these viral deubiquitinating proteins removes ubiquitin from one or more host proteins thus decreasing their degradation and protecting host activities critical to viral pathogenesis or reproduction. In addition to the well-studied ubiquitin system, the ubiquitin super family includes a number of ubiquitin-like proteins including Atg8 [5], Nedd8 [6], ISG15 [7], and SUMO [8,9,10,11] that are all enzymatically conjugated to substrate proteins. Among these ubiquitin-like proteins, SUMO has been shown to modify numerous DNA and RNA virus proteins and is clearly an important host system during viral infection [12,13,14].Over the last 15 years, a new post-translational modification system has been defined that is enzymatically analogous to, but functionally distinct from the classical ubiquitin system [15,16]. This related system involves target protein modification of one or more lysine residues by conjugation of an ubiquitin-like, small polypeptide known as SUMO [17]. The human SUMO-1 gene encodes a 101 amino acid polypeptide with ~50% relatedness to the S. cerevisiae SMT3 protein and ~18% sequence relatedness to ubiquitin [8,10,11]. Vertebrate species have at least 2 additional genes for SUMO1 related proteins, SMT3A (SUMO2) and SMT3B (SUMO3) [18]. The three SUMO genes appear to form two subfamilies, as SUMO2 and SUMO3 share 87% sequence identity compared to only ~50% identity between SUMO2/3 and SUMO1. Less well studied is a fourth SUMO gene whose expression is restricted primarily to the kidneys, dendritic cells, and macrophages [19]; SUMO4 may play a role in diabetes [20], but overall its biological functions and activity are not well defined.Like ubiquitination, sumoylation involves formation of a stable isopeptide bond between the ε-amino group of the target lysine and the carboxyl group of the C-terminal glycine of SUMO. As for ubiquitin, SUMO must first be proteolytically processed at its C-terminus, enzymatically activated, and then covalently attached via a thioester linkage to a cysteine residue of its conjugating enzyme, Ubc9. However, the SUMO proteases, the heterodimeric activating enzyme (Aos1/Uba2, called SAE1/SAE2 in this review), and Ubc9 are all specific for SUMO usage and do not function with ubiquitin [21,22]. Furthermore, transfer of SUMO to target proteins can occur directly from Ubc9, apparently without an absolute requirement for substrate-specific E3-type ligase enzymes that are necessary for ubiquitin addition [23]. As would be expected from this lack of an absolute E3 requirement, numerous target proteins for SUMO1 modification have been identified by their direct interaction with Ubc9 [8,9,21,24,25,26]. However, more recently it has become clear that there are SUMO-specific E3 ligases, and several types have been identified, including members of the PIAS protein family (Protein Inhibitors of Activated Stat) [27,28,29]; RanBP2, a nuclear pore protein [30]; the polycomb protein, Pc2 [31]; TOPORS, topoisomerase I binding protein [32], and members of the TRIM family [33]. Additional SUMO ligases are likely to be discovered and they undoubtedly will have important roles in regulating the functional biology of SUMO targets.Lastly, the sumoylation state of target proteins is not static, but instead reflects a dynamic equilibrium between the forward process of SUMO addition and its removal by cellular desumoylating enzymes known as SENPs [34]. The mammalian SUMO proteases differ greatly in their sequences and are related primarily in the conserved region critical for cysteine protease catalytic activity. Furthermore, individual proteases have been shown to differ in intracellular localization with both nuclear [35,36] and cytoplasmic [37,38,39] species observed. The existence of multiple mammalian desumoylating enzymes, along with the demonstrated differences in intracellular distribution, suggests that desumoylation is likely to be a complex process that contributes to the regulation of activity of SUMO substrates.Functionally, sumoylation is now implicated in a diverse array of critical cellular processes including nuclear processes such as RNA processing, chromatin remodeling, genome maintenance, transcriptional regulation, mitosis, meiosis, differentiation and development, apoptosis, and nucleocytoplasmic transport [40]. More recently, significant non-nuclear functions of sumoylation have been identified in regulation of ion channel activity [41,42] and metabolic pathways [43,44]. Because of this pleiotropic ability to modify numerous proteins and affect a wide range of cellular processes, sumoylation is an attractive target for pathogens to use in modulating the cellular environment to favor pathogen replication and/or maintenance. There are numerous examples of viral proteins that are sumoylated (Table 1), so utilization of this host system by viruses to regulate viral protein function is well documented and will not be further explored in this review (for recent review see [45]). In contrast, recent examples have been described where pathogens alter host sumoylation, either globally or for host specific targets, and the following sections will examine mechanisms by which both viral and bacterial pathogens perturb the host sumoylation process.Sumoylated Viral Proteins.Immunity to pathogens consists of innate and acquired responses that collectively block or diminish infections and usually leads to resolution of acute disease. Upon first exposure to a new pathogen the innate responses, primarily mediated through induction of interferon and cytokines, are a critical initial defense that dampens pathogen growth until pathogen-specific antibodies and T-cells are developed [46]. The innate response is triggered by recognition of pathogen-specific molecules by the Toll-like receptor (TLR) system which then initiates a signaling cascade that activates IRFs 3, 5, and 7 via phosphorylation; the activated IRFs turn on interferon transcription. Two of the TLRs, RIG-1 [47] and MDA5 [48], and an adaptor protein in the TLR signaling pathway known as Pellino-1 [49] have recently been shown to be sumoylated proteins. Sumoylation of RIG-1 enhanced its association with Cardif and led to increased expression of interferon β. Likewise, MDA5 sumoylation also increased interferon production, indicating that sumoylation is a positive regulator of these two TLR pathways and possibly other TLRs. For both MDA5 and RIG-1, sumoylation and the subsequent increased interferon production repressed viral replication, while reducing overall sumoylation led to increased viral replication. These results suggest that viral down regulation of sumoylation would be beneficial to avoid triggering a strong innate response.Downstream of the TLRs, sumoylation of both murine IRF3 and IRF7 in response to viral infection has been reported [50]. Interestingly, sumoylation defective mutants of either protein led to increased interferon gene expression after viral infection suggesting that sumoylation is a negative regulator at this stage of the pathway. However, a subsequent study examining endogenous human IRF3 came to the opposite conclusion [51]. Ran et al. showed that sumoylation of human IRF3 competed with ubiquitination and protected IRF3 from degradation. When sumoylation of IRF3 was elevated by reduction in the levels of the SENP2 SUMO protease it resulted in decreased viral replication, suggesting that sumoylation was a positive regulator of innate immunity. Whether these conflicting results reflect differences in human versus murine IRF3 or differences in the experimental approaches, it is clear that sumoylation is a contributor to the regulatory process that governs the initiation of the interferon response. Consequently, it is likely that viruses have evolved mechanisms to thwart or usurp this upstream portion of the pathway as well as the documented viral effects on downstream effectors such as PKR and 2’,5’-oligoA synthetase [52]. While there are not yet examples of viruses specifically targeting the sumoylation of TLRs or their signaling pathways, several viruses can alter global cellular sumoylation (see sections below) which could impact the functionality of the innate immune response.In addition to classical innate immunity mediated through interferons and cytokines, the concept of intrinsic immunity has developed in recent years [53]. Unlike the induction cascade required to activate innate immunity, intrinsic immunity operates through pre-existing proteins that act to repress viral infection. Several members of the TRIM family of proteins have been shown to participate in intrinsic immunity, and while some can be up regulated by interferon, their anti-viral activity does not directly require interferon [54]. There is now mounting evidence that sumoylation is important in directing TRIM proteins to their targets. For example, the cytoplasmic TRIM5α protein can block certain retroviral infections [55,56]. A recent publication demonstrated that TRIM5α has three potential SUMO-interacting motifs (SIMs) and that the anti-viral activity of TRIM5α requires SIM1 and SIM2, but not SIM3 [57]. Mutations in the murine leukemia virus (MLV) capsid antigen (CA) that blocked CA sumoylation abrogated the ability of TRIM5α to restrict MLV replication. From these results the authors propose that restriction by TRIM5α, at least in part, involves direct binding of TRIM5α to sumoylated CA through the SIM1 and SIM2 regions in TRIM5α. This SUMO-SIM facilitated interaction presumably then leads to some change in capsid structure that disrupts the normal infection process and restricts viral infection.Another member of the TRIM family, PML, has general intrinsic anti-viral activity against both DNA and some RNA viruses [58]. It has generally been observed that cells with decreased or absent PML have enhanced viral replication [59,60,61,62] and that PML-/- mice are more susceptible to certain viral infections [63]. Consistent with an anti-viral role for PML, a large number of viruses specifically target PML NBs for disruption and/or degradation, and abrogation of this viral function typically impairs viral reproduction. The interplay between viruses and PML NBs has been discussed extensively in several recent reviews [45,58,64], however, some recent examples that highlight the role of sumoylation in the virus-PML interaction are presented below and in section 4.4.Elegant studies from Roger Everett’s group found that PML and other NB proteins, such as Sp100 and Daxx, have SIM motifs that are required for recruitment of these proteins to herpes simplex virus (HSV) replication foci [65]. Absence of the SIM motif does not affect PML mobility or normal assembly into NBs, so the SIM motif seems to function primarily to direct the formation of NB protein complexes on viral replication complexes. These NB complexes are transient in wild type HSV infection due to their disruption by the early gene product ICP0. However, an HSV ICP0 null mutant is normally highly defective for replication unless PML is depleted, suggesting that recruitment of the NB proteins creates a repressive environment at the replication foci that would block viral replication unless overcome by ICP0 [65,66]. Reintroduction of wild type PML decreases viral production by the ICP0 mutant while reintroduction of a SIM minus PML does not, strongly implicating the SIM motif in targeting PML to the replication foci [65]. Consistent with the importance of the PML SIM, viral replication foci were found to have a significant deposition of SUMO1 and SUMO2/3 which could serve as the platform for SIM binding. Furthermore, knockdown of the sole SUMO conjugating enzyme, Ubc9, reduces accumulation of SUMO1 and SUMO2/3 signals at the viral foci suggesting that de novo sumoylation is needed for the accretion of the sumoylated proteins at these sites [67]. Ubc9 knockdown also results in greatly enhanced replication of the ICP0 null mutant which is again consistent with sumoylation contributing to intrinsic anti-viral activity. Based on these observations the authors proposed that sumoylated proteins assemble on the viral genomes and serve as the signal to recruit PML and other SIM containing NB proteins to viral replication complexes. Assembly of PML, Sp100, and Daxx would normally inhibit HSV replication if not counteracted by the ICP0 protein (see below). Presumably this SIM-SUMO dependent recruitment could be a general mechanism that accounts for PML anti-viral activity on various other invading viral genomes. What the actual sumoylated proteins are and how they might be directed to the nascent viral replication foci is unknown, but could be related to the viral genomes eliciting the cellular DNA damage response.Not only is the SIM-SUMO interaction important for directing the anti-viral activity of PML NBs, there is increasing evidence that viruses use SIM-SUMO interactions to thwart the effects of PML. PML is itself sumoylated at three sites [68], so PML accumulation at viral replication foci would contribute to the concentration of SUMO groups localized at these foci. Recent studies on herpes varicella-zoster virus (VSV) ORF61 protein showed that ORF61 has a functional SIM motif in the C-terminal domain, and that mutation of this SIM motif significantly impaired the ability of ORF61 to disperse NBs and strongly reduced co-localization of ORF61 with PML NBs [69]. Surprisingly, mutation of this SIM motif did not impair viral replication in permissive cells in vitro, but was required for pathogenesis and skin lesions in a SCID mouse model with human skin xenografts. While the mechanism by which ORF61 disperses PML is still unclear, these results strongly suggest that ORF61 targeting to NBs is mediated through the ORF61 SIM interacting with sumoylated proteins, possibly sumoylated PML.Similar to the ORF61 story, the herpes simplex virus (HSV) ICP0 protein has several putative SIM-like sequences (SLSs), and SLS-4, -5, and -7 contribute to SUMO binding, to degradation of SUMO modified PML isoforms, and to the biological ability of ICP0 to overcome intrinsic anti-viral defenses [67]. Early in infection, unknown SUMO conjugates, along with PML, are recruited to viral genomes that can be visualized as punctate foci in the nucleus. ICP0 is subsequently recruited to these foci in a SUMO2/3-dependent fashion, suggesting that recruitment depends on SIM-SUMO interactions. This ICP0 recruitment does not require PML, so presumably the unknown sumoylated proteins that accumulate with the viral genomes are sufficient to direct ICP0. As for the ORF61 protein, it appears that SUMO moieties are critical signals that mediate localization of ICP0 to the viral genomes in order to counteract the repressive effect of PML and NB proteins. While a number of other viral proteins also disrupt NBs, the role of SIM motifs for targeting these proteins has not been reported; further discussion of NB disruption can be found in section 4.4.In summary, there is growing evidence that many host proteins involved in innate and intrinsic immunity are regulated by sumoylation, thus their function could be dysregulated by viral attacks on the sumoylation system. Alternatively, the SUMO moieties on these host proteins can serve as targeting signals to recruit viral proteins containing one or more SIM sequences. These SUMO-SIM interactions are critical for directing some viral proteins to cellular defense complexes, such as PML NBs, where the viral proteins can subsequently counteract the defense mechanism. Consequently, sumoylation appears to be a highly important component of viral strategies for overcoming host defenses early in infection.The previous section detailed some of the recent work defining the interplay of viruses and sumoylation in the innate and intrinsic anti-viral response. However, given the pleiotropic roles for sumoylation in various cellular processes, it is likely the pathogens modulate sumoylation to alter many other facets of the cellular environment, not just immune response. Theoretically, pathogens could mimic or co-opt any step in the sumoylation process in order to alter the sumoylation status of viral or host proteins, and there are now excellent examples for many steps in the pathway (Figure 1). The following sections will describe examples where viral proteins perturb host sumoylation, either globally or for specific targets.Pathogens and the sumoylation pathway. The enzymology of the sumoylation is shown from the processing of the SUMO precursor to the final conjugation/deconjugation of substrates by the SUMO protease, SENP. Pathogen proteins that act at various steps in the sumoylation pathway are shown in boxes and are discussed in the text. There are now well documented examples of pathogens that express proteins that mimic either SUMO proteases or SUMO ligases. Xanthomonas campestris, a plant pathogen, uses a type III secretion system to inject effector proteins into host cells during infection. Among the injected effectors are two proteins, XopD and AvrXv4, that both decrease overall host sumoylation [70,71,72]. XopD resembles the yeast SUMO protease, Ulp1, can deconjugate SUMO from substrates in vitro and in vivo, and causes an overall decrease in host sumoylation when exogenously expressed [70]. In a subsequent study it was confirmed that XopD is a virulence factor in tomatoes and that virulence at least partially requires the protease activity, though the critical target(s) for desumoylation were not identified [72]. A second Xanthomonas protein, AvrXv4, resembles a cysteine protease and causes desumoylation in plants [71]. However, in vitro SUMO protease activity by AvrXv4 could not be demonstrated raising the possibility that AvrXv4 is not directly a SUMO protease but somehow influences cellular SUMO proteases. Like XopD, AvrXv4 has a role in virulence, and the combined results strongly suggest that Xanthomonas is using SUMO deconjugation to alter the host cell environment to favor bacterial infection.There have been reports of two viral proteins that exhibit SUMO ligase activity, the Kaposi’s sarcoma-associated herpes virus (KSHV) K-bZIP protein [73] and the adenovirus E1B-55K protein [74,75]. K-bZIP is a nuclear transcription factor that is a strong repressor when sumoylated at lysine 158 [76]. Phosphorylation at threonine 111 prevents sumoylation and converts K-bZIP to a strong transcriptional activator [77]. Chang et al. showed that K-bZIP has a SIM motif at residues 73–77 that binds SUMO2/3 but not SUMO1 [73]. They also showed that K-bZIP acts as a SUMO2/3-specific SUMO ligase that enhances sumoylation of its binding partners, including p53 and pRB, and that the ligase activity required the intact SIM motif. Additionally, expression of K-bZIP in an inducible cell line resulted in a global increase in SUMO2/3 conjugates with no effect on SUMO1 conjugates, illustrating that K-zBIP could have wide-spread effects on cellular events through changing the sumoylation status of numerous proteins. Exactly how this benefits viral reproduction is not yet determined, but clearly K-bZIP is both regulated by sumoylation and is able in turn to regulate sumoylation of host proteins.In contrast to K-bZIP, the adenovirus E1B-55K protein appears to be a SUMO1-specific SUMO ligase [74]. It was originally reported that E1B-55K enhanced p53 sumoylation in vivo but lacked SUMO ligase activity in vitro, suggesting that it might be recruiting a cellular ligase to the E1B-55K/p53 complex [75]. However, a subsequent study showed both in vivo and in vitro enhancement of p53 sumoylation confirming that E1B-55K is itself SUMO ligase [74]. This ligase activity may be specific to p53 as E1B-55K did not increase sumoylation of pRB [75]. Interestingly, E1B-55K is itself sumoylated at lysine 104, and mutation of this residue reduces ligase activity on p53. Pennella et al. found that this same SUMO conjugation site mutation also impaired E1B-55K association with PML NBs and the E1B-55K-dependent export of p53 to cytoplasmic aggresomes [74]. Additionally, they demonstrated that sumoylation of p53 reduced its intracellular mobility and increased its tethering to NBs resulting in enhanced nuclear export of p53. Based on these observations they proposed that the SUMO ligase activity of E1B-55K is a mechanism that contributes to repression of p53 function by increasing p53 sumoylation which facilitates the association of p53 with NBs and its subsequent export from the nucleus.STUbLs (SUMO Targeted Ubiquitin Ligases) are a group of ubiquitin ligases that contain SIM motifs that target the ligases to sumoylated proteins [78]. A recent report by Boutell et al. demonstrated that the herpes simples ICP0 protein has properties suggestive of a STUbL [67]. ICP0 has 7 predicted SIM-like sequences (SLSs), and at least some of them are functional for SUMO binding. In yeast two-hybrid assays ICP0 interacts in a SLS-4 dependent manner with SUMO2/3 but not SUMO1. In contrast, in in vitro pull-down assays the C-terminal portion of ICP0 binds SUMO1 and not SUMO2/3; SLS-5 appeared to be important for this SUMO1 interaction, but other sequences may also contribute. The discrepancy between the two-hybrid and pull-down assay results was not resolved, but together they support the interpretation that ICP0 may make interactions with all the SUMO isoform. In addition to SUMO binding, ICP0 can specifically ubiquitinate poly-SUMO chains in vitro. This activity requires both the RING domain and an intact SLS-4 sequence. ICP0 SLS-4 mutants retain full ubiquitinating activity on other substrates, so the SIM motif appears necessary only for directing ICP0 to sumoylated substrates. Consistent with the in vitro results, expression of ICP0 in a stable cell line resulted in a decrease in global SUMO conjugates, including sumoylated forms of PML, and this activity was reduced though not entirely eliminated in an SLS-4 mutant. These results suggest that ICP0 is ubiquitinating sumoylated proteins and targeting them for proteasomal degradation. Thus, it appears that the SLSs of ICP0 can redirect the intrinsic ubiquitin ligase activity to sumoylated proteins, such as PML, which may contribute to ICP0’s ability to overcome intrinsic cellular resistance to HSV. Given the growing number of known viral ubiquitin ligases [79], it would be surprising if others were not also STUbLs.A large number of pathogen proteins have been shown to target the sumoylation system in order to modulate overall sumoylation levels in the host cell. Many pathogen express proteins that target Ubc9, the SUMO conjugation enzyme, but in some cases other sumoylation enzymes are also targeted. The one known example of a pathogen protein that targets the SUMO activating enzyme (SAE) is the Chicken Embryo Lethal Orphan (CELO) adenovirus GAM1 protein [80]. GAM1 is critical for CELO replication [81], and it was quickly shown that GAM1 expression decreased sumoylation of the HDAC1 histone deacetylase [82]. A subsequent study revealed that GAM1 expression leads to a dramatic decrease in total SUMO conjugated products in the cell, that GAM1 binds the SAE1/2 SUMO activation enzyme and inhibits its activity in vitro, and also strongly reduced intracellular levels of SAE1, SAE2, and Ubc9 [80]. Boggio et al. further demonstrated that GAM1 binds the cullin proteins, Cul2 and Cul5, and recruits the Cul2/5-EloB/C-Roc1 ubiquitin ligase complexes to SAE1/2 [83]. Invitro the recruitment of the ubiquitin ligase causes enhanced ubiquitination of SAE1and in vivo a GAM1 mutant that cannot bind the cullins is unable to decrease SAE1/2 levels, implicating ubiquitin mediated proteasomal degradation as the mechanism by which the SUMO activating enzyme is reduced in the presence of GAM1. However, SAE2 is not directly ubiquitinated in the ligase complex, so the loss of SAE2 appears to be due to its destabilization when not in complex with SAE1. The net result of GAM1 action on SAE1 is dramatic reduction in levels of both SAE1 and SAE2 with the resultant cessation of de novo sumoylation. Sumoylation is generally associated with transcriptional repression [84], so global loss of sumoylation should result in a much more transcriptionally robust cellular environment for the virus [85].There are two known examples of viral proteins targeting SUMO ligases, the human papillomavirus (HPV) E6 protein [86] and the Ebola virus VP35 protein [87]. The HPV E6 binds the PIASy SUMO ligase to inhibit sumoylation of PIASy substrates [86]. E6 did not induce degradation of PIASy so the inhibition of activity appears to be a direct result of binding. This binding and inhibitory activity was restricted to the high risk HPV 16 E6 protein and was absent in the low risk HPV11 E6 protein. Interestingly, the ability of E6 to overcome PIASy induced senescence was not dependent upon p53 so must be operating through other PIASy targets, such as pRB. The ability of the 16E6 protein to inhibit sumoylation-promoted induction of senescence through the pRB pathway would clearly be of value for viral infection and possibly transformation.In contrast to the HPV 16E6 protein, the Ebola VP35 protein stimulates the ligase activity of its target, PIAS1. PIAS1 is an endogenous SUMO ligase for IRF7. Sumoylation of IRF7 decreases its transcriptional activity on the interferon β promoter and may be part of a normal feedback process to attenuate the interferon response and limit inflammation. In co-immunoprecipitations VP35 was found complexed to both PIAS1 and IRF7, and VP35 expression led to enhanced sumoylation of IRF7 that reduced its transcriptional activity. When co-expressed with a PIAS1 inactive mutant VP35 was unable to enhance sumoylation of IRF7, indicating that the VP35 effect on IRF7 is mediated through PIAS1 alone. Normally IRF7 is sumoylated at lysine 406, but VP35 induced sumoylation at multiple lysines, even in a lysine 406 mutant, suggesting that VP35 is causing promiscuous sumoylation of this target to ensure its transcriptional repression. Similar effects were seen with IRF3. Consequently, it appears that Ebola is using VP35 to down regulate the innate immune response by hypersumoylating two transcription factors involved in interferon induction to reduce their transactivation capacity and decrease interferon production.While there are only a limited number of viral proteins known to target either SAE or SUMO ligases, pathogen proteins affecting Ubc9 are more numerous, suggesting that Ubc9 is a very effective cellular target for manipulating sumoylation either globally or for specific substrates. Note that most sumoylated proteins, including pathogen proteins, interact with Ubc9 as part of the sumoylation process, however, the focus in this section will only be on interactions that alter or potentially alter Ubc9 function. Examples have now been documented for both global sumoylation increases and decreases mediated by different pathogens, so both scenarios apparently can provide benefit in a pathogen-specific manner. The two examples of pathogen-dependent decreases in sumoylation via effects on Ubc9 are the Listeriamonocytogenes LLO protein [88] and the HPV E6 protein [89]. L. monocytogenes is a facultative intracellular human pathogen associated with foodborne illnesses. Infection of HeLa cells with L. monocytogenes results in a general decrease in overall SUMO1 and SUMO2/3 conjugates. This decrease does not require bacterial entry into the cells, but is dependent on a known virulence factor, the listeriolysin O (LLO). Administration of purified LLO to cells also decreased sumoylation, but had no effect on invitro sumoylation indicating that LLO was not directly inhibiting enzyme activity. Cell culture studies revealed that LLO was reducing intracellular Ubc9 levels with no effect on SAE, and a similar LLO-dependent decrease in Ubc9 was observed in an infected mouse model. Over expression of SUMO1 or SUMO2 in infected HeLa cells led to increased overall sumoylation and significantly decreased numbers of intracellular bacteria at seven hours post-infection, indicating that sumoylation is detrimental to bacterial reproduction and must be reduced to allow optimal pathogen growth. Mechanistically, the reduction of Ubc9 was not at the transcriptional level and could not be prevented by proteasome inhibition. However, an aspartyl protease inhibitor partially restored Ubc9 levels in the presence of LLO, suggesting that LLO may be acting through a cellular protease that can target Ubc9. Related toxins from Clostridium perfringens and Streptococcus pneumonia also reduced Ubc9 levels, indicating that host sumoylation may be generally restrictive to other bacterial pathogens and that these pathogens have evolved mechanisms to counteract this host system.The high risk HPV E6 proteins, 16E6 and 18E6, also reduce both intracellular Ubc9 levels and overall sumoylation; low risk 11E6 had no effect on Ubc9 or sumoylation, but other low risk E6 proteins were not tested so the generality of this observation is not known [89]. High risk E6 proteins are known to induce proteasomal degradation of several host proteins via direct binding of the target and recruitment of a host ubiquitin ligase called E6AP [90]. Consistent with a possible proteasomal degradation mechanism, both 16E6 and 18E6 bound Ubc9 in pull-down assays. Additionally, the in vivo reduction in Ubc9 levels was E6AP dependent, and there was no effect at the transcript level. However, standard proteasomal inhibitors were relatively ineffective in restoring Ubc9 levels so the mechanism of degradation remains uncertain. Furthermore, the critical sumoylated substrates have not been identified and the direct advantage to the virus of reduced sumoylation has not been established, so the biological contribution of this E6 activity is uncertain. Nonetheless, the fact that E6 targets two sumoylation enzymes, Ubc9 and PIASy (discussed above), suggests that modulation of sumoylation is important to the viral life cycle. Interestingly, another HPV protein, the minor capsid protein L2, increases global host sumoylation, though the mechanistic basis for this increase has not been explored and may not involve Ubc9 [91]. Since L2 is a late protein and E6 an early protein, it is possible that both up and down regulation of sumoylation are required at different stages of the viral cycle.As opposed to the global reduction in sumoylation by LLO and E6, the Epstein-Barr virus (EBV) LMP1 oncoprotein targets Ubc9 and causes enhanced global sumoylation [92]. LMP1 is the latent membrane protein and possesses six transmembrane domains and a 200-amino acid C-terminal cytoplasmic tail containing three CTAR motifs. Bentz et al. showed that LMP1 binds Ubc9 through CTAR3 and that the interaction requires enzymatically active Ubc9. The LMP1-Ubc9 interaction leads to an increase in total SUMO conjugates with SUMO1, SUMO2, or SUMO3, and this increase is eliminated by a CTAR3 deletion. Abrogating the increase in sumoylation with the CTAR3 deletion does not cause an effect on host cell growth, but does reduce the ability of LMP1 to stimulate cell migration in a scratch assay, indicating that LMP1’s ability to increase sumoylation does have biological implications. Mechanistically, LMP1 did not cause any change in Ubc9 protein levels so the authors proposed that LMP1 may be acting as a SUMO ligase, though the critical targets for the migration effect remain to be determined. A similar global increase in sumoylation was also observed during influenza A virus infection of HEK293, A549, MDCK, and Vero cells [93]. The viral protein mediator was not identified, and there was no change in intracellular levels of Ubc9, SAE1, or SAE2, so the viral target that causes this increase in sumoylation is unknown. However, the increase in sumoylation could not be mimicked by interferon treatment alone and did not occur with UV inactivated virus, suggesting that one or more viral products produced during active infection were responsible. It will be interesting to see if this effect is somehow mediated through Ubc9 as for LMP1 or if it involves a different mechanism.An intriguing but more limited viral effect on Ubc9 has recently been reported for the adenovirus E1A oncoprotein [94]. Interaction of E1A to Ubc9 was first reported in 1996 [25], and Yousef et al. extended those studies by exploring the interaction in detail [94]. The authors found that the Ubc9 interaction motif on E1A is the sequence EVIDLT in the conserved region 2 (CR2) domain. While this sequence resembles a SIM motif, sumoylation of Ubc9 was not required for the interaction with E1A. Furthermore, unlike the previous examples discussed, E1A interaction with Ubc9 did not appear to affect overall sumoylation in vivo nor did it affect in vitro sumoylation of two test substrates, HDAC4 or E2-25K proteins. Analysis of Ubc9 mutants indicated that E1A binds to Ubc9 in the N-terminal region involving sequences that overlap but are not identical to those involved with SUMO binding by Ubc9. Since SUMO binding by Ubc9 is critical for formation of SUMO chains (poly-sumoylation), E1A was tested for an effect on poly-sumoylation in a yeast assay. Ulp2 deleted yeast cannot grow at 37 °C due to toxic accumulation of poly-sumoylated proteins, but wild type E1A is able to restore efficient growth while a Ubc9 non-binding E1A mutant cannot. The authors concluded from this result that while E1A does not appear to grossly alter the mono-sumoylation pattern that it can inhibit poly-sumoylation which may affect the function of specific host proteins. They also speculated that E1A binding might affect the ability of Ubc9 to be sumoylated at lysine 14. Sumoylation at this residue alters Ubc9 substrate specificity [95], so this could also have effects on sumoylation of host proteins. How these subtle E1A effects on Ubc9 contribute to viral fitness will require further study.Lastly, while a specific viral protein mediator has not yet been established, it is worth mentioning another example by which host sumoylation could be perturbed. Coxsackievirus B5 (CVB5) infection of HeLa cells does not alter Ubc9 levels, but causes a dramatic dispersal of Ubc9 from perinuclear aggregates into a diffuse signal throughout the cell [96]. How re-localization affects SUMO conjugates, either globally or for specific substrates, was not examined, but it seem likely that this dispersal of Ubc9 is another viral mechanism for altering sumoylation of host proteins. In summary, it is clear from the examples discussed that both viral and bacterial pathogens have evolved diverse mechanisms to target Ubc9 and influence the host SUMOeome.The previous section described pathogens and pathogen proteins that targeted sumoylation enzymes to cause broad effects on multiple host SUMO substrates. This section will focus on how pathogens modulate sumoylation of specific target proteins. As might be anticipated, the target proteins identified to date include critical host cell regulators of growth or pathogen resistance, such as p53, pRB, PML, and Daxx. All of these host proteins are known to be regulated by sumoylation, so it not surprising that pathogens have developed strategies to dysregulate these proteins through modulation of their sumoylation status.One of the important host cell proteins whose sumoylation is specifically impacted by pathogen proteins is the cell cycle regulator, pRB [97]. pRB is a tumor suppressor that inhibits cell cycle progression by blocking E2F activity until phosphorylation of pRB by cyclin kinases at late G1/early S phase releases E2F [98]. Ledl et al. showed that pRB is sumoylated at lysine 720 in the B-box motif of the so called pocket region, and that the hypophosphorylated form of pRB is preferentially sumoylated [97]. Lysine 720 is part of a “lysine cluster” that contributes to interaction with LxCxE containing proteins such as the adenovirus E1A protein and the papillomavirus E7 protein. Both E1A and E7 bind pRB and displace E2F to stimulate cellular entry into S phase which is critical for viral replication. As suspected from the known binding surfaces, addition of E1A protein to an in vitro sumoylation reaction prevents pRB sumoylation, and this inhibition requires direct E1A-pRB interaction. Similarly, in vivo expression of either E1A or E7 inhibits pRB sumoylation. Lastly, they showed that sumoylation of pRB decreases its repressive activity on an E2F driven promoter, indicating that sumoylation is an actual regulator of pRB and that E1A or E7 inhibition of sumoylation would have a functional effect. Whether or not this effect on pRB function is an irrelevant consequence of E1A or E7 protein binding or actually contributes to creation of the permissive environment has not been investigated.p53 is another cellular tumor suppressor that can elicit both cell cycle arrest and apoptosis in response to infection, so it is a primary target for many viral proteins that seek to limit p53’s functions [2]. As discussed in section 4.1, the adenovirus E1B-55K protein is a SUMO ligase that specifically targets p53 and enhances its sumoylation with SUMO1 [74]. This viral enhanced sumoylation contributes to p53 localization and tethering in PML NBs which results in p53 export and accumulation in cytoplasmic aggregates where it is nonfunctional. This clearly constitutes viral hijacking of the sumoylation system to suppress the activity of a host defense protein.Another interesting and so far unique example of a viral protein influencing the sumoylation of a specific host protein is that of the Kaposi’s sarcoma-associated herpesvirus (KSHV) viral protein kinase (vPK/ORF36) and the host KAP-1 protein [99]. KAP-1 is a transcriptional repressor that recruits chromatin remodeling proteins, such as HDACs, to inactivate cellular promoters [100]. KAP-1 is sumoylated at three lysines, and the conjugated SUMO moieties contribute to binding of the chromatin remodeling proteins and hence the repressive function of KAP-1 [101]. KAP-1 is also phosphorylated, and phosphorylation of serine 824 reduces sumoylation at all three lysines and thus decreases the repressive activity of KAP-1 [102]. Chang et al. subsequently presented evidence that KAP-1 is involved in the switch between latent and lytic growth for KSHV [99]. Importantly, they showed that the viral kinase could phosphorylate KAP-1 leading to decreased KAP-1 sumoylation and decreased occupancy of KAP-1 on viral promoters. These results point to a model where KSHV uses phosphorylation of KAP-1 by the vPK as a mechanism to antagonize KAP-1 sumoylation and overcome KAP-1 repression, thus favoring lytic replication over latency.Finally, many viruses have developed mechanisms to disrupt sumoylation of PML and associated proteins such as Daxx. Given the important role of PML NBs in innate immunity it is to be expected that viruses would need to overcome this defense in order to establish a productive infection. The finding, that PML assembly into NBs and recruitment of associated proteins is highly dependent upon PML sumoylation [103,104], offers a rationale for why many viruses attack the sumoylation status of PML. There is currently a large list of viral proteins that are known to disrupt NBs, and this is a particularly prominent feature of the herpesvirus family, including the BZLF protein of Epstein-Barr virus [105], the LANA2 protein of Kaposi’s sarcoma-associated herpes virus [106], human cytomegalovirus IE1 protein [107], the herpes simplex ICP0 [108], and the varicella zoster virus ORF61 protein [69]. Somewhat surprisingly the different herpes viruses utilize different mechanisms to disrupt PML sumoylation. For example, the EBV BZLF protein (also known as Zebra, Z, or Zta) is efficiently sumoylated at lysine 12 and appears to prevent PML sumoylation by out competing PML for a limited supply of free SUMO [105]. Consistent with this model, a lysine 12 mutant of BZLF is impaired in dispersing PML bodies when BZLF is limited compared to PML [109]. In contrast, as discussed in section 4.2, the herpes simplex ICP0 protein appears to be a SUMO-targeted ubiquitin ligase [67]. ICP0 contains SIM motifs that direct it to PML where the ICP0 ubiquitin ligase activity conjugates ubiquitin to PML and enhances its proteasomal degradation. Similar to ICP0, the varicella zoster virus ORF61 protein also possesses SIM motifs that are critical for interaction and dispersal of PML NBs [69]. However, the ORF61 protein alone is not sufficient to degrade PML, suggesting that its known ubiquitin ligase activity is not directly involved in modifying PML, so the precise mechanism for PML dispersal remains uncertain. Like ICP0 and ORF61, the LANA2 protein of KSHV has a SIM motif (residues 474–477) that is critical for interaction with and destruction of PML NBs in B cells [110]. Consistent with a SUMO-SIM interaction, the association of LANA2 and PML not only requires the LANA2 SIM motif, but also lysine 160 of PML which is a known sumoylation site. In addition to disruption of NBs, LANA2 causes PML levels to decrease via proteasomal degradation. These effects are associated with a LANA2 mediated increase in PML modification by SUMO2 with a concomitant increase in ubiquitination of PML. SUMO2/3 modification of PML is known to promote degradation of PML due to ubiquitination by the cellular RNF4 ubiquitin ligase [111], suggesting that this may be the pathway triggered by LANA2 expression. A subsequent study found that LANA2 is sumoylated preferentially by SUMO2 and that a sumoylation deficient mutant of LANA2 was impaired for NB disruption [106]. This observation raises the possibility that SUMO moieties on LANA2 may be interacting with SIM motifs on PML or other proteins to further facilitate the formation of complexes which lead to PML degradation. Lastly, the human cytomegalovirus (HCMV) has two early proteins, IE1 and IE2, that are both localized to PML NBs [107,108]. IE1 induces disruption of NBs without degradation of PML [112], but with the loss of sumoylated forms of PML [113,114]. However, IE1 has no SUMO protease activity in vitro, and purified IE1 does not inhibit sumoylation of PML in vitro [115]. A model proposed by the authors is that direct interaction of IE1 with PML induces disaggregation of NBs that subsequently exposes released PML to cellular SUMO proteases. In this scenario the desumoylation of PML would be indirectly mediated by this viral early protein. The situation with IE2 is even less clear. The initial observation was that IE2 expressed alone co-localized with NBs though this did not lead to NB disruption [107]. While IE2 is sumoylated [116], interacts with both Ubc9 [117] and PIAS1 [118], and has a SIM motif [119,120], a study by Sourvinos et al. found that targeting of IE2 to NBs was via association with viral genomes [121]. Consequently, while IE2 is important for viral replication there is little evidence that it plays a role in overcoming the antiviral effect of PML, and this role may be primarily the purview of IE1. In summary, members of the herpesvirus family express early proteins that target PML NBs, and most of these proteins utilize SIM motifs to facilitate their interaction with sumoylated PML. Many of these PML binding herpesvirus proteins subsequently affect the sumoylation status of PML to promote dissemble of the NBs, though they do so through diverse mechanism.PML NBs include a variety of other cellular proteins, many of which are also sumoylated such as Sp100 [122] and DAXX [123]. An early report indicated that HSV ICP0 reduced the sumoylation of Sp100 as well as PML [108], and a subsequent study found that ICP0 also induced proteasomal degradation of Sp100 [124]. While the mechanistic basis for the reduction in Sp100 sumoylation has not been explored, it is tempting to speculate that it could be through the same STUbL activity that promotes PML degradation. Likewise, early proteins from other herpesvirus family members that can modulate the sumoylation PML may have similar effects on other sumoylated proteins in the NBs. In addition to early proteins, there is now one well documented study indicating that a cytomegalovirus tegument protein, pp71, enhances sumoylation of Daxx [125]. Sumoylation of Daxx requires direct binding of pp71 but does not require Daxx to be in PML NBs. Furthermore, pp71 does not lead to a general increase in cellular sumoylation so this effect appears restricted to Daxx or a very limited number of substrates. Pp71 did not interact with Ubc9 so it is unclear if pp71 is acting as a SUMO ligase for Daxx or stimulating Daxx sumoylation through some less direct mechanism. Functionally, pp71 could still induce degradation of a Daxx SUMO site mutant so sumoylation is not required for that activity. Additionally, no effect of sumoylation on Daxx transcriptional repression was observed so the biological significance of this pp71 mediated enhancement in sumoylation remains obscure.Overall, the conclusion from this section is that viruses have evolved many strategies to dysregulate host cell proteins through modulation of their post-translational modification with the SUMO paralogs. This has been most thoroughly explored for PML and related NB proteins that have anti-viral activity, but clearly occurs for other important cellular regulatory proteins such as p53 and pRB. Given the growing appreciation of how widely sumoylation impacts many cellular processes, it is highly likely that other important viral attacks on host protein sumoylation will be identified in the future.From modest beginnings as a possibly minor modification system, sumoylation has emerged as a major regulatory network that involves hundreds to thousands of proteins through direct SUMO conjugation and SIM motif-mediated interactions with sumoylated proteins. Sumoylation is now known to be involved in major cellular events including transcriptional regulation, cell cycle control, DNA repair, RNA processing, chromatin remodeling, and nucleocytoplasmic trafficking. Viruses and certain bacterial pathogens must utilize many of these cellular processes to facilitate their own gene expression and genome replication. Consequently, it is not surprising that many pathogens have evolved proteins that are regulated by sumoylation or can usurp certain aspects of the host sumoylation system to reprogram the cellular environment to be more permissive for pathogen persistence or reproduction. The recent finding that sumoylation is involved in regulating aspects of intrinsic and innate immunity, provides another rationale for why pathogens, particularly viruses, target the sumoylation system. The ability to avoid or reduce initial host defenses is obviously critical to establishing a productive infection, and dysregulating sumoylation generally or for specific targets appears to be a common mechanism for pathogens to thwart these defenses.As presented in this review, there are now a multitude of examples of how pathogens impact the sumoylation system. These range from increasing or decreasing sumoylation of single proteins, usually via direct binding of a pathogen protein, to global increases or decreases in sumoylation. Single protein targets typically are key cellular growth regulatory factors or are critical for host immune response. Global changes in sumoylation induced by pathogens are more dramatic, but also more difficult to understand functionally as identifying the critical targets versus irrelevant proteins is challenging. Mechanistically, some pathogen proteins act by mimicking sumoylation enzymes whereas others function by binding to and altering the activity of the endogenous host sumoylation enzymes. An emerging theme is that many of these interactions are mediated or enhanced through SIM-SUMO binding, and in some cases each binding partner has one or more SIM motifs and can also be sumoylated. This raises the possibility of combinatorial interactions that may subtly influence complex affinity, stability, or composition and thus have functional consequences. The next few years are likely to reveal additional important nuances about how pathogens utilize sumoylation to their own benefit as well as identifying many exciting new targets.I gratefully acknowledge and thank all the members of my laboratory over the last 15 years who have developed, advanced, and maintained our interest in sumoylation.

Med-MDPI/biomolecules/biomolecules-02-02-00228.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).orf256 is a wheat mitochondrial gene associated with cytoplasmic male sterility (CMS) that has different organization in various species. This study exploited the orf256 gene as a mitochondrial DNA marker to study the genetic fingerprint of Triticum and Aegilops species. PCR followed by sequencing of common parts of the orf256 gene were employed to determine the fingerprint and molecular evolution of Triticum and Aegilops species. Although many primer pairs were used, two pairs of orf256 specific primers (5:-94/C: 482, 5:253/C: 482), amplified DNA fragments of 576 bp and 230 bp respectively in all species were tested. A common 500 bp of nine species of Triticum and Aegilops were aligned and showed consistent results with that obtained from other similar chloroplast or nuclear genes. Base alignment showed that there were various numbers of base substitutions in all species compared to S. cereal (Sc) (the outgroup species). Phylogenetic relationship revealed similar locations and proximity on phylogenetic trees established using plastid and nuclear genes. The results of this study open a good route to use unknown function genes of mitochondria in studying the molecular relationships and evolution of wheat and complex plant genomes.Hexaploid bread wheat (Triticum aestivum) is a member of Triticeae tribe, which also includes barley (Hordeum vulgare) and rye (Secale cereale) as well as other diploid and tetraploid wheats. Meiotic studies indicated that the general evolution of the Triticeae tribe has been defined by divergence at the diploid level from a common diploid ancestor and convergence at the polyploid level involving the diverged diploid genomes [1]. Cytological and molecular studies provided information on the identity of donors and the patterns of genome evolution of the Triticum/Aegilops species [2]. The Triticum and Aegilops genera contain 13 diploid and 18 polyploid species [3]. T. monococcum includes the cultivated form T. monococcum ssp. monococcum and the wild form, T. boeoticum. There are two tetraploid wheat species: Triticum temopheevii (AAGG genome) and Triticum turgidum (AABB genome). Finally, there are two hexaploid wheats: Triticum zhukovskyi (AAAAGG) and T. aestivum (AABBDD), including several subspecies [3,4].T. aestivum is hexaploid with a genome constitution of AABBDD, and was formed about 8,000 years ago from hybridization between T. turgidum (AABB) and A. tauschii (DD) [5,6,7]. The A genome originated with T. urartu (AA), which is closely related to T. monococcum (AA). A. speltoides, A. bicornis, A. searsii, and A. sharonensis appear to have diverged from a common ancestor at about the same time [8]. The grass family (Poaceae) diverged about 50–80 mya into the subfamilies Pooideae (tribe Triticeae containing wheat, barley, rye, Aegilops sp.), Panicoideae (tribe Maydeae containing maize), and Bambusoides (tribe Oryzeae containing rice) [9,10,11]. Maize and sorghum diverged about 16.5 mya [12]; wheat and barley diverged about 10–15 mya [9], with wheat and rye diverging about 7 mya [13]. The cytoplasms of T. aestivum, T. temopheevii, and T. turgidum originate from an ancestor like A. speltoides [14].It is suggested that the ancestor Aegilops speltoides species (S genome) was the donor of what became the B genome of the bread and durum wheats [15]. It is believed that A. speltoides is the B genome donor [16] and the maternal donor of polyploid wheats [17,18].Nuclear genes have been used in molecular phylogenetic analysis. Sequence alignment of nuclear genes encoding plastid acetyl-CoA carboxylase (ACCase) and plastid 3-phosphoglycerate kinase (PGK) were used in molecular phylogenetic analysis of the Triticum and Aegilops species. This included A, D, and S diploids and A genome polyploids using a system based on sequences of large fragments [4,19,20]. On the other hand, receptor-like kinase, Lrk, genes were used to study hexaploid wheat evolution from its progenitors, yet the study showed high conservation in gene content and organization [21]. Therefore, molecular evolution studies over a narrow time window with highly conserved genes is not an advantage because changes in DNA sequence and rearrangements are minimal.The chimeric open reading frame, orf256, is located upstream of coxI in fertile, cytoplasmic male sterile (CMS), and fertility restored (FR) mitochondria from Tt [22,23]. The 5' flanking sequence from −228 to −1 and the first 33 nucleotides of the coding sequence of the orf256 are identical to those of coxI of Ta, but the rest of the orf256 sequence is not related to that of coxI [22]. The orf256 sequence was detected in various species of wheat relatives and progenitors, but was expressed as RNA only in Tt and Aegilops speltoides.Previous studies on orf256 showed some interesting features including (1) the close evolutionary history of T. aestivum and T. temopheevii, (2) the absence of orf256 in the mitochondrial DNA of T. aestivum, its presence in T. temopheevii, and the presence of a related sequence in rice, (3) the specific transcriptional and translational characteristics of orf256 depending on the source of the nucleus and the relationship to cytoplasmic male sterility, and (4) the lack of a known function for orf256. This gives a good opportunity to follow changes in its sequence, its location, its rearrangement, and its presence or absence in Triticum and Aegilops species. These molecular characteristics of orf256 suggest that this is a rapidly changing gene and make it a suitable molecular handle for evolutionary studies. In this study, orf256 was used as a molecular tool to establish a DNA fingerprint and phylogenetic relationship among Triticum and Aegilops species and their evolutionary changes.Various specific primers (Table 1) were designed on the sequence of orf256 gene to cover different parts of the gene [24].Nucleotide sequence of primers that were used to detect the orf256 sequence using PCR. Primers with bold face font gave common positive PCR results with all tested species.Wheat seeds were surface sterilized [13,25]. About 20 g of clean wheat seeds were soaked for 20 hr in 100 mL of 10 ppm ampicillin (Sigma) solution. The antibiotic solution was drained off and 100 mL of 0.1% silver nitrate (w/v) was added. Seeds were shaken vigorously for 10 min, and the silver nitrate solution was replaced with 100 mL of 0.5% of NaCl solution. After 10 min of vigorous shaking, seeds were rinsed three times with sterile, deionized water. Sterilized seeds were spread on 0.1% water agar in a sterilized plastic container and kept in the dark for 7–10 days at room temperature. Shoots were harvested and used directly for mitochondrial isolation or freeze dried for genomic DNA isolation.10-day-old shoots were freeze dried in (Alpha 1–2 LO plus Christ, Vacuubrand, Germany). Dried shoots were ground in a coffee grinder to fine powder and used for DNA isolation.Wheat mitochondria were isolated according to Song and Hedgcoth [26]. Mitochondrial pellets were stored at −20 °C.DNA was isolated from 10-day-old wheat shoots and or wheat mitochondria of Triticum and Aegilops species (Table 2). DNA was isolated using plant DNA isolation kit (Qiagen, California, USA) following manufacturer instructions. DNA concentration was estimated and used as PCR template. DNA samples were visualized on 1–2% agarose.Triticum and Aegilops species that were used in this study.Polymerase Chain Reaction (PCR) was used to amplify diagnostic fragments of orf256 using different combinations of primers. PCR was undertaken in 50 µL total volume containing 5 µL of 10X PCR buffer, 4 µL 25 mM MgCl2, 1 µL (10 ng) of DNA, 1 µL (100 ng, 125 picomole) of each primer (forward and reverse), 1 U of Taq DNA polymerase. PCR amplification conditions were initial denaturation at 95 °C for 5 min, denaturation at 95 °C for 1 min, annealing at 50 °C for 30 sec for 35 cycles, extension at 72 °C 1 min, and final extension at 72 °C for 5 min.The common PCR fragments obtained were amplified in all Triticum and Aegilops species, especially T. turgidum, and were purified and sequenced [24]. Ten samples were sequenced for each species to eliminate the heteroplasmy possibility of mitochondrial genomes.The obtained DNA sequences of the orf256 amplified fragments were aligned using CLUSTALW [27]. The phylogenetic relationship among Triticum and Aegilops species was established using PHYLIP program on the Pasteur Institute Server [28].PCR was used to amplify DNA fragments from Triticum and Aegilops species. Using primer pair 5:-94/C: 482 (Table 1), PCR product of 576 bp was amplified (Figure 1) including Triticum turgidum, whereas using primer pair 5':253 and C: 482 resulted in the amplification of 230 bp fragment in all species tested (Figure 2). Other primer combinations (Table 1) amplified various fragments from different species except Triticum turgidum; therefore, we limited the comparison to these two fragments.PCR product (576 bp) amplified using primer pair 5': −94 and C: 482. M: 100 bp DNA ladder; 1: Tt, 2: Ttu, 3: Tm, 4: Tb, 5: Asp, 6: Ab, 7:Ase, 8: At, 9: Ash, 10:Sc, 11: Ta. Full scientific names are shown in Table 2.PCR product (230 bp) amplified using primer pair 5':253 and C: 482. 1: Tt, 2: Ttu, 3: Tm, 4: Tb, 5: Asp, 6: Ab, 7: Ase, 8: At, 9: Ash, 10:Sc, 11: Ta. Full scientific names are shown in Table 2.The large PCR fragment (576 bp) obtained with primers 5': −94 and C:482 was cleaned and sequenced from the nine Triticum and Aegilops species. The nine DNA sequences obtained were used for multiple alignment using ClustalW2 (Figure 3). Multiple alignments revealed many differences among the nine sequences used in this study. Generally, the 5' third of the aligned sequences showed the most drastic and significant differences, including cluster of deletions or single deletions in some species as well as base substitutions. The middle part of the sequence has fewer changes, whereas the 3' third is more conserved among the species under study. Alignment of 500 bp showed various numbers of base substitutions compared to Secale cereale (out of group species) (Table 3). It showed 49 base substitutions in T. temopheevii and T. turgidum; 50 base substitutions in T. monococcum; 47 base substitutions in T. boeoticum, A. speltoides, A. bicornis; 25 base substitutions in A. searsii; 22 base substitutions in A. tauschii; and 21 base substitutions in A. sharonensis (Table 3).Summary of PCR product size obtained and the number of base substitutions in Triticum and Aegilops species used in this study compared to S. cereal sequence. Multiple alignment of 500 bp of orf256 of Triticum and Aegilops species.The longest orf256 sequence obtained from T. turgidum is 576 bp using primer pairs 5: −94/C: 482. Only 500 bp were used (76 bp were eliminated) because of gaps to establish a consensus phylogenetic tree. The tree was established using PHYLIP software on the Pasteur Institute website [28]. The consensus tree was calculated by the UPGMA method. Bootstrap values were calculated as percentages of 1000 trials. Secale cereale was used as outgroup species. Six data sets were included in the calculation of the consensus tree using the nine species. Set one included species A. speltoides, A. sharonesis, A. bicornis. A. searsii, T. temopheevii, T. turgidum. Set two included species A. tauschii, T. boeoticum, and T. monococcum. Set three included species T. temopheevii and T. turgidum. Set four included species T. boeoticum and T. monococcum. Set five included species A. speltoides, A. sharonesis, A. bicornis, A. searsii. Set six included species A. sharonesis, A. bicornis, and A. searsii.The consensus tree was established by making one thousand trials (Figure 4). The tree has two clades, A and B. Clade A which has the same location (bootstrap) in one thousand trials contains two branches, C and D. Branch C contains one species; A. speltoides and sub-branch E which contains three species A. searsii, A. bicornis, A. sharonesis. Branch D contains two species; T. temopheevii and T. turgidum. Clade B that has the same location (bootstrap) in 975 trials has one species; A. tauschii and one sub-branch F which included two species; T. monococcum, T. boeoticum (Figure 4).Consensus phylogenetic tree of Triticum and Aegilops species based on the common 500 bp of orf256 sequence and one thousand trials. Bootstraps (the numbers on the branches) indicate the number of times the partition of the species into the two sets, which are separated by that branch, occurred among the trees, out of 999.99 trees.DNA distances among studied species were calculated using DNAbars software on the Pasteur Institute website [28,Table 4]. T. temopheevii and T. turgidum were the closest species with DNA distance about 0.2. Also, minimum distances occurred between A. serseaii and A. bicornis, A. bicornis and A. sharonesis, T. monococcum and T. boeoticum with distances of 0.2, 0.2, and 0.4 respectively. The highest distance was between T. temopheevii and A. speltoides to A. tauschii with DNA distance of 33.78. S. cereale is the outgroup species. T. temopheevii and A. speltoides were the most separated species of S. cereal with DNA distance of 32.54, although they were located on different sub-branches of the consensus phylogenetic tree, whereas T. boeoticum and A. taushii were the closest species to S. cereale with DNA distance of 4.11, although they were located on different sub-branches of the consensus phylogenetic tree.DNA distances among Triticum and Aegilops species (1000 trials). Distance Matrix was calculated using the Jukes-Cantor correction method. Base positions 123 in the codon and gap weighting 0.0 were used.T. turgidum gave negative results with other primer pair combinations (Table 1). Results previously obtained from other studies also suggested that this species has a partial orf256 sequence [13]. Common 500 bp were used to study the similarity among the nine different Triticum and Aegilops species. The phylogenetic relationship among studied species, although different, was still consistent with results obtained from previous studies which used other plastid and nuclear genes. In a study using 3 phosphoglycerate kinase (pgk-1) gene of Triticum and Aegilops, results revealed that some species showed similarity of location on the phylogenetic tree. T. temopheevii and T. turgidum along with T. aestivum (not included in this study since it does not have orf256 in its mitochondria DNA) showed closer location on the phylogenetic tree using acetyl-coA carboxylase (ACC-1) and 3-phosphoglycerate kinase (PGK-1) [4,29]. They were mapped on one sub-branch (sub-branch D, Figure 4). A. speltoides showed independent location from other Aegilops or Triticum species using the same genes (PGK-1) [4,29]. The present study showed a similar pattern because it is positioned on a separate branch with bootstrap of 55 (Figure 4). A. searsii, A. bicornis, and A. sharonesis showed a closer location in the present study. They were located on sub-branch E with bootstrap 52 (Figure 4). They showed similar relatedness [4] using 3-phosphoglycerate kinase (PGK-1) while they did not show this close proximity on the phylogenetic tree using acetyl-coA carboxylase (ACC-1) and 3-phosphoglycerate kinase genes (PGK-1) [29]. Triticum monococcum and T. boeoticum also showed close proximity location on phylogenetic trees using 3-phosphoglycerate kinase (PGK-1) [29], yet they did not show this close proximity on phylogenetic trees using the same gene ((PGK-1) [4]. T. taushii is located on an independent branch in this study (Figure 4), but in other studies, it showed close proximity with T. aestivum using ACC-1 and PGK-1 genes [4,29]. A partial sequence of WAG-2 gene was used to study the molecular evolution of wheat and its relatives. Marked variations were reported in single nucleotide polymorphisms (SNIPS) and indel numbers. Similar topology of phylogenetic trees using WGA-2 gene and the orf256 genes were obtained. For example, A. tauchii was located on one separate clade (clade III) using the WGA-2 gene. A similar location on the phylogenetic tree was obtained using orf256 gene (30). A. speltoides and T. turgidum had close phylogenetic topology on trees established using WGA-2 and orf256 genes. This supports the established idea that A. taushii (DD) is the source of D genome of Triticum aestivum (AABBDD).From the data obtained from this study and similarities of our results with results obtained using other nuclear and plastid genes, it can be concluded that the orf256 represents a suitable molecular tool to study the relationship among Triticum and Aegilops species. Also, this introduces one more mitochondrial gene to study bioinformatic relationships among species with complex genomes which could lead to resolving their evolution at the molecular level. Orf256 and other genes could be used in monitoring gene transfer among cellular organelles, especially the nucleus and mitochondria, and tracking their evolutionary changes.Wheat mitochondrial gene orf256 was used to study the phylogenetic and the evolutionary relationship among Triticum and Aegilops species. The results obtained were consistent with those obtained using plastid and nuclear genes. Also, the phylogenetic tree obtained from this study gave similar locations to many Triticum and Aegilops species which used plastid and nuclear genes. Data conclude that the orf256 gene of wheat mitochondrial DNA is a good molecular tool to study bioinformatic analysis of Triticum and Aegilops genomes.The work was supported by Taif University; grant number 1-432-748.

Med-MDPI/biomolecules/biomolecules-02-02-00240.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).SUMO (small ubiquitin-related modifier) conjugation is a reversible three-step process of protein post-translational modifications mediating protein-protein interactions, subcellular compartmentalization and regulation of transcriptional events. Among divergent transcription factors regulated by SUMOylation and deSUMOylation, the androgen receptor (AR) is of exceptional significance, given its established role in prostate carcinogenesis. The enzymes of the SUMO pathway can have diverse effects on AR transcriptional activity, either via direct modification of the AR or through modification of AR co-regulators. Accumulating in vitro and in vivo evidence implicates the SUMO pathway in AR-dependent signaling. Prostate cancer cell proliferation and hypoxia-induced angiogenesis are also regulated by the SUMO pathway, through an AR-independent mechanism. Thus, an important role has been revealed for members of the SUMO pathway in prostate cancer (PCa) development and progression, offering new therapeutic targets.Prostate cancer (PCa) is the most common cancer and the second leading cause of death from cancer in males in most western countries. The androgen receptor (AR) is one of the most important factors in PCa development [1]. The transcriptional AR activity modulated by positive or negative regulators plays a critical role in controlling the growth and survival of PCa cells. Enhanced AR activity is essential for cancer cell growth because PCa, in most cases, will regress in response to androgen removal therapy [1,2]. Despite this exquisite sensitivity to androgen deprivation therapy which renders it the most endocrine-sensitive solid neoplasm, advanced disease eventually progresses to castration-resistant PCa. However, recent evidence has shown that cancer progression at this stage is, again, often mediated by AR signaling, so that subsequent AR targeting may further contribute to disease control and, eventually, survival improvement [3]. Given the central role of AR in PCa development and progression, regulators of AR transcriptional activity may have significant effects on these processes.SUMOylation represents one such important post-translational modification system that regulates the activity of many transcriptional regulators [4]. The continually growing list of SUMO modified proteins comprises various nuclear receptors, including AR as well as transcriptional activators, coactivators, and corepressors [5]. The biological functions of SUMOylation include protein subcellular translocation, subnuclear structure formation, modulation of transcriptional activity, regulation of protein-protein interactions, regulation of cellular metabolism in physiology and disease as well as regulation of several intracellular signaling pathways [6]. SUMOylation depends upon the activity of small ubiquitin-related modifier (SUMO), a protein moiety that is conjugated to a specific lysine residue on target proteins [4]. Three SUMO family members exist, SUMO-1/Smt3C, SUMO-2/Smt3A, and SUMO-3/Smt3B, and all are ubiquitously expressed in mammals [7,8]. At the amino acid level, SUMO-2 and SUMO-3 are 87% identical but only 50% identical to SUMO-1 [7]. Although they exhibit low homology in amino acid sequence, SUMO-1 and ubiquitin are structurally related and share significant similarity in secondary and tertiary structures in higher eukaryotes [9]. Therefore, it is not surprising that the processes of SUMOylation and ubiquitination are mechanistically similar [5]. Like ubiquitination, the conjugation of SUMO is mediated by a series of enzymatic reactions catalyzed by E1, E2, and E3 enzymes that are distinct from those enzymes that catalyze ubiquitination [7]. The SUMO E1 enzymes SAE1 (SUMO-activating enzyme) and SAE2 activate SUMO and transfer it to the E2 enzyme Ubc9, which then directs the conjugated SUMO to its target substrates [7]. Ubc9 then catalyzes formation of an isopeptide bond between SUMO-1 and the ε-amino group of lysine in the target protein. The specificity for the target protein of SUMOylation is thought to reside with Ubc9 itself. SUMOylation typically occurs at a specific sequence, ψKXE, on the target protein where ψ is a hydrophobic residue and X is any amino acid. In vitro evidence has indicated that Ubc9 is sufficient for binding to the SUMO acceptor site and efficiently transferring SUMO to selected targets [7]. However, recent evidence shows that a specific E3 ligase might be required for efficient SUMOylation in vivo. Three classes of proteins have been identified to have SUMO E3 ligase activity: the protein inhibitor of activated STAT (PIAS) family proteins [10,11], the polycomb protein Pc2 [12], and RanBP2 (Ranbinding protein 2) [13]. The PIAS proteins are reported to act as SUMO-E3 ligases for the SUMO-1 conjugation to AR in vivo and in vitro [14,15].SUMOylation is readily reversible in the cell because the isopeptide bond created between the C-terminal glycine in SUMO and the epsilon amino group in the acceptor lysine can be cleaved by SUMO proteases, which are also termed SENPs (SENtrin specific Proteases) or deSUMOylases. There are six SENP enzymes in mammals, each containing a highly conserved 200 amino acid catalytic domain that mediates deSUMOylation [16]. The amino and carboxyl terminal domains vary between SENPs and play roles in subcellular localization and perhaps substrate recognition [17].SUMOylation and deSUMOylation processes are responsible for transcriptional regulation of various initiated signals, including androgen-mediated transcription. However, there is an established correlation between enhanced androgen-dependence and prostate carcinogenesis and the most striking proof for this is the use of expression of the AR-regulated prostate-specific antigen (PSA) gene as a biologic marker for the diagnosis and treatment of PCa [18,19,20]. The multiple interconnections between members of the SUMO pathway, AR-mediated and AR-independent propagation of PCa are reviewed here, with implications for their ultimate clinical significance.The AR can be modified by SUMOylation, preferentially by SUMO-1. Two major SUMOylation sites (K386 and K520) have been identified within the AR [21]. The biological effect of AR SUMOylation was explored by mutating one (or both) SUMOylation sites and measuring androgen-induced transcription. AR containing the Lys 386 to Arg substitution either alone, or together with a Lys 520 to Arg substitution, showed a 2–3-fold enhancement of androgen-dependent transcription on promoters containing multiple androgen response elements (AREs) [21,22,23]. The activity of AR containing the Lys 520 to Arg substitution alone was similar to wild type AR. The data suggest that SUMOylation of AR primarily at Lys 386 reduces the transcriptional activity of AR. The underlying mechanism for this effect has not been defined [17].Besides SUMOylation, AR is also subjected to phosphorylation and acetylation. Ubc9, the SUMO E2 enzyme, binds the AR within the hinge region [24] that includes the site of direct acetylation of lysine residues at a conserved KLKK motif. However, the SUMOylation of the AR was unaffected by the mutation of the AR acetylation site in vitro, probably suggesting that Ubc9 stimulation of AR transcriptional activity may be independent of its ability to catalyze SUMO-1 conjugation [25,26].AR activity is also regulated by the PIAS family proteins [27]. The PIAS family is composed of several homologous proteins, including PIAS1, PIAS3, PIASxα, PIASxβ, and PIASy. AR-dependent transcription is repressed by PIAS1 and PIASxα in the presence of exogenous SUMO-1 and PIAS RING finger-like domain and enhanced in the absence of SUMOylation. PIAS3 inhibits AR transactivation in LNCaP and HeLa cells but enhances AR activity in HepG2 and AR-overexpressing LNCaP cells [15,28,29,30]. Repression of AR-dependent transcription by PIAS1 and PIASxα is dependent on the ectopic expression of SUMO-1, indicating that the SUMOylation of AR is crucial for AR-dependent transcriptional repression [14]. PIASy-mediated enhancement of AR-dependent transcriptional repression, however, seems to be independent of SUMOylation [31,32].A novel PIAS-like protein, named hZimp10, was identified as an additional AR co-regulator. hZimp10 co-localized with AR and SUMO-1 at replication foci throughout the S phase, and it was capable of enhancing SUMOylation of AR in vivo. Studies using SUMOylation-deficient AR mutants suggested that the augmentation of AR activity by hZimp10 is dependent on the AR SUMOylation. A link was revealed between hZimp10-mediated enhancement of AR SUMOylation and modulation of AR-mediated transcription, although it remains uncertain whether hZimp10 acts directly as an E3 ligase [32,33].The SUMOylation acceptor sites of the AR have also been mapped to a previously identified inhibitory motif called the synergy control (SC) motif [34]. The SC motif has also been found within the negative regulatory regions of many different transcription factors [34] and their SUMO acceptor sites have been defined within these SC motifs [26]. Mutation of these SUMO acceptor sites in these factors leads to an increase in the transcriptional activation. These findings suggested that SUMOylation-dependent repression is a common regulatory mechanism in transcriptional control [26].But what is the molecular mechanism through which SUMOylation regulates AR transcriptional potential? It was shown that a conserved motif localized at the hinge region of the AR is responsible for acetylation and that the extent of AR SUMOylation is independent of its acetylation status [35], excluding the possibility of SUMOylation antagonizing the same lysine residues for acetylation. Furthermore, the same research group reported that PIAS family proteins PIAS1 and PIASxα could enhance AR SUMOylation but could not alter the subnuclear localization of the AR [35], indicating that SUMOylation of the AR is irrelevant to AR subnuclear distribution. In addition, it was also demonstrated that SUMOylation does not affect its DNA-binding capability [23].A mechanistic insight underlying SUMO-dependent transcriptional repression of the AR was provided by the implication of Daxx, initially identified as a cytoplasmic signaling molecule linking Fas receptor to Jun N-terminal kinase in Fas-mediated apoptosis [36], in this process. Daxx has recently been demonstrated to act within the nucleus to regulate gene expression, inhibiting the activity of several transcription factors. The finding that Daxx selectively binds to SUMOylated AR in vitro and in vivo, combined with evidence of suppression of AR transcriptional activity by wild type, but not mutant SUMO, strongly suggest that AR SUMOylation is involved in regulating Daxx-dependent transcriptional repression [37]. Accordingly, mutation of SUMO-conjugated sites in AR resulted in a loss of Daxx interaction and an increase of AR transcriptional activity [26].The most recent data on SUMO-mediated repression of AR transcriptional activity involves histone deacetylase 4 (HDAC4) binding. Intriguingly, this was found to depend on SUMOylation, rather than deacetylation, of the AR. HDAC4 increases the level of AR SUMOylation in both whole-cell and cell-free assay systems, raising the possibility that the deacetylase may act as an E3 ligase for AR SUMOylation. Knock-down of HDAC4 increases the activity of endogenous AR and androgen induction of PSA expression and PCa cell growth, which is associated with decreased SUMOylation of the receptor. Overall, HDAC4 has emerged as a new positive regulator of AR-mediated transcription, revealing a deacetylase-independent, SUMOylation-dependent mechanism of HDAC action in PCa cells [38].Further enhancing the diversity of SUMOylation-induced effects on AR signaling, and being at odds with the previously observed negative effect of SUMO-1 conjugation on AR-initiated transcription, SUMO-3 may have a negative or strongly positive effect on AR, depending on the type of PCa cells. In primary prostate epithelial cells, PrEC, and the PCa cells, PC-3, SUMO-3 has a weak negative effect on AR transcriptional activity. In contrast, SUMO-3 and its close relative SUMO-2 strongly enhance transactivation by endogenous AR in LNCaP cells. This positive effect is observed in both androgen-dependent and androgen-independent LNCaP cells. Mutational analysis of AR and SUMO-3 demonstrated that the SUMO-3-mediated transcriptional activation does not depend on either the previously-identified SUMOylation sites of AR or the covalent conjugation of SUMO-3 to target proteins [15]. These results suggest a novel mechanism for elevating AR activity through the switch of SUMO-3 from being a weak negative regulator in normal prostate cells to a strong positive regulator in PCa cells. This further implies that SUMO-2 and SUMO-3 stimulate the proliferation of PCa cells that is independent of AR SUMOylation, making it mechanistically distinct from the SUMO-1-dependent repression of AR activity. SUMO-3 may thus have an important role in promoting prostate carcinogenesis [15].In addition, strong clinical data regarding immunohistochemical expression of SUMO pathway member proteins in human specimens further support a critical role for protein SUMOylation in PCa development and progression. Endogenous basal PIAS3 expression was reported to localize in the nucleus in a majority of epithelial and endothelial cells. Increased PIAS3 expression was observed in 100/103 samples examined in a variety of human cancers including prostate, lung, breast, colorectal, and brain tumors. Differential PIAS3 expression and the specific patterns might therefore be useful as a molecular tumor marker [39].Likewise, it was found that in primary PCa, Ubc9 expression is increased compared with normal tissue, whereas in metastatic prostate tissues, it is decreased in comparison with their corresponding normal and primary adenocarcinoma tissues. Ubc9 expression correlates negatively with tumor Gleason score. The authors conclude that expression of Ubc9 is directly associated with progression of PCa, since it was high in prostatic intraepithelial neoplasia (PIN) cells and even higher in primary adenocarcinomas [40]. To explain this, the authors postulate that given that the N-terminal half of the AR hinge region is essential for the interaction with Ubc9, and AR transcription is enhanced by coexpression with Ubc9 [24], it is possible that Ubc9 contributes to PCa progression by modulating AR stability and/or activation. A correlation between level of Ubc9 expression and the presence/extent of host-immune infiltrate in primary PCa was another key finding of this Ubc9-based TMA analysis [40].AR SUMOylation is a dynamic process and is reversed by SENP1, which promotes AR-dependent transcription in PCa cells. Nucleocytoplasmic shuttling is an important parameter that should be always considered in SUMOylation and deSUMOylation events related to AR. Nuclear import of AR is not sufficient for SUMOylation, because constitutively nuclear apo-ARs or antagonist-bound ARs are only very weakly modified by SUMO-1 in comparison with agonist-bound ARs. Of the SUMO-specific proteases (SENP)-1, -2, -3, -5, and -6, only SENP1 and SENP2 are efficient in cleaving AR-SUMO-1 conjugates in intact cells and in vitro. Both SENP1 and -2 are nuclear and found at sites proximal to AR. Their expression promotes AR-dependent transcription, but in a promoter-selective fashion. SENP1 and -2 stimulated the activity of holo-AR on compound ARE-containing promoters. The effects of SENP1 and -2 on AR-dependent transcription were dependent on catalytic activity and required intact SUMO acceptor sites in AR, indicating that their coactivating effects are mainly due to their direct isopeptidase activity on holo-AR. In PCa cells, ectopic expression of SENP1, but not that of SENP2, increased the transcription activity of endogenous AR. Silencing of SENP1 attenuated the expression of several AR target genes and blunted androgen-stimulated growth of LNCaP cells. These results indicate that SENP1 reverses AR SUMOylation and helps fine-tune the cellular responses to androgens in a target promoter-selective manner [22]. Intriguingly, SENP1’s ability to enhance AR-dependent transcription is not mediated through deSUMOylation of AR, but rather through its ability to deconjugate histone deacetylase 1 (HDAC1). Thus, SENP1-dependent deSUMOylation of HDAC1 reduces its deacetylase activity and repressive activity to AR-dependent transcription [41,42,43].SENP1 levels were enhanced in cells treated with IL-6, which is a positive regulator of the activation of AR-dependent transcription through the activation of the MAPK and JAK/STAT pathways [44,45]. The combination of both a synthetic androgen (R1881) and IL-6 profoundly enhanced SENP1 expression by more than seven-fold, compared to either compound alone, and this was further supported by an enhanced production of PSA. Conversely, inhibition of SENP1 expression by siRNA reduces androgen-induced PSA production in LNCaP cells [46]. Thus, it is likely that SENP1 induction is essential for androgens and IL-6 to induce PSA secretion.These in vitro findings translate to an altered clinical phenotype, depending on the levels of SENP1 expression, as evidenced by induction of PIN-like structure formation in SENP1 transgenic mice that were older than 4 months [2]. Elevated SENP1 expression has also been detected in human PCa at the PIN stage [20]. SENP1 messenger RNA was increased in 29 of 43 cases of high grade PIN (67%). Similarly, SENP1 expression was increased in 26 of 43 PCa samples (60%). Thus, SENP1 expression is preferentially increased during the development of PCa in the majority of cases. Collectively, these studies indicate that overexpression of SENP1 is likely to play a significant role in PCa development [20].Androgen treatment of PCa cells stimulates AR SUMOylation within 15 minutes and reaches a maximum level by 1 hour [22]. The similarity between these kinetics and androgen-induced nuclear import of AR, and the fact that SUMOylation enzymes E1 and E2 are highly concentrated in the nucleus, suggested that AR SUMOylation might occur only in the nucleus after nuclear import. However, this is not the case, based on the observation that AR can be SUMOylated in either the cytoplasm or nucleus, with similar efficiencies [22]. The androgen-mediated induction of AR SUMOylation could therefore reflect a conformational change that enhances accessibility of the modification site or the interaction of AR with SUMOylation enzymes. Under steady state conditions, only a small fraction of the total AR pool is conjugated with SUMO [22]. Investigation of the protein expression profile of mouse prostate by the use of subcellular fractionation, 1-DE (one-dimensional gel electrophoresis) protein separation and mass spectrometry, also demonstrated that both free SUMO-2/3 and SUMO-1 are particularly abundant in the prostate and their levels are subject to tight control by the androgen 5R-dihydrotestosterone (DHT) [47].DeSUMOylation is also subject to tight control by androgens, as exposure of LNCaP cells to androgen treatment enhances SENP1 transcription. This androgen-mediated augmentation of SENP1 is an AR-dependent event, as reflected by the absence of the effect both when the androgen receptor antagonist bicalutamide is present, and in AR-negative PCa PC-3 cells. A specific ARE on the promoter of the SENP1 gene is required for SENP1 induction by androgens [48]. Thus, elevation in SENP1 mRNA levels in PCa cells is selective and requires the activation of the AR. Upon activation, AR binds this ARE and activates SENP1 transcription. Therefore, SENP1 regulates androgen-AR signaling through a positive feedback mechanism, as follows: The androgen-activated AR binds a specific response element located proximal to the SENP1 promoter. SENP1 promoter activity is enhanced by this activated AR, and thereby, SENP1 mRNA levels are significantly elevated. SENP1 up-regulation completes the positive feedback loop by potentiating AR-dependent transcription and cell proliferation [49]. It is intriguing to speculate that androgen ablation therapy is initially effective in treating PCa due to its ability to decrease SENP1 expression. It could therefore be hypothesized that therapeutic agents designed to selectively lower SENP1 levels might be more effective than androgen ablation therapy in the treatment of advanced PCa. Like androgen ablation, reduction of elevated SENP1 levels lowers AR activity [49,50] and PCa cell proliferation. However, unlike androgen ablation, selective down-regulation of SENP1 could modulate these two events without depleting prostate epithelial cells of androgen. By not altering androgen levels, this SENP1-targeting agent would not prompt the development of androgen-dependent cancer cells [49].Emerging data implicate SUMOylation in regulating AR activity by modulating co-regulator activity. Two such factors include p68 (DDX5) and p72 (DDX17), which are members of the DEAD-box RNA helicase family that can unwind double stranded RNA and contribute to the remodeling of ribonucleoprotein complexes. These activities of p68/p72 are required for efficient RNA splicing and microRNA processing. p68/p72 perform functions that are independent of their helicase activity. This is especially common in their role as transcriptional coactivators, where p68/p72 regulate various transcription factors, including AR [50]. One conserved lysine residue in the N-terminus of p68 and p72 RNA helicase (K53 and K50, respectively) has been identified to be a target for SUMOylation [51,52,53]. SUMOylation of p68/p72 has a variety of consequences: It doubles protein stability of p68, whereas it only slightly increases the protein half-life of p72. This may translate into a significant clinical effect, as p68 was found to be overexpressed in prostate tumors and is capable of coactivating AR [54], the key player in prostate tumorigenesis, and could thereby facilitate prostate tumorigenesis [50].Alien thyroid hormone receptor-interacting protein (TRIP15)/subunit 2 of the signalosome complex (CSN2) binds to the amino-terminus of AR with the receptor SUMOylation sites being involved. Both the AR N terminus with the first 328 aa and the SUMOylation sites are involved in the regulation of the Alien-AR interaction, as deletion of the inhibitory domain or mutating the SUMOylation sites of AR leads to an increase of AR-mediated transactivation. Thus, SUMOylation might have a possible role for binding to Alien, although Alien binds non-SUMOylated AR at least in vitro, suggesting the notion that SUMOylation per se is not involved for binding of Alien to AR [55]. Alien has characteristics of a co-repressor as it is recruited to AR in the presence of the AR antagonist cyproterone acetate (CPA). Furthermore, cellular localization of Alien is changed towards a predominant nuclear localization upon treatment of PCa cells with CPA. Notably, stable expression of Alien in LNCaP cells inhibits both endogenous PSA expression and proliferation of these cells in the presence of CPA, but not in the presence of an AR agonist. These findings underline the importance of co-repressors for inhibition of PCa cell growth by androgen antagonists [55].A multi-C2H2-type zinc finger protein (ZNF), ZNF451, was found to interact with both the SUMO E2 conjugase Ubc9 and SUMO-1, SUMO-2 isoforms. Further, ZNF451 interacts in a SUMO-1-enhanced fashion with AR. Although it lacks an intrinsic transcription activation function, ablation of endogenous ZNF451 in LNCaP PC cells significantly decreases expression of several AR target genes. Thus, ZNF451 acts as a transcriptional co-regulator. Both ZNF451 and AR are expressed in PCa cells. SUMOylation status of ZNF451 determines its subnuclear localization, suggesting that movable ZNF451 may influence transcription by modulating SUMO-directed trafficking of co-activator or co-repressor proteins between nucleoplasm and nuclear bodies (NBs). It is of note that AR itself evades NBs, and therefore, a “courier,” such as ZNF451, is likely to be involved in the release of co-activators from nuclear storage sites. Furthermore, ZNF451 may also tether transcriptional co-repressors, such as HDAC4, and thereby relieve repression: that is, de-repress AR activity. These data characterize ZNF451 as a novel SUMO-associated co-regulator protein that regulates androgen signaling [56].Pontin, a component of chromatin-remodeling complexes, is also SUMO-modified, and SUMOylation of pontin is an active control mechanism for the transcriptional regulation of pontin on AR target genes in PCa cells. Biochemical purification of pontin-containing complexes revealed the presence of the Ubc9 SUMO-conjugating enzyme that underlies its function as an activator. Intriguingly, DHT treatment significantly increased the SUMOylation of pontin, and SUMOylated pontin showed further activation of transcription of a subset of nuclear receptor-dependent target genes and led to an increase in proliferation and growth of PCa cells. These data clearly define a functional model and provide another link between SUMO modification and PCa progression. It would therefore be tempting to explore the possibility that malignant progression of PCa cells might be dependent on SUMOylation of pontin [57].To summarize, multiple levels of regulation of AR-dependent transcription by the SUMO pathway have been characterized in recent years, involving either direct interaction of SUMO pathway member proteins with AR or indirect control of AR-associated co-regulators (Figure 1). Simplified schema of interconnections between small ubiquitin-related modifier (SUMO) pathway proteins and other factors regulating androgen receptor (AR)-dependent transcription.SUMOylation regulates the activity of prostate carcinogenesis through a variety of mechanisms, some of which are unrelated to AR signaling. For example, the Sp family of transcription factors, which contains three proteins, Sp3, M1 and M2, with differing capacities to stimulate or repress transcription. This was demonstrated in a series of experiments monitoring the activity of the natural promoter of PSA in its natural cellular milieu (prostate epithelial cells) in conjunction with expression vectors encoding wild-type and point-mutated Sp proteins. A yeast two-hybrid screen to identify Sp3-binding proteins resulted in the identification of Ubc9 as an M2-binding protein, and sequence analyses identified consensus SUMOylation motifs within several Sp family members. Western blots probed with anti-Sp3 detected a high molecular weight Sp3 isoform that is stabilized by a SUMO-1 hydrolase inhibitor, and this protein is also bound by anti-SUMO-1 antiserum. Transient transfection assays with epitope-tagged-SUMO-1 and GFP-SUMO-1 fusion proteins confirmed that Sp3, M1 and M2 proteins are SUMOylated in vivo. Substitution of arginine for lysine at one putative site of SUMOylation, lysine551, blocked SUMOylation of all Sp3 isoforms in vivo and led to a marginal increase in Sp3-mediated transactivation in insect and mammalian cells. In contrast, introduction of this amino acid substitution within M1 converted it into a potent transcriptional activator. Therefore, Sp3 isoforms are SUMOylated in vivo and this post-translational modification plays an important role in the regulation of Sp3-mediated transcription. Although SUMOylation appears to only modestly reduce trans-activation by Sp3, SUMOylation of M1, and presumably M2, is required for a mechanism of transcriptional repression that is insensitive to HDAC inhibition (by trichostatin A). Regardless of the precise mechanism by which SUMOylation converts M1 and M2 into transcriptional repressors, it is evident that Sp3 isoforms are differentially regulated by SUMOylation, and the abundance of SUMOylated M1/M2 proteins determines overall levels of Sp3-mediated transcription [58].Provided SENP1 can strongly increase AR transcriptional activity, an effect of SENP1 on proliferation was demonstrated in androgen-dependent PCa, as endogenous SENP1 silencing by SENP1 siRNA in LNCaP cells restricted cell growth. However, similar results were observed in PC-3, an androgen-independent PCa cell line, suggesting that SENP1 might play a role in regulating cell proliferation through an AR-independent pathway. Also, the number of SENP1-silenced PC-3 cells in the G1 phase was significantly increased but was decreased in the S and G2/M phases, suggesting that SENP1 may promote G1–S phase transition in PC-3 cells. This slowing of the proliferation of SENP1-silenced PC-3 cells may thus be due to an alteration in G1–S phase progression. Silencing SENP1 expression decreased the expression of cyclin D1, but not cyclin E, in PC-3 cells. Conversely, stably transfected SENP1 in LNCaP cells enhanced cyclin D1 expression and cell proliferation. Interestingly, cell proliferation is dependent on cyclin D1 expression induced by SENP1. Thus, the regulation of cyclin D1 expression by SENP1 is another means through which PCa cell growth is regulated. The induction of cyclin D1 expression by SENP1 depended on its catalytic activity, as the mutation of SENP1 catalytic domain disrupted its activity on cyclin D1 transcription and cell proliferation [2].Hypoxia-inducible factor (HIF) 1α stabilization, enhanced vascular endothelial growth factor (VEGF) production, and angiogenesis is regulated by SENP1 during prostate carcinogenesis. In the absence of SENP1, HIF1α is actively SUMOylated and subsequently degraded under hypoxic conditions [1,2]. SENP1 alters VEGF levels by directly regulating HIF1α stability during fetal development [1,2]. Furthermore, SENP1 transgenic mice exhibited increased expression of HIF1α with progression of the dysplasia. The enhanced HIF1α stability in the SENP1 overexpressing mice produced elevated VEGF expression. Consequently, it is not surprising that angiogenesis was readily observed in these SENP1 transgenic mice compared with age-matched wild-type mice. In two lines of SENP1 transgenic mice, the hyperplasia further progressed to develop PIN [2]. Also, high-grade PIN was observed in the transgenic mouse line with the higher level of the SENP1 transgene. Enhanced proliferation of prostate epithelia was observed in the SENP1-overexpressing mice, and concurrently, pro-oncogenic factors: specifically the androgen receptor (AR) and cyclin D1, were elevated [2]. Thus, SENP1 participates in the development of prostate neoplasia via facilitating both pro-growth and angiogenic pathways. In the well-defined PCa mouse model TRAMP, high-grade PIN is accompanied by an increase in HIF1α levels, which is, in turn, required for initiation of the angiogenic switch [59]. Consistently, SENP1 transgenic mice exhibit an induction of HIF1α with the initial onset of PIN (or low-grade PIN) at 4 months of age, suggesting that SENP1 regulation of HIF1α occurs early in prostate pathogenesis. Also, SENP1 overexpression initiates the HIF1α pathway in the prostates of transgenic mice as indicated by the elevation of HIF1-regulated VEGF protein levels at low-grade PIN (4 months of age) and even more dramatic VEGF elevation at 12 months of age, when the SENP1 transgenic mice concurrently exhibit an increase in microvessel density [2]. Given that HIF1α is currently being evaluated as a prognostic marker for PCa aggressiveness, it is intriguing to speculate that because SENP1 modulates HIF1, SENP1 may be an equally good marker. This fosters the need for more comprehensive studies to evaluate the potential of SENP1 as a prognostic marker in human PCa [2]. Intriguingly, HIF1 has also been demonstrated to be SUMO-1 modified even in the presence of SENP-1. Work by Berta et al., 2007 reported that when HIF1α is conjugated to SUMO, its transcriptional activity is decreased and that this is not mediated by a change in the protein's half-life [60]. Most importantly, HIF1α also regulates SENP1 as a transcriptional factor, thus contributing to formation of a positive feedback loop which is important in VEGF production, essential for angiogenesis in endothelial cells [61].Assessment of tissue from human PCa patients indicates elevated mRNA levels of both SENP1 and the SUMO-2/3 deconjugating enzyme, SENP3. SENP3 regulates the transcriptional activity of HIF1α via deSUMOylation of the coregulatory protein p300. Through this mechanism, overexpression of SENP3 facilitates the expression of HIF1α-regulated VEGF, which is critical for vascular development. Induction of SENP3 can be mediated via reactive oxygen species (ROS), as the latter can inhibit the ubiquitin-proteosomal mediated degradation of SENP3 to increase SENP3 protein levels [62,63]. Hence, the induction of SENP3 directly contributes to cancer progression, providing another attractive alternative target for therapy, possibly most notably in cancers with increased ROS levels, including PCa [63,64]. The aforementioned mechanisms of SUMO pathway-mediated AR-independent induction of PCa growth and angiogenic signaling are depicted in Figure 2. Progression of PCa to androgen-independent growth involves many oncogenic signaling pathways, some of which are regulated by SUMOylation, as described above. Furthermore, since aberrant AR signaling (regulation of which is intimately associated with the SUMO pathway) is an important driving force of hormone resistance in PCa, the impact of the SUMO pathway may play a role in the development of androgen-refractory PCa. Further data from testing of androgen-refractory PCa cells and tissues are needed to confirm this hypothesis, however.AR-independent promotion of prostate cancer (PCa) cell proliferation and angiogenesis by the SUMO pathway.Inevitably, many questions remain unanswered with respect to the various roles played by SUMO pathway members in PCa development and progression. For instance, does the expression of one or multiple SUMOs and SENPs fluctuate with the onset of a given carcinoma, or do specific combinations of SUMOs and SENPs alterations, regarding their function and/or localization, drive the initiation and progression of cancer? It is also difficult to tell yet whether changes in the level of SUMOylation of given transcriptional regulators persist throughout the different stages of PCa. It is possible that both SUMO conjugation and deconjugation are critical for PCa evolution. For example, one arm may be favored at an early stage to initiate tumor growth, while the other arm may be favored at more advanced stages for cancer metastasis [63].The potential use of elevated SENP1 or/and SENP3 levels in the prostate gland as prognostic markers could possibly identify individuals with an increased risk of developing PCa. In addition, assessment of their levels in PCa samples might serve as diagnostic markers for enhanced angiogenesis and, consequently, more aggressive carcinomas. Additional studies using tissue arrays to validate the use of either SENP1 or SENP3 as biomarkers for PCa are warranted.Attempts to target the deSUMOylation process by developing specific inhibitors for the family of SENPs are being made, although no pharmacologic agent is currently available. The first report of a successful design of SENP1 inhibitors came only recently, and involves a series of SENP1 inhibitors based on a benzodiazepine scaffold which showed inhibitory activity as high as IC50 = 9.2 μM. The compounds with the best SENP1 inhibitory activity were tested against PCa cells (PC-3) to evaluate their ability to inhibit cancer cell growth in vitro, with IC50 values as low as 13.0 μM detected [65]. Hopefully, progression of a selective inhibitor of either SENP1 or SENP3 to in vivo testing as well as in the setting of a dedicated clinical trial might lead to an effective therapeutic approach to restoring balance to the aberrant SUMO pathway of PCa patients.

Med-MDPI/biomolecules/biomolecules-02-02-00256.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Post-translational modifications of proteins are essential for cell function. Covalent modification by SUMO (small ubiquitin-like modifier) plays a role in multiple cell processes, including transcriptional regulation, DNA damage repair, protein localization and trafficking. Factors affecting protein localization and trafficking are particularly crucial in neurons because of their polarization, morphological complexity and functional specialization. SUMOylation has emerged as a major mediator of intranuclear and nucleo-cytoplasmic translocations of proteins involved in critical pathways such as circadian rhythm, apoptosis and protein degradation. In addition, SUMO-regulated re-localization of extranuclear proteins is required to sustain neuronal excitability and synaptic transmission. Thus, SUMOylation is a key arbiter of neuronal viability and function. Here, we provide an overview of recent advances in our understanding of regulation of neuronal protein localization and translocation by SUMO and highlight exciting areas of ongoing research.SUMOylation is the covalent attachment of a member of the SUMO family of proteins to one or more lysine residues on target proteins. SUMOylation is best characterized for nuclear proteins involved in genome integrity, nuclear structure and transcription [1,2] but it is now clear that SUMOylation is also important for extranuclear signal transduction, trafficking and modification of cytosolic and integral membrane proteins. Since the first report of SUMOylation as a modification of the nuclear pore component RanGAP [3,4], several hundred SUMOylation substrates have been characterized and many more putative targets have been identified by proteomic studies [5,6].Neurons are highly-specialized morphologically complex polarized cells that exhibit constant constitutive and activity-dependent directed protein transport. The signaling pathways that orchestrate protein trafficking are necessarily sophisticated and multilayered and it has become evident that for many neuronal proteins SUMOylation is an important factor in regulating their localization and function under both physiological and pathophysiological conditions [7,8]. In this review we focus on recent advances regarding the effects of SUMOylation on protein trafficking in neurons (Table 1).Consequences of SUMOylation of proteins of different cell compartments.In the nucleus, SUMOylation is mainly associated with transcriptional regulation and DNA damage repair. SUMO modification of transcription factors or their interaction partners can lead to an increase or decrease in transcriptional activity of the cell [24], while SUMOylation of a number of components of the DNA damage repair machinery is required for their localization to points of DNA damage [25,26]. In neurons, SUMOylation is also emerging as an important regulator of sub-nuclear organization and protein trafficking within the nucleus (Figure 1).Translocation of nuclear and nucleo-cytoplasmic proteins following SUMOylation. In the nucleus, SUMO modification leads to the formation of Cajal bodies, which are positive for coilin and SUMO. Other SUMO related nuclear structures are PML bodies, which contain SUMOylated PML and SUMOylated BMAL1. Caspases -2/-7/-8, GSK3beta and FAK translocate to the nucleus on SUMOylation. Abbreviations: PML: promyelocytic leukemia; BMAL: brain and muscle aryl hydrocarbon receptor nuclear translocator-like; GSK: glycogen synthase kinase; FAK: focal adhesion kinase.Cajal bodies (CBs) are small nuclear structures present in metabolically active mammalian cells, especially neurons, where their size is often increased compared to other cell types [27]. They are characterized by the presence of the two main components, the p80 coilin and the survival motor neuron (SMN) protein. CBs are involved in the processing of replication-dependent histone mRNAs and in the biogenesis of ribonucleoproteins associated with pre-mRNA, pre-rRNA processing and splicing [27]. CBs are transcription-dependent and high numbers of Cajal bodies are indicative of a transcriptionally active cell with a high cellular mass [28]. SUMO1 and Ubc9 transiently co-localize to Cajal bodies in undifferentiated neuron-like UR61 cells [29]. Although not yet validated as SUMO substrates, both coilin and SMN possess high probability SUMOylation sites and coilin interacts with the SUMO E3 ligase PIASγ [30]. Interestingly, this association occurs only in undifferentiated cells and cells that have been exposed to stress [29]. Therefore, it has been hypothesized that SUMO1 translocation to CBs is dependent on events during neuronal differentiation and forms part of a stress response in differentiated cells. These results are consistent with reports that changes in protein SUMOylation play an important role in neuronal development and survival [8,31,32,33]. Hence, it seems likely that CBs represent a potential target of neuroprotection by SUMOylation.Another nuclear structure abundant in SUMO is the PML body, which are highly enriched in the promyelocytic leukemia (PML) protein. SUMOylation is prerequisite for the PML body formation since only SUMOylated PML is able to recruit other PML body proteins such as Daxx [9,10,34]. PML bodies are involved in a variety of processes including viral defense, stress response and genome stability [35,36,37]. Although PML bodies appear to be absent from most neurons [38,39] they are present in human dorsal root ganglion neurons (DRGN) [40]. Interestingly, in DRGNs affected by acute inflammatory demyelinating polyneuropathy (AIDP) associated with Guillain-Barre syndrome [41], the number of PML bodies increases with the severity of the dymyelination [40]. Thus, PML bodies are potentially involved in the DRGN response to axonal damage and injury. This is supported by the finding that in AIDP neurons, the glucocorticoid receptor (GR), itself a SUMO substrate, is localized to PML bodies, whereas in healthy cells GR does not localize to any nuclear foci [42,43]. This re-localization might play a role on the transcriptional regulation of genes expressed upon axonal damage.All organisms are affected by a circadian rhythm, which spans a period of about 24 hours [44]. In mammalians, the circadian rhythm is regulated by a master clock in the suprachiasmatic nuclei of the hypothalamus which is entrained by the natural light-dark cycles [44,45]. Disruption of the circadian rhythm can be involved in bipolar disorder, sleeping disorders and dementia [45]. The transcription factor BMAL1 is an essential component of the clock [44,46] and can be SUMOylated by SUMO1, SUMO2 or SUMO3 [11,12]. Under physiological conditions, BMAL1 is predominantly modified by SUMO2/3, which is involved in the localization of BMAL1 to PML bodies where it functions as a transcription factor [11]. In addition, SUMOylation also triggers ubiquitin-mediated periodic degradation of BMAL1 at PML bodies [11,12]. Although SUMOylation has been thought to antagonize proteasomal degradation of proteins, the discovery of SUMO-targeted ubiquitin ligases (StUbls) has shown that SUMO and ubiquitin can collaborate to promote protein degradation [47]. StUbls are present at PML bodies [48] so it is likely that an as yet unidentified StUbl regulates BMAL1 SUMOylation and degradation.The activity-dependent removal and insertion of AMPA receptors from and into the post-synaptic membrane underlies long-term depression (LTD) and long-term potentiation (LTP) [49,50]. Although AMPA receptors themselves do not appear to be SUMO substrates [19], synaptic SUMOylation has been implicated in LTP [51]. For example, glycine-induced chemical LTP leads to an increased co-localization of Ubc9 and SUMO1 in dendrites [52], suggesting activity-dependent loading of Ubc9 with SUMO. A possible SUMO target known to regulate AMPA receptor surface expression is the activity-related cytoskeletal-associated protein Arc/Arg3.1 [18]. It has been suggested that SUMOylation causes Arc to relocate into dendrites [18] where interaction of Arc with the cytoskeleton is involved in the establishment and maintenance of LTP and synaptic scaling [7,18].Kainate receptors (KARs) play important roles in the regulation of synaptic transmission and neuronal excitability [53,54,55,56]. The KAR subunit GluK2 is SUMOylated in response to agonist stimulation, which leads to endocytosis of GluK2-containing kainate receptors [19]. Although both, kainate and NMDA-stimulation, lead to KAR internalisation, NMDA-induced endocytosis does not involve GluK2 SUMO modification [19]. Agonist-induced or NMDA-induced endocytosis events are respectively associated with GluK2 degradation or recycling [57]. Thus, it is tempting to speculate that SUMOylation might target GluK2 towards lysosomal degradation following endocytosis, whereas non-modified GluK2 is recycled and can be subsequently re-inserted into the membrane. More recently, details of how SUMOylation of GluK2 is regulated and the importance of this modification for KAR plasticity have emerged. It has been shown that kainate stimulation leads to PKC-mediated phosphorylation of GluK2 which, in turn, promotes its SUMOylation and internalization [58]. Notably, this phosphorylation-mediated SUMOylation of GluK2 is required for the removal of GluK2-containing KARs during KAR LTD at mossy fibre synapses [59], providing the first direct example of a requirement for SUMO-mediated protein trafficking in a form of synaptic plasticity (Figure 2a).The group III family of metabotropic glutamate receptors (mGluRs) comprises mGluR4 and mGluR6-8 [60]. They are expressed throughout the brain and, with the exception of mGluR6, are mostly presynaptic at glutamatergic and GABAergic terminals, where they act to regulate presynaptic release [60]. In yeast two-hybrid assays the c-termini of mGluR8a and 8b interact with Ubc9 and SUMO1, in addition to the SUMO E3 ligases PIAS1, PIASγ, PIASxβ. Subsequently, it was shown that PIAS1 interacts with the c-termini of all group III mGluRs [21] and that the mGluR8 c-terminus can also interact with the E3 ligases Pc2 and PIAS3L [20]. In addition, full-length mGluR8b can be SUMOylated in HEK293 cells co-transfected with SUMO1 [20].However, despite these data, no direct co-immunoprecipitation of SUMOylated endogenous mGluRs from neurons or brain has yet been achieved and the functional consequences of mGluR SUMOylation remain to be determined. Because of this it has been questioned if group III mGluRs are actually true SUMO targets. For example, although mGluR7 is SUMOylated at a specific lysine residue in vitro, SUMO modification in neurons was not detected and no functional or trafficking differences were observed upon over-expression of a non-SUMOylatable mutant of mGluR7 [61]. Thus, while it remains possible that endogenous mGluRs do undergo SUMOylation, the number of receptors SUMOylated at any given time may be very small and the functional consequences very subtle.Influence of SUMOylation on receptor trafficking and protein transport. (A) At the pre-synapse, de-SUMOylation of the cannabinoid receptor CB1 is proposed to lead to its internalization. In contrast, SUMOylation of group III mGluRs may lead to internalisation and/or degradation. Agonist-induced phosphorylation of GluK2 in post-synaptic kainate receptors causes endocytosis. Potential SUMO regulation of AMPA receptors may be via SUMO modified interacting proteins; (B) In the axon, the RNA binding protein La is SUMOylated, triggering its retrograde axonal transport through binding to dynein. Abbreviations: mGluR: metabotropic glutamate receptor.The cannabinoid receptor 1 (CB1) is the most widely-expressed G-protein coupled receptor in the brain [62]. CB1 is distributed largely presynaptically and, upon agonist activation, acts to suppress neuronal excitability and neurotransmitter release through coupling to Gi and Go G proteins [63]. In primary rat cortical neurons the CB1 agonist Δ9-THC has been reported to increase levels of un-conjugated SUMO1, an effect blocked by the selective CB1 antagonist AM251. The authors propose that some of this increase in free SUMO1 derives from de-SUMOylation of CB1 and the tumor suppressor p53 following treatment with Δ9-THC [22] and that de-SUMOylation of CB1 causes its internalization while de-SUMOylation of p53 hinders its nuclear export [22,64]. Further validation is required but these observations would suggest that CB1 is unusual in being highly SUMOylated under basal conditions, in contrast to the vast majority of reported SUMO substrates, and that its trafficking and translocation are induced by activity-dependent deSUMOylation.It is well-established that local protein translation occurs in axons and dendrites [65]. This requires the packaging of mRNAs and components of the translation machinery into granules and transport to designated translation sites [66]. The RNA-binding protein La recognizes RNA transcripts containing a 5’-UTR terminal oligopyrimidine (TOP) element and acts as an RNA chaperone during transport [67]. Neurons express several mRNAs containing the TOP motif, including the known La target grp78/BiP [68]. La can move along the microtubule network in retrograde or anterograde directions by binding dynein or kinesin, respectively [23]. SUMOylation of La appears to be the switch that determines the direction of La transport. SUMOylated La only binds to dynein and therefore is subject to retrograde transport towards the nucleus. Conversely, non-SUMOylated La binds kinesin and undergoes anterograde transport [23]. This is an intriguing observation and further work should define if SUMOylation might act as a general switch that can determine the polarity of target protein transport (Figure 2b).The serine/threonine kinase GSK3β plays a central role in many cell pathways including energy metabolism, Wnt signalling, neuronal development, inflammation, tumorigenesis and cell death [69,70,71,72]. In neurons, GSK3β has been reported to be involved in regulating the balance between LTP and LTD [73] and has been implicated in a number of neurodegenerative disorders. Like many of the proteins it phosphorylates, GSK3β can be SUMOylated [13]. Wild type GSK3β is present throughout the cytoplasm and nucleus whereas a SUMOylation-deficient GSK3β mutant is excluded from the nucleus and present only in the cytoplasm [13]. It remains to be established if deSUMOylated GSK3β is actively exported from the nucleus or if the SUMO moiety acts as a nuclear localization signal. Whatever the mechanism, it appears again that SUMOylation acts as a switch to bring about translocation/re-localization of the substrate protein, potentially playing a central role in synaptic plasticity and neuronal function.FAK is an important mediator of signals between the extracellular matrix and the cytoplasm and is involved in controlling cell motility, shape and adhesion [74,75]. In neurons, FAK is important for neuronal migration and axon pathfinding [76,77,78]. FAK autophosporylates at Tyr397 to create a binding site for interaction partners including other kinases such as Src and Fyn [79]. SUMOylation of FAK at Lys152 increases autophosphorylation and promotes its nuclear localization [17,80]. However, the nuclear function of FAK remains unclear and further work is required to define how this nucleo-cytoplasmic shuttling of FAK is mediated.The caspase family of cysteine proteases mediate apoptotic neuronal death [81] but members of this family have also been reported to play physiological roles in AMPA receptor trafficking and synaptic plasticity [82]. In general, caspases can be divided into two subgroups, namely the initiator caspases and the effector caspases. Members of both subgroups have been reported to be SUMOylated, which promotes their nuclear localization [14,15,16]. Caspase-2 and caspase-7 are localized in nuclear speckles, identified as PML bodies for caspase-2 [15,16], whereas caspase-8 shows a more dispersed nuclear distribution [14]. It has been hypothesized that this translocation of caspases to the nucleus acts on specific targets in circumstances of cell stress such as neuronal hypoxia in the brain [15]. Interestingly, among the caspases there is no common domain containing the putative SUMOylation site. For the initiator caspases the SUMOylation sites are located within the death effector domain or the caspase recruitment domain [14,16] but the SUMOylation site of the effector caspase-7 is within the p20 subunit of the activated caspase [15]. As yet, very little functional data has been reported but SUMOylation of caspase-2 appears to increase its maturation from procaspase-2 [16]. Clearly, there is plenty of potential for future investigation.Protein SUMOylation is a major regulator of neuronal function and dysfunction. It is involved in many diverse cell pathways and, in many cases transient SUMOylation initiates the translocation of substrate proteins between compartments within the cell. Misregulation of these SUMO-regulated processes can disrupt proper protein translocation and has been implicated in a wide range of neuronal diseases. Thus, for these substrates, SUMOylation can be viewed as a mobilization factor that orchestrates appropriate protein–protein interactions to mediate relocation and instigate downstream functional consequences. As we have pointed out, understanding of the mechanisms, targets and consequences of SUMOylation in neurons is at a very early stage. It is clear, however, that this is an important and exciting field that will undoubtedly shed new light on fundamental aspects of neuronal function and dysfunction.We are grateful to the MRC, BBSRC, Wellcome Trust and ERC for financial support.

Med-MDPI/biomolecules/biomolecules-02-02-00269.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).The secretion of insulin by pancreatic islet β-cells plays a pivotal role in glucose homeostasis and diabetes. Recent work suggests an important role for SUMOylation in the control of insulin secretion from β-cells. In this paper we discuss mechanisms whereby (de)SUMOylation may control insulin release by modulating β-cell function at one or more key points; and particularly through the acute and reversible regulation of the exocytotic machinery. Furthermore, we postulate that the SUMO-specific protease SENP1 is an important mediator of insulin exocytosis in response to NADPH, a metabolic secretory signal and major determinant of β-cell redox state. Dialysis of mouse β-cells with NADPH efficiently amplifies β-cell exocytosis even when extracellular glucose is low; an effect that is lost upon knockdown of SENP1. Conversely, over-expression of SENP1 itself augments β-cell exocytosis in a redox-dependent manner. Taken together, we suggest that (de)SUMOylation represents an important mechanism that acutely regulates insulin secretion and that SENP1 can act as an amplifier of insulin exocytosis.Insulin secretion from pancreatic islet β-cells is tightly regulated by many factors including nutrients, hormones, and neurotransmitters. Increased insulin release in response to glucose is central to energy homeostasis, and is impaired in diabetes [1]. At least two pathways control glucose-stimulated insulin secretion in β-cells and are usually referred to as ‘triggering’ and ‘amplifying’ pathways [2,3,4] (Figure 1). The triggering pathway is characterized by events leading to an increase in intracellular Ca2+ following the metabolic generation of ATP, inhibition of ATP-sensitive K+ (KATP) channels, and subsequent activation of voltage-dependent Ca2+ channels (VDCCs) [5]. The secretory response to Ca2+ can be amplified by several receptor-mediated and metabolic signals [2,3,4]. Key among these is the generation of cAMP through the action of incretin hormones such as glucagon-like peptide-1 (GLP-1) [6,7,8], while several candidate metabolic signals have been proposed [9,10]. Among these, NADPH derived from the mitochondrial export of malate and citrate is perhaps the strongest candidate [9,10,11,12]. The downstream mechanism by which NADPH affects insulin secretion is unclear, but may involve a glutaredoxin-dependent pathway [13,14] which presumably transduces the generation of these reducing equivalents into an action on the secretory machinery through the cycling of reduced/oxidized glutathione couples [15].Potential sites of regulation by SUMOylation in the mechanism of glucose-stimulated insulin secretion from pancreatic β-cells. The general mechanism regulating insulin secretion through triggering and amplifying pathways is shown. SUMOylation (S) may control insulin gene transcription, mitochondrial (Mito.) function, GLP-1 receptor signaling/localization, Kv2.1 channel function and downstream exocytosis in response to metabolic signals.SUMOylation, the post-translational modification of a target protein by the covalent attachment of SUMO (small ubiquitin like modifier), modulates target activity, protein-protein interactions, and sub-cellular localization [16,17]. Much work has been devoted to understanding the nuclear effects of SUMOylation, and indeed in pancreatic β-cells SUMOylation of the transcription factors Pdx-1 and MafA regulates their localization and control of insulin gene transcription [18,19]. Recently, extra-nuclear roles have been demonstrated, including the regulation of synaptic transmission [20,21], ion channel activity [22,23,24,25], and mitochondrial function and morphology [26,27,28]. Thus, the SUMOylation pathway is emerging as a regulatory mechanism that can control many key cellular functions.SUMOylation is readily reversible through the action of the sentrin-specific SUMO proteases (SENPs), of which there are at least six identified in mammalian cells (SENP1-3, 5-7) [29]. While the mechanisms regulating SENP activity remain unclear [29,30], recent reports suggest that the enzyme is regulated by the intracellular redox state through the formation of di-sulfide bonds between SENP1 monomers (at least two cysteine residues are implicated, C603 and C613) [31]. This is suggested to modulate the cellular response to reactive oxygen species [32], and could be particularly relevant to pancreatic β-cell function since redox changes are implicated both in the normal physiology of insulin secretion and in diabetes pathophysiology [33,34]. Here we discuss the effects of SUMO1, and the role of (de)SUMOylation (particularly through SENP1), in the control of insulin secretion. Furthermore, we present evidence that SENP1 is an important mediator of insulin exocytosis in response to the key metabolic signal, NADPH.Up-regulation of SUMOylation in pancreatic islets or β-cells inhibits insulin secretion stimulated by glucose [35] or activation of the GLP-1 receptor [36]. This occurs in the absence of changes in VDCC activity and intracellular Ca2+ responses, suggesting that much of the upstream mechanism (i.e., glucose metabolism, electrical activity and Ca2+ entry) controlling insulin secretion remains intact [35]. In order to be secreted, insulin granules must first be trafficked to the plasma membrane [37,38]. This process as well is intact following SUMO1 over-expression in insulin-secreting cells. In fact, cells over-expressing SUMO1 have more secretory granules localized to the plasma membrane (Figure 2). Thus, while the insulin granule trafficking and glucose-dependent Ca2+ responses are intact, insulin release is blunted following SUMO1 over-expression [35,36]. This would suggest that SUMO1 acts far downstream in the secretory pathway, likely by inhibiting the exocytosis of insulin granules in response to the intracellular Ca2+ signal [35]. As such, the increased secretory granule density at the plasma membrane of insulin secreting cells (INS-1 insulinoma cells are shown in Figure 2) is a ‘traffic jam’ resulting from efficient plasma membrane targeting of insulin granules which are then unable to undergo exocytosis and release their cargo.Since it is possible that the up-regulation of SUMO1 expression could inhibit insulin secretion by down-regulating the expression of insulin itself [18,19], it is important to note that we observed no change in insulin content upon over-expression of SUMO for 24–48 hours. While a more chronic over-expression of SUMO1 may result in decreased insulin expression, and indeed was observed in insulinoma cells stably over-expressing GFP-SUMO1 [36]. The SUMO1-dependent inhibition of insulin exocytosis certainly occurs acutely however, and is likewise rapidly reversible, as the direct intracellular dialysis of recombinant SUMO1 into β-cells blocks exocytosis within 1–2 minutes and this can be reversed (with a similar time course) by the direct infusion of SENP1 [35]. These findings are important for two reasons: (1) the regulation of exocytosis in these cells is dependent on SUMO-conjugation to a target (i.e., is reversible by SUMO cleavage); and (2) the effects of SUMOylation on exocytosis are very rapid, precluding a role for changes in gene expression per se.SUMOylation inhibits the exocytosis of secretory granules, without affecting granule trafficking to the plasma membrane. (A) SUMOylation (S) acts downstream of insulin granule targeting to the plasma membrane to inhibit Ca2+-dependent exocytosis. Up-regulating SUMO1 blocks downstream exocytosis and results in a build-up of secretory granules at the plasma membrane. This can be seen by total-internal reflection fluorescence (TIRF) imaging of cortical actin and secretory granules labeled with NPY-mCherry within ~100 nm of the plasma membrane. Representative images are shown in (B), and quantified data is shown in (C). *** p < 0.001 compared with control.There are several points at which SUMOylation may impact insulin secretion (Figure 1). SUMOylation can control nuclear signaling in β-cells [39], including regulation of the key insulin gene transcription factors MafA and Pdx1 [18,19]. Somewhat further ‘downstream’, SUMOylation has been demonstrated to control mitochondrial fission and function [26,28]. Although unexplored in the context of insulin secretion and β-cell function, this would be expected to modulate the generation of metabolic signals that are essential for insulin secretion. While our recent findings suggest that mitochondrial function may be more-or-less intact following SUMO1 over-expression given normal islet Ca2+ responses [35], a more detailed analysis of mitochondrial activity and morphology would be required to support a lack of effect of SUMO1 on β-cell mitochondria.One recent report demonstrated the SUMOylation-dependent trafficking of the G-protein-coupled GLP-1 receptor [36]. A key factor in promoting postprandial insulin secretion, GLP-1 is secreted from intestinal L-cells following a meal and acts to augment the β-cell secretory response to circulating glucose primarily through the generation of cAMP (reviewed extensively [6,7,40]). Rajan et al. [36] showed that SUMOylation of the GLP-1 receptor prevents its cell surface trafficking, resulting in impaired cAMP generation and insulin secretion in response to GLP-1. Intriguingly, these authors also demonstrated the up-regulation of SUMO mRNA (and that of the SUMO-conjugating enzyme Ubc9) following exposure of mouse islets to high glucose, raising the possibility that increased SUMOylation could contribute to the reduced insulin secretion observed in diabetes.Given that SUMO1 can inhibit insulin exocytosis downstream of the sites mentioned above, we have focused our attention on potential SUMOylation targets at the exocytotic site. Our initial work focused on the voltage-dependent K+ (Kv) channel Kv2.1, which is highly expressed in rodent and human β-cells and mediates action potential repolarization [41,42,43,44], since related K+ channels are regulated by SUMOylation [22,23]. Indeed, both cloned Kv2.1 and the native channel in rodents and humans is inhibited by SUMOylation [24] of a C-terminal lysine (K470) [25]. While inhibition of Kv2.1 currents in itself cannot account for the ability of SUMO1 to block exocytosis, since our exocytosis measurements (such as those in Figure 3, Figure 4, Figure 5) are carried out under conditions where the cell membrane potential is ‘clamped’, it is interesting to note that the Kv2.1 channel can regulate β-cell exocytosis independent of its electrical function, through a direct interaction with the exocytotic protein syntaxin 1A [45,46]. It will thus be interesting to determine whether SUMOylation of Kv2.1 alters its role in β-cell exocytosis and its interaction with syntaxin 1A, which we have also recently identified to itself be SUMOylated in pancreatic islets (unpublished). Finally, as discussed in further detail below, our original screen for exocytotic proteins that interact with SUMO1 in insulin-secreting cells identified a likely candidate to be synaptotagmin VII [35], which is the primary Ca2+-sensor for insulin exocytosis [47] and possesses two candidate SUMOylation sites; one each in close proximity to its two Ca2+-binding domains.Amplification of β-cell exocytosis by NADPH. (A) Mouse β-cells were maintained at a non-stimulatory glucose concentration (1 mM) for 2 hours prior to intracellular dialysis of NADPH/NADP+ (at molar ratios of either 1:10 or 10:1) and whole-cell patch-clamp measurement of exocytosis; (B) Exocytosis is measured as the increase of cell surface area (capacitance) that occurs following membrane depolarization, opening of voltage-dependent Ca2+ channels, and subsequent fusion of secretory granules with the plasma membrane; (C) The exocytotic response to a series of ten membrane depolarizations remained low under the reduced NADPH condition (1:10), but is significantly amplified by elevation of NADPH (10:1) even when glucose remains low (1 mM).SENP1 is required for NADPH-dependent amplification of exocytosis. (A) Mouse β-cells were maintained at a non-stimulatory glucose concentration (1 mM) for 2 hours prior to intracellular dialysis of NADPH/NADP+ at a molar ratio 10:1 and whole-cell patch-clamp measurement of exocytosis; (B,C) Representative traces (B) of exocytotic responses and averaged data (C) in control cells (Ad-sh-scramble) or following SENP1 knockdown (Ad-sh-SENP1). ** p < 0.01 and *** p < 0.001 compared with Ad-sh-scramble.SENP1 amplification of β-cells is redox-dependent. (A) Mouse β-cells, infected with adenovirus expressing GFP (Ad-GFP) or SENP1 (Ad-SENP1, which co-expresses GFP), were maintained in non-stimulatory glucose (1 mM) for 2 hours prior to whole-cell patch-clamp. In some experiments, H2O2 (200 µM) was infused directly into cells at the time of the experiment; (B) Immunoblotting of protein lysates from INS-1 832/13 cells demonstrates the up-regulation of monomeric SENP1 (with Ad-SENP1) and its subsequent dimerization in the presence of H2O2 (200 µM); (C,D) Over-expression of SENP1 recapitulates the effect of NADPH to amplify β-cell exocytosis (C), an effect that is lost under oxidizing conditions (D). * p < 0.05 compared with Ad-GFP.The cellular and molecular pathways that modulate SUMOylation at the exocytotic site are unknown, but our work suggests that SUMOylation acts as a ‘brake’ on exocytosis, perhaps to prevent unwanted insulin secretion when circulating glucose levels are low. The implication of this is interesting, as it suggests that the acute facilitation of exocyosis is regulated by removal of SUMO-conjugates (i.e., deSUMOylation). Indeed, synaptotagmin VII appears to be basally SUMOylated at non-stimulatory glucose concentrations [35]. Following glucose stimulation, we observed a transient deSUMOylation of synaptotagmin VII (which could be replicated by SENP1 over-expression). Indeed, deSUMOylation is required for the glucose-dependent amplification of insulin exocytosis since block of deSUMOylation (by either up-regulating the conjugating enzyme Ubc9, or knockdown of SENP1) prevents glucose-dependent amplification of β-cell exocytosis. Furthermore, up-regulation of SENP1 is itself able to amplify the β-cell exocytotic response to Ca2+, recapitulating the effect of glucose-stimulation (seen also in Figure 5). Thus, the deSUMOylating enzyme is both required and sufficient in glucose-dependent amplification of β-cell exocytosis, although the molecular and metabolic pathways linking glucose-stimulation and deSUMOylation are unknown.One possible link between glucose-stimulation and the enhancement of insulin exocytosis is NADPH, which can directly augment β-cell exocytotic responses [13,14]. We show here that NADPH is able to enhance the exocytotic response of β-cells even in the absence of stimulatory glucose (Figure 3). In these cells NADPH is dialysed directly into mouse β-cells (at a molar ratio of ether 1:10 or 10:1 with NADP+) through a patch-clamp pipette, following which the cell is subjected to series of membrane depolarizations to activate Ca2+ channels, allowing Ca2+ influx and the stimulation of exocytosis [48]. In this case, exocytosis is monitored as increased cell surface area (called capacitance) upon the fusion of secretory granules.To determine whether SENP1 is required for the NADPH-dependent amplification of insulin exocytosis a knock-down approach was used [35]. NADPH-dependent amplification of exocytosis was observed in β-cells infected with a control adenovirus (Ad-sh-Scramble; n = 28) (Figure 4). Using a SENP1-targed shRNA adenovirus (Ad-sh-SENP1) described previously [35], we find that NADPH fails to enhance exocytosis following SENP1 knock-down, where the response was impaired by 57% compared to the control (n = 21, p < 0.01) (Figure 4). Conversely, SENP1 up-regulation augments exocytosis in mouse β-cells in the absence of a glucose-stimulus or NADPH (n = 12) compared with that seen upon expression of GFP only (n = 8, p < 0.05) (Figure 5). We saw no difference in VDCC activity following either SENP1 knockdown or over-expression (not shown). Since intracellular redox state is suggested to regulate SENP1 activity, possibly by controlling SENP1 dimerization [31,32], we examined whether SENP1 amplification of exocytosis was dependent upon intracellular redox. For this, hydrogen peroxide was used, which we show to promote SENP1 dimerization in INS-1 832/13 insulinoma cells (Figure 5B). When 200 µM of H2O2 were added in the pipette solution SENP1 over-expression was no longer able to amplify the exocytotic response of mouse β-cells (n = 13) compare to the GFP control (n = 11) (Figure 5).Our recombinant adenoviruses expressing either SENP1 (Ad-SENP1) or a SENP1-targeted shRNA (Ad-sh-SENP1) were described and characterized previously [35]. These co-express GFP to allow identification of infected cells. Adenoviruses expressing GFP alone (Ad-GFP) or a scrambled shRNA sequence (Ad-sh-Scrambled) served as controls.Pancreatic islets were isolated from male C57/BL6 mice by collagenase (1 mg/mL) digestion and handpicked as previously [35,50]. Islets were then dispersed to single cells by shaking 11 min in Ca2+-free media (138 mM NaCl, 5.6 mM KCl, 1.2 mM MgCl2, 5 mM HEPES, 3 mM glucose, 1 mM EGTA, 1 mg/mL albumin). Cells were plated into 35 mm culture dishes in RPMI 1640 with L-glutamine, 10% FBS, and 100 units/mL penicillin/streptomycin. When primary cells and cell lines were infected with Ad-GFP or Ad-SENP1 they were cultured 2 days prior to experiments while when infected with Ad-sh-Scrambled or Ad-sh-SENP1 they were cultured for 3 days to allow for SENP1 knockdown. All animal experiments were approved by the animal care and use committee at the University of Alberta.INS1 832/13 insulinoma cells (a gift from Prof. C. Newgard, Duke University) were maintained in RPMI1640 medium with 11.1 mM glucose and 2 mM L-glutamine, supplemented with 10% FBS, 10 mM HEPES, 100 U/mL penicillin/streptomycin, 1 mM sodium pyruvate, 50 μM β-mercaptoethanol, in a humidified atmosphere (5% CO2, 37 °C).INS1 832/13 cells infected with Ad-GFP or Ad-SENP1 were incubated in KRBH buffer (135 mMNaCl, 3.6 mMKCl, 0.5 mM MgCl2, 0.5 mM NaH2PO4, 10 mM HEPES, 2 mM NaHCO3, 1.5 mM CaCl2, 0.1% BSA, 11.1 mM glucose, pH = 7.4) with or without 1 mM hydrogen peroxide (H2O2, Sigma) for 30 min in the dark at 37 °C. Cells were harvested and lysed in Cell-Lytic-M buffer (Sigma-Aldrich), supplemented with DTT and protease inhibitor cocktail (Bio Basic Inc.) and kept on ice for 30 min. Lysates were separated in 8% SDS-PAGE and transferred to a polyvinylidenedifluoride (PVDF) membrane (Millipore, Billerica, MA) followed by blocking with 5.0% nonfat dry milk in TBST (150 mM NaCl, 50 mM Tris, and 0.1% Tween 20, pH = 7.4) for 1 h at room temperature. The primary antibodies were a mouse monoclonal anti-SENP1 (C12) (1:700, Santacruz) and mouse monoclonal anti-β-tubulin (1:40,000, Sigma; not shown) at 4 °C overnight. After washing with TBST, the membrane was incubated with peroxidase-linked anti-mouse IgG (Amersham Bioscience; 1:5,000) for 1h at room temperature. Detection was done by an enzymatic chemiluminescence (ECL) kit (Amersham Bioscience) and exposure to X-ray film (Fujifilm, Tokyo, Japan).Changes in membrane capacitance were monitored in whole-cell configuration using EPC10 patch-clamp amplifier controlled with PatchMaster software (HEKA Electronik, Lambrecht, Germany). Experiments were performed at 32–35 °C. Extracellular solution contained (in mM) 118 NaCl, 20 TEA, 5.6 KCl, 1.2 MgCl2, 2.6 CaCl2, 1 glucose and 5 HEPES (pH 7.4 with NaOH). The intracellular solution contained (in mM) 125 Cs-glutamate, 10 CsCl, 10 NaCl, 1 MgCl2, 0.05 EGTA, 5 HEPES, 0.1 cAMP, and 3 MgATP (pH 7.15 with CsOH). For some experiments pipette solution was supplemented with β-NADPH and β-NADP+ at 100/10 μM; or H2O2 at 200 µM. The resistances of patch pipettes, pulled from borosilicate glass and coated with sylgard was 4–6 MΩ when filled with pipette solution. Data was normalized to initial cell size and expressed as fF/pF. Mouse β-cells were identified by the absence of voltage-gated Na+ current from a holding potential at −70 mV. Data analysis was with FitMaster software (HEKA Electronik, Lambrecht, Germany) and Origin 7.0 and was statistically evaluated with two-tailed, unpaired Student's t-test or one way-ANOVA followed by Scheffe's post-hoc test. All data are expressed as means ± SEM and p < 0.05 was considered significant.Insulin exocytosis is a process finely tuned by at least two glucose-dependent pathways: triggering and amplifying. The mechanisms underlying the triggering pathway are well known, whereas the dynamics of the amplifying pathway have still not been revealed. Nevertheless they are strictly interrelated: amplification of β-cell exocytosis is of no consequence until exocytosis is triggered by Ca2+ [3]. Moreover, amplification of insulin release is glucose-dependent and many metabolic signals have been proposed to regulate its occurrence. NADPH is one of the major candidates involved in this process [49]. Our previous work [24,35], and data presented here, demonstrate that the deSUMOylating enzyme SENP1 plays a pivotal role in insulin exocytosis in response to glucose and metabolically-derived signals. One can conclude that NADPH-dependent amplification of β-cell exocytosis is strictly dependent on triggering Ca2+ and is mediated by SENP1.Proposed model for the regulation of insulin exocytosis by deSUMOylation at the exocytotic site. Metabolically derived reducing equivalents, in the form of NADPH, are proposed to amplify insulin exocytosis in part by promoting deSUMOylation of several targets at the exocytotic site. Potential targets include syntaxin 1A, synaptotagmin VII, and the voltage-dependent K+channel Kv2.1. Several questions remain, not the least of which include the molecular mechanism linking NADPH to SENP1 activity.Albeit speculative to indicate which proteins are the SUMO1 targets at the plasma membrane, the exocytotic Ca2+-sensor synaptotagmin VII represents a top candidate. Additionally, novel SUMO-targets such as Kv2.1 and syntaxin 1A have been proposed to coordinate insulin exocytosis per se. A proposed model is shown in Figure 6 whereby the metabolic generation of NADPH, perhaps through its role in determining β-cell redox state, acts through SENP1 to mediate the deSUMOylation of a number of targets at the exocytotic site. This removes the ‘brake’ on exocytosis, thus amplifying the secretory response to elevated intracellular Ca2+. Of course, much remains to be determined and our understanding of pancreatic β-cell SUMOylation in both the long-term and acute regulation of insulin secretion is in its infancy. Further study of these mechanisms, and the pathophysiological role for SUMOylation, will provide a new layer to our insight into insulin secretion in health and diabetes.The authors thank Nancy Smith and James Lyon for their technical support and Dr. Quan Zhang for his advice and suggestions. This work was funded by an operating grant to PEM from the Canadian Institutes of Health Research (CIHR MOP244739) and the National Science and Engineering Research Council. PEM holds a Scholarship from Alberta Innovates-Health Solutions (AI-HS) and the Canada Research Chair in Islet Biology. EV was supported in part by the Gesellschaft von Freunden und Förderer der TU Dresden e.V.

Med-MDPI/biomolecules/biomolecules-02-02-00282.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Biomolecules are composed primarily of the elements carbon, nitrogen, hydrogen, oxygen, sulfur, and phosphorus. The structured assembly of these elements forms the basis for proteins, nucleic acids and lipids. However, the recent discovery of a new bacterium, strain GFAJ-1 of the Halomonadaceae, has shaken the classic paradigms for the architecture of life. Mounting evidence supports the claim that these bacteria substitute arsenic for phosphorus in macromolecules. Herein, we provide a brief commentary and fuel the debate related to what may be a most unusual organism.There exists a wonderful dichotomy in science. On one hand, we have universal laws and principles that are constant: the force of gravity; Newton’s laws of motion; the principles of thermodynamics. On the other hand, scientists discover new things on a daily basis, from the “discovery of the accelerating expansion of the universe through observations of distant supernova” [1] which is incomprehensibly large compared to humankind, to the discovery of quasicrystals in chemistry, which exist on the atomic level. Every once in a while, these new discoveries uproot notions of our world which have existed for hundreds of years, effectively redefining what we understand about the natural world. As such, science has traditionally held that life depends on six essential bioelements: carbon, hydrogen, oxygen, nitrogen, sulfur, and phosphorus. These bioelements, with the help of trace elements - namely metals and metalloids - compose all lipids, proteins, and nucleic acids and thus, are the building blocks of life. The complex interactions between molecules, until recently, were thought to be intricate to the point that there could not be any deviation from the classic norm. Dr. Felisa Wolfe-Simon and her recent discovery of a strain of bacteria which may substitute arsenic for phosphorous in biomolecules challenge this norm.Since the physical constructs of life are based upon essential bioelements, scientists discerned that these six specific elements must be used without exception. Although it is broadly accepted that six aforementioned elements form the basis of all biomolecules, examples of substitutions with trace elements are not absent in the literature. For instance, copper can take the place of iron in certain mollusks and arthropods to facilitate oxygen transport. Cadmium can substitute for zinc in enzyme families [2]. In both of these examples, the substituting element shares a number of commonalities with the parent element, namely orbital configuration, electronegativity, and quantity of valence electrons. These similarities are largely due to the elements belonging to the same periodic group. In consideration of this unique relationship between trace elements, some have hypothesized that the major bioelements may be interchangeable as well. Dr. Wolfe-Simon’s discovery in California may provide evidence to support these hypotheses.While working in Mono Lake, California, a hypersaline and alkaline lake, Dr. Wolfe-Simon stumbled upon a new bacterium; one, she argues, that can use arsenic in place of phosphorus in all of its metabolic functions. Arsenic is a chemical analog of phosphorus, just as copper is to iron, mentioned above. It sits just below phosphorus on the periodic table, making it a non-metal, and implicating similar atomic radii, the same number of valence electrons, and nearly identical electronegativity and orbital configurations. The similarity between phosphorus and arsenic translates to other molecules they form, importantly phosphate (PO43−) and arsenate (AsO43−). It is the similarity between these two derivatives that makes arsenic dangerous; thus its implication in many infamous homicides and questionable deaths over the last 2000 years. Evidence of the similarity between the structure, synthesis, and hydrolysis of both arsenate and phosphate-containing compounds is described by Moore in a study of ADP-arsenate. She cites that the equilibrium constants for the synthesis of both arsenate and phosphate esters is almost identical, implicating that such similarities in enzymatic activity show an intimate relationship between the two derivatives. The study also posits that the rate constant for both the synthesis and hydrolysis of glucose-6-arsenate is 105-fold larger than the rate constant for the same reactions in phosphate, indicating that cellular processes actually favor the use of arsenate-containing compounds to that of the phosphate counterpart [3].Arsenic and phosphorus derivatives are so similar to each other, that arsenate can be incorporated into metabolic pathways usually requiring phosphate. The consequences of this substitution are not felt until the later stages of these pathways, as arsenate is quite unstable, compared to phosphate. A notable example of the problems with arsenate substitution is seen in glycolysis, where the interchanging of arsenate forms products that easily hydrolyze to form the next intermediate which consequently inhibits the production of ATP molecules typically formed in this reaction. Due to this, arsenate is an “uncoupler of glycolysis”, which has major energy consequences [4]. Arsenate can also disrupt the conversion of pyruvate to acetyl CoA, preventing the initiation of the Krebs cycle, leading to additional loss of precious energy molecules [5]. This is the primary cause of arsenic’s toxicity.Resistance to arsenic, especially in bacterial isolates, is not unheard of. A 2010 study by Huang et al. indicated that twelve isolates from various families including Pseudomonas, Naxibacter, Acinetobacter, Mesorhizobium, Enterobacter, Methylobacterium, Bacillus, and Caryophanon displayed arsenic resistance in contaminated soils around Pteris vittata habitats. P. vittata, a Chinese brake fern, is considered an arsenic hyperaccumulator which tolerates levels of arsenic significantly higher than most plants. Its tolerance of markedly increased levels of arsenic and related compounds is a result of the symbiotic relationship between the plant roots and the microbes that colonize the rhizosphere surrounding the plant [6].The isolates mentioned above are not only arsenic resistant, but three of the isolates - one each of the Naxibacter, Mesorhizobium, andPseudomonas species - were shown to reduce arsenate to arsenite due to the presence of a reduction-detoxification mechanism coded by the arr operon. Not only does arsenic resistance occur in microbes adapted to hyperarsenic environment, but metabolic processing of arsenic has been shown in select isolates [6].If an organism had such appropriate mechanisms as a reduction-detoxification mechanism as seen in the Huang study [6] to cope with the instability of arsenic and its derivative compounds, especially considering the known examples of trace element substitution, arsenic could be used in the place of phosphorus. This is exactly what Dr. Wolfe-Simon and her research team set out to test. Operating under the impression that organisms found in Mono Lake, a lake with atypical chemical conditions-namely high levels of arsenic, may be unique in terms of cellular processes and adaptations, the research team isolated GFAJ-1, a bacterium from the family Halomonadaceae [1]. In the laboratory, Wolfe-Simon and her team worked to mimic the natural environment of the bacterium by introducing lake sediments into an artificial medium at pH 9.8, supplementing with glucose, trace minerals, and vitamins, and adding specified concentrations of arsenate ranging from 100 μM–5 mM. The team did not add phosphate to the medium, and went through multiple dilution transfers to reduce any potential residual phosphorus. The medium contained slight amounts of phosphate (3 μM), the result of trace impurities in the added nutrients. The negative control in the experiment would later render this remaining phosphate negligible in terms of supporting bacterial growth [2].Wolfe-Simon began growing the bacterium on agar plates and transferred a colony to the artificial medium described above. Progressively increasing the concentration of arsenate to identify the optimal level for growth in GFAJ-1, she found the optimal concentration is 40 mM AsO43− with no added phosphate. Interestingly, the bacterium grew and propagated at increased rates when phosphate was added, but no growth was observed when the colony was added to a plate with no added phosphate or arsenate. This data is crucial, because it suggests that GFAJ-1 exhibits arsenic-dependent growth but is not an obligate arsenophile [2].In order to discern whether or not the bacterium was taking AsO43− from the medium, the research team measured the amount of intracellular arsenic through inductively coupled plasma mass spectrometry, a type of spectrometry capable of detecting concentrations of metals below one part per trillion (10−12) [7]. The bacterium displayed 10× the amount of arsenic as phosphorus in the +As/−P -grown cells. The detected phosphorus is likely the result of scavenged phosphorus from trace phosphate impurities in the reagents. The levels of residual phosphorus were far below the minimum amount needed to support growth, indicating once again, that this presence of phosphorus was not sufficient to support bacterial growth. These results were confirmed with x-ray analyses and high-resolution secondary ion mass spectrometry [2].After confirming the presence of arsenic inside the bacterium, Dr. Wolfe-Simon investigated the intracellular location of the arsenic. Using radiolabeled arsenate, her team observed arsenic in lipids, proteins, and nucleic acids, indicating much more than just an intracellular accumulation of arsenate. Further analyses that considered the bond lengths between arsenic and other atoms as a comparison to normal bond lengths between phosphorus and those atoms confirmed integration into various cellular biomolecules [2].So why is this significant? What could the discovery of a bacterium named with the acronym for “Give Felisa A Job” mean for science? The answer is simple: everything. This discovery, if corroborated and further studied, would be absolutely groundbreaking to all of science. This seemingly small substitution of arsenic for phosphorus would shake the pillars that all of life science sits upon. Not only would this discovery open the door up to the search for thousands if not millions of other species that can make said substitutions, but it would redefine the principles of biology that have been taught for decades. Since the discovery of DNA and the double helix by Watson and Crick with the help of Rosalind Franklin in 1953, scientists have recognized the code of life as being stable, but extremely specific. The nucleotide bases, the non-covalent interactions that allow for base-specific bonding, the 10.5 bases per helical turn, the rigid and stable backbone, and the sugar pucker all lend to the notion of maintaining an excruciatingly specific composition. The substitution of arsenic for phosphorus and its subsequent change from phosphate to arsenate changes all of that. The effects span from nucleic acids to energy molecules to the exoskeletons of chitinous arthropods and bones of animals [8].Due to the extensive implications of the information published in Dr. Wolfe-Simon’s paper, a multitude of critics have arisen to object to the findings. Many have disagreed with the claims of arsenic substitution made in the publication, citing errors in testing ranging from improperly performing tests to gathering insufficient data from the lack of testing. Others have pointed out problems with the controls set up by the team, indicating that presence of any phosphate in the media disqualified the results. Still, others have debated whether the intracellular arsenic has been incorporated into the biomolecules or simply present in vacuoles within the bacterium [9]. At any rate, members outside of the scientific community may view the criticisms and other events that have transpired as superfluous, vindictive, and outright scathing.Dr. Wolfe-Smith responded to many of these criticisms in interviews with various scientific and media groups. She explained that her team published the data they had at such a quick rate because her laboratory did not have all of the necessary tools to perform many of the more intensive tests that critics have called upon. By publishing the information as quickly as possible, she hoped that other labs, now with their curiosity piqued, would come forward seeking to collaborate further on this project. She has also argued that the tests performed in the published manuscript are indeed sound, and calls upon others to replicate her results [10].It is no surprise that this discovery has come under what some may consider a brutal attack over the last year; the proposed repercussions almost beg of it. This experiment has the potential to change a dogma of biology: that the original six bioelements are essential to life. The current discussion contours that of the one surrounding the realization that prions do not follow the same biochemical patterns that are requisite of all other life. Whereas all organisms, from humans to obligate intracellular parasites - like viruses - follow the DNA to RNA to protein pathway, prions are hypothesized to contain no nucleic acids, and replicate solely with proteins. This breaks what is heralded as the central dogma of biology, and was subjected to heavy attacks from the scientific community at large when Tikvah Alper and John Griffith brought prions to the scientific community in the early 1960’s.Dr. Wolfe-Simon’s discovery faces the same scrutiny today; just as Alper and Griffith in 1963, Darwin in 1859, and Copernicus in 1543. This discussion and debate is the culmination of science. Science is defined as the observation and testing of natural phenomena, but requires so much more than that. Without attacks heralded by intellectuals from a multitude of disciplines and the ensuing pursuit of evidence and answers, breakthroughs in science do not occur. To accept postulations and hypotheses of this magnitude without requesting - if not demanding - further research and explanation is irresponsible to the field of science. Scientific discovery hinges on discussion among peers, it always has and it always will. Those things characterized as “great scientific achievements” have withstood years and sometimes decades of criticisms. They have been accompanied by banishment, disgrace, and legal repercussions for the respective proposers. They have been defended tooth and nail from factions of the religious, scientific, and philosophical communities. Unfortunately, some have died as a result of these discussions, and the majority of intellectuals who had hands in bringing about these scientific breakthroughs died long before the debates were settled and their hypotheses deemed accurate.What does all this mean for the reader? Dr. Wolfe-Simon is being subjected to the same procedure as all those who have come before her. Her hypothesis is certainly one that raises numerous questions and demands more evidence, but the implications of it have the potential to shake the foundation of biology as we have known it for centuries. We encourage continued debate, discussion, and research. We acknowledge that we may not have answers for another two, or five, or ten years, but we trust that science will operate successfully as it has for thousands of years: through healthy debate and discussion.This work was supported by funding to Mark A. Brown from the National Science Foundation (1060548). 

Med-MDPI/biomolecules/biomolecules-02-02-00288.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).LRRs (leucine rich repeats) are present in over 14,000 proteins. Non-LRR, island regions (IRs) interrupting LRRs are widely distributed. The present article reviews 19 families of LRR proteins having non-LRR IRs (LRR@IR proteins) from various plant species. The LRR@IR proteins are LRR-containing receptor-like kinases (LRR-RLKs), LRR-containing receptor-like proteins (LRR-RLPs), TONSOKU/BRUSHY1, and MJK13.7; the LRR-RLKs are homologs of TMK1/Rhg4, BRI1, PSKR, PSYR1, Arabidopsis At1g74360, and RPK2, while the LRR-RLPs are those of Cf-9/Cf-4, Cf-2/Cf-5, Ve, HcrVf, RPP27, EIX1, clavata 2, fascinated ear2, RLP2, rice Os10g0479700, and putative soybean disease resistance protein. The LRRs are intersected by single, non-LRR IRs; only the RPK2 homologs have two IRs. In most of the LRR-RLKs and LRR-RLPs, the number of repeat units in the preceding LRR block (N1) is greater than the number of the following block (N2); N1 » N2  in which N1 is variable in the homologs of individual families, while N2 is highly conserved. The five families of the LRR-RLKs except for the RPK2 family show N1 = 8 − 18 and N2 = 3 − 5. The nine families of the LRR-RLPs show N1 = 12 − 33 and N2 = 4; while N1 = 6 and N2 = 4 for the rice Os10g0479700 family and the N1 = 4 − 28 and N2 = 4 for the soybean protein family. The rule of N1 » N2 might play a common, significant role in ligand interaction, dimerization, and/or signal transduction of the LRR-RLKs and the LRR-RLPs. The structure and evolution of the LRR domains with non-LRR IRs and their proteins are also discussed. LRR (leucine rich repeat) regions are present in over 14,000 proteins in the data bases-PFAM, SMART, PROSITE, and InterPro [1,2,3,4]. LRR-containing proteins have been identified in viruses, bacteria, archaea, and eukaryotes. Arabidopsis thaliana and Oryza sativa subsp. japonica (rice) contain over 700 and 1,400 LRR proteins, respectively [5]. Most LRR proteins are involved in protein-ligand and in protein-protein interactions; these LRR proteins include plant immune response and mammalian innate immune response [6,7,8,9,10]. Most LRR repeating units are 20–30 residues in length. All LRR units can be divided into a HCS (highly conserved segment) and a VS (variable segment). The HCS part consists of an 11 residue stretch, LxxLxLxxNxL, or a 12 residue stretch, LxxLxLxxCxxL, in which “L” is Leu, Ile, Val, or Phe, “N” is Asn, Thr, Ser, or Cys, and “C” is Cys, Ser or Asn [7,11,12,13,14]. Eight classes of LRRs have been characterized by different lengths and consensus sequences of the VS part of the repeats. They are “RI-like”, “CC”, “Bacterial”, “SDS22-like”, “plant specific (PS)”, “Typical”, “TpLRR”, and “IRREKO”. Plant specific LRRs (class: PS-LRR) are 23 to 25 residues long and contain a conserved consensus sequence of the VS part, SGxIPxxLxxLxx, in which “S” is Ser or Thr, “G” is Gly or Ser, “I” is Ile or Leu, and “L” is Leu, Ile, Val, Phe, or Met, and “x” is any amino acid [14]. The structures of polygalacturonase inhibiting protein (PGIP) and brassinosteroid insensitive 1 (BRI1), which have PS-LRRs, are available [15,16,17].LRR-containing proteins from plants have diverse overall structures and functions. Several classes contain LRR-containing receptor-like kinases (LRR-RLKs) [18,19], LRR-containing receptor-like proteins (LRR-RLPs) [20], nucleotide binding site LRR (NBS-LRR) proteins [21,22] and PGIPs [23,24,25]. They provide an early warning system for the presence of potential pathogens and activate protective immune signaling in plants [26,27,28]. In addition, they act as a signal amplifier in the case of tissue damage, establishing symbiotic relationships and effecting developmental processes.Evolution of plant, disease resistance (R) genes that encode an LRR region has been studied by many researchers [18,22,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45]. The generations of R genes are proposed to be mainly due to gene duplication, genetic recombination, diversifying selection, sequence divergence in the intergenetic region, composition of the transposable elements, gene conversion, and unequal crossover [41,42,43].Non-LRR, island regions (IRs) interrupting LRRs are widely distributed; they are referred to as “islands” or “loop outs” [46,47]. A large number of plant LRR proteins have non-LRR IRs which are called LRR@IR proteins; they include LRR-RLKs and LRR-RLPs [46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61]. Some experimental studies on the function of non-LRR IRs within LRR@IR proteins have been performed [62,63,64]. TLRs 7, 8, and 9 out of Toll-like receptors (TLRs) are also LRR@IR proteins [65,66,67]; TLRs initiate an innate immune response [68,69,70,71]. A method—LRRpred—identify the repeat number of LRRs and phasing (that is, what segment or residue corresponds to the beginning of a repeating unit) was developed, which incorporates protein secondary structure prediction [65,72]. LRRpred predicts the repeat number and phasing of LRRs to be completely consistent with, or almost so, with those revealed by structure analyses [72]. Furthermore, to identify non-LRR IRs, a method (called LRR@IRpred) utilizing LRRpred was developed and used to find LRR@IR proteins from organisms other than plants [47]. The present article reviews 19 families of plant LRR@IR proteins identified by LRR@IRpred and describes some features of their LRR domains. The structure, function and evolution of the LRR domains as well as the LRR@IR proteins are discussed.All of the LRR domains in one protein form a single continuous structure and adopt an arc or horseshoe shape [73]. Three residues at positions 3 to 5 in the HCS, LxxLxLxxNxL or LxxLxLxxCxxL, form a short β-strand. On the inner, concave face there is a stack of the parallel β-strands and on the outer, convex face there are a variety of secondary structures such as α-helix, 310-helix, polyproline II helix, or a tandem arrangement of β-turns, which are connected by two loops. Most of the known LRR structures have caps, which shield the hydrophobic core of the first LRR unit at the N-terminus and/or the last unit at the C-terminus. In extracellular proteins or extracellular regions, the N-terminal and C-terminal caps frequently consist of Cys clusters including two or four Cys residues; the Cys clusters on the N- and C-terminal sides of the LRR arcs are called LRRNT and LRRCT, respectively [8,9,10].The crystal structures of PS-LRR domains of Phaseolus vulgaris PGIP and A. thaliana BRI1 (an LRR@IR protein) have been determined [15,16,17]. The structure of the BRI1 LRR domain forms a right-handed superhelix composed of 25 PS-LRRs (Figure 1A) [16,17]; most of these 25 PS-LRRs are 24 residues long. The helix completes one full turn, with a rise of ~70 Å. The concave surface is formed by α- and 310 helices that produce inner and outer diameters of ~30 and ~60 Å, respectively. The consensus sequence LxGx(I/L)P at positions 11 to 16 likely forms a second β-strand, which characterizes the fold of the PS-LRRs. Thus, the structural LRR units may be represented by β-β-310. BRI1 has both an LRRNT with Cx6C and an LRRCT with Cx6C; both the LRRNT and LRRCT form two disulfide bonds. The disulfide bonds contribute to the stability of the N-terminal cap structure (N-Cap) consisting of one β-strand and two α-helices and the C-terminal cap structure (C-Cap) consisting of two short helices.The crystal structures of LRR domains of A. thaliana transport inhibitor response 1 (TIR1) and coronatine-insensitive protein 1 (COI1) (that are F-box proteins) are also available [74,75,76]. TIR1 has 18 LRRs of various lengths (from 22 to 35 residues) of which 13 are noncanonical, imperfect LRRs and have long β-strands of 4–6 residues. Most VS parts adopt α-helix. Thus, the structural LRR units may be represented by β-α. The TIR1 LRR domain form a right-handed superhelix of one full turn, which is represented by one closed ring, as well as the BRI1 LRR domain [74,75]. The top surface of the TIR1 superhelix has three long intra-repeat loops (loop-2 in LRR2, loop-12 in LRR12 and loop-14 in LRR14). The loop-2 plays a pivotal role in constructing the auxin- and substrate-binding surface pocket by interacting with the nearby concave surface of the TIR1 LRR structure. The COI1 LRR domain adopts a very similar structure to that of TIR1 [76]. Similarly, three long intra-repeat loops are involved in the bindings of hormone (jasmine) and polypeptide substrates [76].Three-dimensional structures of the PS-LRR domains of BRI1 and PGIP. (A) BRI1 [3RGZ]; (B) PGIP [1OGQ]. The LRRs are colored blue, the cap structures at the N-terminal and C-terminal side orange, the non-LRR IR in BRI1 pink, and the disulfide bonds yellow. All figures were prepared with PYMOL.Plant LRR@IR proteins found through previous research by Matsushima et al. [47] and by use of keywords in the references are described. Homologs of an individual protein family from various plant species were collected by the following procedures. First, LRRs in a representative LRR@IR protein of each family were identified by LRR@IRpred; the number of repeat units in the preceding LRR block (N1), its number in the following block (N2), and the non-LRR IR sequence of the LRR region were determined. Second, database searches using the amino acid sequences of the non-LRR IR and one LRR unit at the N-terminal and C-terminal IR region were performed by FASTA at the Bioinformatic Center, Institute for Chemical Research, Kyoto University on February 15, 2012. Third, PS-LRR proteins with highly significant similarity (E-value < 10−10) were identified and then they were regarded as putative homologs in which the results of amino acid sequence alignments of full lengths and non-LRR IRs, and their domain architecture, were taken account of. Finally, LRRs in the homologs of each family were identified by LRR@IRpred. When a candidate region is not an LRR unit and its length is longer than average length of the repeating unit of LRRs, it was defined as a non-LRR IR. The following sequence analyses were also carried out: signal sequence analysis by the program SignalP (http://www.cbs.dtu.dk/services/SignalP/) [77], transmembrane predictions by TMHMM (http://www.cbs.dtu.dk/services/TMHMM/) [78], and the identification of other characteristic regions by SMART (http://smart.embl-heidelberg.de/smart/set_mode.cgi? GENOMIC = 1) [2].Finally, the 19 families of 344 LRR@IR proteins are described (Supplementary Table S1). The 19 families are grouped into LRR-RLKs, LRR-RLPs, and intracellular proteins. At least one protein in each family has clear experimental evidence for its existence or expression data (such as existence of cDNA(s), RT-PCR or Northern hybridizations) of the existence of a transcript. TMHMM predicts that A. thaliana RSYR1 and RPP27 contain a transmembrane region at the N-terminal side (Supplementary Table S1). However, orthology or domain structure was taken account of, and then these two proteins were regarded as LRR-RLKs. SignalP predicts no signal peptide in A. thaliana At1g74360 and soybean putative disease resistance protein. Similarly, these proteins were regarded as an LRR-RLK and an LRR-RLP, respectively. LRR-RLKs count 165/233/239 proteins from A. thaliana, 292/357 proteins from O. sativa subst. Japonica (rice) and 440 from Popula trichocarpa (poplar) [42,79,80]. LRR-RLPs count 90 LRR-RLPs from rice (O. sativa) and 48/56 from A. thaliana [42,46]. There are LRR-RLKs and LRR-RLPs having no non-LRR IRs, such as FLS2, Xa21, and TMM [81]. LRR- containing receptor-like cytoplasmic kinases (LRR-RLCKs) that lack an extracellular domain have no non-LRR IRs [79,82].The present review could not describe all families of LRR@IR proteins in plants because of a limited survey of LRRs having non-LRR, IRs which comes from LRR@IRpred. LRR-RLKs have an extracellular LRR region with an N-terminal signal peptide, a single transmembrane-spanning region, and an intracellular serine-threonine kinase region [18,19],. Transmembrane kinase 1 (TMK1), brassinosteroid insensitive 1 (BRI1), A. thaliana At1g74360 protein, phytosulfokine receptor (PSKR), tyrosine-sulfated glycopeptide receptor 1 (PSYR1), and LRR receptor-like serine/threonine-protein kinase RPK2 are members of theLRR-RLKs family. The LRR-RLKs are LRR@IR proteins in which the LRRs are intersected by a single non-LRR IR; only RPK2 has two IRs (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). Schematic representation of six LRR-RLKs having LRR domains intersected by non-LRR island regions. Arabidopsis thaliana TMK1 [TMK1_ARATH]; A. thaliana BRI1 [BRI1_ARATH]; Daucus carota PSKR [PSKR_DAUCA]; A. thaliana PSYR1 [PSYR1_ARATH]; A. thaliana At1g74360 [Y1743_ARATH]; A. thaliana RPK2 [RPK2_ARATH]. Nineteen families of plant LRR proteins having LRR domains intersected by non-LRR island regions. a “N1” is the repeat number of LRRs of the first LRR block in the homologs of each family. b “N2” is the repeat number of LRRs of the second LRR block in the homologs of each family. c “N1/N2” is average values. d The LRR domain in Arabidopsis RPK2 contains two non-LRR IRs. The number “13” is the sum of repeat number of LRRs of the first and second LRR blocks. The number “8” is the repeat number of the third LRR block.The transcript concentration of O. sativaTMK1 increase in the rice internode in response to gibberellins [83]. Nicotiana tabacumTMK1 mRNA accumulation in leaves was stimulated by CaCl2, methyl jasmonate, wounding, fungal elicitors, chitins, and chitosan [84]. TMK1 orthologs were identified from 14 plant species and its paralogs are present in 10 species, including A. thaliana, Glycine max, and O. sativa (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). Also G.max Rhg4, which is a soybean cyst nematode resistance gene [85], was identified as a TMK1 homolog; while G.max Rhg1 [C9VZY3] contains 13 PS-LRRs of 24 residues in which only LRR6 is 29 residues long. The TMK1 homologs contain 13 LRRs intercepted by a 56 to 76-residue, non-LRR IR. The number of repeat units in the preceding LRR block (N1) is greater than the number of the following block (N2), which means N1 » N2 with N1 = 10 and N2 = 3. The non-LRR IRs have a cluster of four Cys residues with the pattern of Cx6−7Cx29−30Cx6−11C and a conserved motif of Lx8Yx7−8WxG where “Y” is Tyr or Phe, “W” is Trp, and “G” is Gly; this motif is similar to Yx8KG found in many LRR-RLPs [46]. An LRRNT (with Cx6C) is observed, but not an LRRCT. Putative C-Cap regions are rich in Gly, Ser, and Pro residues. BRI1/SR160 is a receptor complex for brassinosteroids that are necessary for plant development, including expression of light- and stress-regulated genes, promotion of cell elongation, normal leaf and chloroplast senescence, and flowering [86,87,88,89,90,91,92]. BRI1 orthologs were identified from 24 species and its paralogs are also present in 10 species. The BRI1/SR160 homologs contain 21–26 LRRs with a single non-LRR IR. The N1 value is relatively variable among species and is 10–22, while N2 = 4; N1 » N2 (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). A. thaliana BRI1 contains 25 LRRs interrupted by a 70-residue IR between LRR21 and LRR22. The non-LRR IR, together with LRR22, binds brassinosteroids [62]. The non-LRR IRs of the BRI1 homologs are 68–70 residues long and have a cysteine cluster of Cx25−26C and have a conserved motif of R(I/V/M/L)Y. An LRRNT (with Cx6C) and an LRRCT (with Cx6C) were observed; only soybean BR [C6ZRS8] and Ricinus communis LRR-RLK [B9T4K2] have LRRNTs with Cx26−27C. The LRRCT regions are rich in His, Arg, and Lys residues, and thus are basic.PSKR is a PSK receptor that regulates, in response to PSK binding, a signaling cascade involved in plant cell differentiation, organogenesis, and somatic embryogenesis [55,63,93,94]. PSKR orthologs and paralogs were identified (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). The PSKR homologs contain LRRs with a 36 to 38-residue, non-LRR IR. N1 = 17 − 18 and N2= 4 (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). The non-LRR IRs have a conserved motif of (Y/F)x5−12Yx5F. Most LRRCT regions are basic. Daucus carota PSKR contains 22 LRRs intersected by a 36-residue IR between LRR17 and LRR18. An LRRNT (with Cx33CCx6C) that is similar to that in PGIP [15] and LRRCT (with Cx6C) are observed. A 15-residue region within the non-LRR IR is a binding site of PSK [63]. The corresponding regions in the homologs are relatively variable.A. thaliana RSYR1 regulates, in response to tyrosine-sulfated glycopeptide binding, a signaling cascade involved in cellular proliferation and plant growth [95]. The RSYR1 homologs from seven species contain 21–22 LRRs with a 37-residue, non-LRR IR (N1 = 17 − 18 and N2= 4) (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). The non-LRR IRs have a conserved motif of Yx2LPVFx4Nx4Qx2−3QLSxL. The LRRNT (with four, five, or seven Cys residues) and the LRRCT (with Cx7C) are observed. The LRRCT regions are basic. A. thaliana At1g74360 is a BRI1-related protein (Figure 2 and Table 1, and Supplementary Table S1 and Figures S1). Putative orthologs and paralogs were identified from 10 species. The At1g74360 family contains 21–22 LRRs with a single IR. The N1 value is relatively conserved among species; N1 = 16 − 17, while not N2= 4 but N2= 5. The non-LRR IRs of 76-residue are longer than those in BRI1 and have a cysteine cluster with the pattern of Cx25Cx16C. The IRs are highly conserved among the homologs. A. thaliana RPK2 is a key regulator of anther development (e.g., lignifications pattern), including tapetum degradation during pollen maturation (e.g., germination capacity) [96,97,98] and contributes to shoot aptical meristerm homeostasis [99,100]. The RPK2 homologs from Arabidopsis lyrata subsp. Lyrata, Populus trichocarpa, and R. communis contain 21–22 LRRs with two non-LRR IRs. The first IR is between LRR9 and LRR10. The second IR is between LRR13 and LRR14 (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). The second IRs are highly conserved among homologs. There are an LRRNT (with Cx6Cx25Cx16C) and an LRRCT (with Cx11C). The LRRCT region is rich in Ser and Pro residues. Sawa and Tabata [101] have reported the RPK2 homologs from other plant species-Musa acuminate, O.sativa Japonica Group, Vitis vinifera, Sorghum bicolor, Physcomitrella patens, and Marchantia polymorpha.LRR-RLPs have a short cytoplasmic tail instead of the kinase region in LRR-RLKs (Figure 3) [20]. LRR-RLPs are involved both in resistance of plant–pathogen interactions and development [34,102]. Tomato Cf genes confer resistance to the fungal pathogen Cladosporium fulvum [43,56,103,104]. Tomato Verticillium wilt disease resistance gens (Ve1)and Ve2, apple HcrVf2, Arabidopsis RPP27 are involved in resistance to Verticillium, Venturia, and Peronospora, respectively [105,106,107]. Furthermore, the tomato LeEIX initiates defense responses upon elicitation with a fungal ethylene-inducing xylanase (EIX) of non-pathogenic Trichoderma from tomato that confer resistance against the fungal pathogen Cladosporium fulvum [108,109]. The clavata2 (CLV2) functions in both shoot and root meristems of Arabidopsis  [58,110,111,112] and also affects autoregulation of nodulation of pea and Lotus japonicus [113,114]. Zea maysfascinated ear2 is involved in meristem development [59]. A. thaliana RLP2 is involved in the perception of CLV3 and CLV3-like peptides, that act as extracellular signals regulating meristems maintenance [64]. The LRR-RLPs are all LRR@IR proteins in which the LRRs are intersected by a single non-LRR IR (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1).Schematic representation of 11 LRR-RLPs having LRR domains intersected by non-LRR island regions. Currant tomato Cf-9 [Q40235I]; Currant tomato Cf-2.1 [Q41397]; Tomato Ve1 [Q94G61]; Apple HcrVf1 [Q949G9]; A. thaliana RPP27 [Q70CT4]; Tomato EIX1 [Q6JN47]; A. thaliana CLV2 [Q9SPE9]; Maize fascinated ear2 [Q940E8]; A. thaliana RLP2 [RLP2_ARATH]; Oryza sativa Os10g0469700 [Q337L7]; Soybean disease resistance protein [C6ZS07].Tomato Cf-9/Cf-4 homologs were identified from six species. Elicitor-inducible LRR receptor-like protein (EILP) from N. tabacum  [115] was identified as ortholog of tomato Cf-9/Cf-4. The number of N1 is 18 to 22, while N2 keeps 4, and the non-LRR IRs are 40–44 residues long (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1) and have a conserved motif of MKx3Ex6Yx5−8Yx7TKG in which hydrophilic residues are conserved. The EILP protein also contains 27 LRRs with N1 = 23 and N2 = 4. Most of the homologs have LRRNT consisting of six Cys residues with the pattern of Cx24−29Cx13−23CCx6Cx12−13C. However, peru 1 and peru 2 have an LRRNT of four Cys’s with the pattern of Cx47CCx6C [116]. The C-terminal side of the LRRCT is rich in Glu and Asp residues and thus is acidic. Tomato Cf-2/Cf-5 homologs were identified from two species (Lycopersicon esculentum, and L. pimpinellifolium). The number of N1 is highly variable; N1 = 20 − 33, while N2 keeps 4, and the non-LRR IRs are 37–41 residues long. The IRs are hydrophilic. The variability of N1 has been reported by other researches in between the paralogs and orthologs [43,46,103,104] (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). Interestingly, the N-terminal LRRs include tandem repeats of the super-motif of two highly conserved LRRs; for example, LxxLxLxxNxLSGxIPxxIGYLRS and LxxLxLSxNxLNGxIPxxFGxLxN in currant tomato Cf-2.1 [103].Tomato Ve orthologs and paralogs were identified from twelve species including Solanum neorickii, S. aethiopicum, Mentha longifolia, and M. spicata [105,117,118]. The Ve homologs contain 32–34 LRRs intercepted by a 44 to 49-residue, non-LRR IR with N1 = 28 − 30 and N2 = 4(Figure 3 and Table 1, and Supplementary Table S1 and Figure S1).The non-LRR IRs have a conserved motif of YYx8K(G/R) and are relatively hydrophilic.Apple HcrVfs (Homologs of Cladosporium fulvum resistance genes of Vf region) are scab resistance genes [119,120]. Mentha longifoliaHcrVfs are orthologs of tomato Ve genes [105,117,118]. The HcrVfs paralogs contain 32–34 LRRs intercepted by a 41 to 46-residue, non-LRR IR with N1 = 22 − 28 and N2 = 4 (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). The non-LRR IRs have a conserved motif of VTKGxExEYx(K/E)ILxFxKxxDLSCNF in which hydrophilic residues are conserved. The C-terminal side of the LRRCT is rich in Gly and Pro residues. A. thaliana RPP27 homologs were also identified from A. lyrata. The LRR@IR proteins contain 16–30 LRRs intercepted by a 61 to 71-residue, non-LRR IR with N1 = 11 − 26 and N2 = 4 (Figure 2 and Table 1, and Supplementary Table S1 and Figure S1). The IRs have a conserved motif of FxxKxRYD. The C-terminal side of most LRRCT regions is acidic. Tomato LeEIX1 and LeEIX2 contain 31 LRRs intercepted by a 47 to 49-residue, non-LRR IR with N1 = 27 and N2 = 4 (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). The C-terminal side of the LRRCT is acidic.A. thaliana CLV2 homologs were identified from 11 species. The CLV2 homologous proteins contain 22 LRRs intercepted by a 41 to 43-residue, non-LRR IR with N1 = 18 and N2 = 4 (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). The IRs have a conserved motif of LxFxYxL. The C-terminal side of most LRRCT regions is acidic. A. thaliana CLV1 is an LRR-RLP but not LRR@IR protein. Z. mays fascinated ear2 is an ortholog of Arabidopsis CLV2. The homologs were also identified from O. sativa subsp. Japonica, and indica, and S. bicolor. The fascinated ear2 homologous proteins contain 17–18 LRRs intercepted by a 41 to 42-residue, IR with N1 = 10 − 14 and N2 = 4 (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). The IRs and the LRRCT regions are rich in Gly. Both regions may be flexible. A. thaliana RLP2 contains 23 LRRs that are intercepted by a 44-residue, IR with N1 = 18 and N2 = 4 (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). There are an LRRNT and an LRRCT. The extracellular region including the 23 LRRs is homologous to that in A. thaliana PSYR1 [121].O. sativa Os10g0469700 is an LRR@IR protein; the function is unknown (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). The homologs from four species contain 10 LRRs with a single IR with N1 = 6 and N2 = 4. The non-LRR IRs with 39–40 residues is represented by MKxP(K/E)IxSSx2−3LDGSxYQDRIDIxWKGx3FQx4L.A putative disease resistance protein from soybean [C6ZS07] is an LRR@IR protein (Figure 3 and Table 1, and Supplementary Table S1 and Figure S1). The homologs were identified from four species and contain 8–32 LRRs with a single IR with N1 = 4 − 28 and N2 = 4. The N1 number is highly variable in both the paralogs and orthologs. The IRs have a conserved motif of Yx2Sx5Kx7(R/K)I.A. thaliana TONSOKU(TSK)/MGOUN3(MGO3)/BRUSHY1(BRU1), which is localized in the nucleus and is preferentially expressed in the shoot apex than in the leaves and stems, is required for cell arrangement in root and shoot apical meristems and involved in structural and functional stabilization of chromatin [122,123,124]. The TONSOKU protein may represent a link between response to DNA damage and epigenetic gene silencing [125]. Potential homologs of A. thaliana TONSOKU have been identified in eight species. The UniProKB database describes that A. thaliana TONSOKU contains three LRRs and eight TPRs, while the data bases - InterPro, Gene3D, SMART and PROSITE-identify only TPR. LRR@IRpred identifies 14 LRRs with a single IR; N1 = 13, N2 = 1 (Figure 4 and Table 1, and Supplementary Table S1 and Figure S1) [47]. The LRRs are not “plant-specific” motifs but presumably “RI-like” motifs. Thus, the structural LRR units may be represented by β-α instead of β-β-310. The LRR domain is predicted to adopt a typical horseshoe shape seen in ribonuclease inhibitor [126]. The non-LRR IRs are 70–131 residues long and are rich in Ser and Gly. The IRs may be unstructured or flexible. A. thaliana  MJK13.7 is considered to be intracellular protein. The function is unknown. A. thaliana MJK13.7 homologs were identified from 11 species. The homologs contain 20 LRRs intersected by a single IR; N1= 12, N2 = 8 (Figure 4 and Table 1, and Supplementary Table S1 and Figure S1). All of the non-LRR IRs are 60–62 residues long and have conserved Lys residues at five positions. The consensus of the LRRs is LxxLxLxxNxLxxLPxxLxxLxx of 23 residues that are present in many proteins from bacteria to human (data not shown). The LRR motif does not belong to PS-LRR and the structure of the LRR domain is not available. However, the LRR motifs are contained in part of the LRR domains in toll-like receptor 1 (TLR1) and glycoprotein Ibα (GpIbα) of which the crystal structures are available [127,128,129,130]. Four LRRs are IKVLDLHSNKI KSIPKQVVKLEA and LQELNVASNQL KSVPDGIFDRLTS in TLR1, and LGTLDLSHNQL QSLPLLGQTLPA and LDTLLLQENSL YTIPKGFFGSHL in GpIbαe. The structures revealed that the LRRs may be characterized by extended conformations at the bold sequences [127,128,129,130]. Schematic representation of two plant intracellular LRR@IR proteins having LRR domains intersected by non-LRR island regions. A. thaliana MJK13.7 [Q9M7W9]; A. thaliana TONSOKU [Q6Q4D0].Moreover, A. thaliana MJK13.7 forms a family with its homologs from insect species, Strongylocentrotus purpuratus, Nematostella vectensis, and Paramecium tetraurelia and LRRC40 from vertebrates species [47]. The S. purpuratus protein has 163 residues containing two repeats of 64 residues each [47]. Rice blast resistance gene Pi-ta encodes an NBS-LRR protein with 928 residues [44,45]. The Pi-ta protein [Q9AY26] lacks a canonical LRR [44]. The C-terminal region contains highly imperfect LRRs with 10 repeats of various lengths (from 16 to 75 residues) based on the consensus LxxLxxL. The Pi-ta protein appears to be an LRR@IR protein. LRR@IRpred predicts 13 LRRs of 20–54 residues with one non-LRR, IR between LRR6 and LRR7(Supplementary Figure S1). The secondary structure prediction prefers α-helix in the VS’s. The Pi-ta LRR domain might adopt a similar structure to those of TIR1 and COI1 [74,75,76]. Most plant LRR@IR proteins that are LRR-RLKs or LRR-RLPs keep the rule of N1 » N2;N1 = 10 − 30 and N2 = 3 − 5 (Table 1). The same rule of N1 » N2is observed in other LRR@IR proteins of toll receptors and toll-related proteins from insect species, that have one single transmembrane-spanning region and an intracellular Toll IL-receptor (TIR) domain as well as TLRs instead of the kinase region in LRR-RLKs [131]. Most toll receptors and toll-related proteins contain 21–30 LRRs interrupted by a single non-LRR IRs of 81–120 residues with N1 » N2; N1 = 17 − 24 and N2 = 4 − 6 (data not shown). Fritz-Laylin et al. [46] have performed sequence analysis of 90 LRR-RLPs of rice (O. sativa) and 56 Arabidopsis (A. thaliana). Many LRR-RLPs contain 18–28 LRRs intercepted by a 30 to 80-residue, single IR with N1 » N2; N1 = 14 − 24 and N2 = 4 [46]. The non-LRR IRs in plant LRR@IR proteins may be classified into two groups; one group is non-LRR IRs having cysteine clusters, while the other has no cysteine clusters. The IR cysteine clusters are characterized by Cx6−7Cx29−30Cx7−11C in A. thaliana TMK1 homologs, Cx25C in BRI1 homologs, and Cx25Cx16C in At1g74360 homologs. The other non-LRR IRs frequently have a conserved motif of Yx8KG which are observed in the homologs of A. thaliana TMK1, tomato Cf-9/Cf-4, tomato Cf-2/Cf-5, tomato Ve, M. longifolia HCrVf, A. thaliana CLV2, and Z. mays fascinated ear2, and O. sativa Os10g0469700. Non-LRR IRs in many LRR-RLPs from Arbidopsis and rice contain a conserved motif of Yx8KG [46]. Most of the LRRNTs consist of two, four, or six Cys residues of which the patterns are Cx6−7C,Cx23−34CCx6C, and Cx24−29Cx13−23CCx6Cx12−13C. They probably form one, two, and three disulfide bonds, respectively. The LRRCTs consist of two Cys’s with the pattern of Cx4−29C which probably form one disulfide bond (Supplementary Table S1 and Figure S1). The disulfide bonds should contribute to the structural stabilization of the N-terminal and C-terminal caps.The structure of a non-LRR IR is available in A. thaliana BRI1 (Figure 1A). The BRI1 LRR domain forms a superhelix with 25 LRRs. The 70-residue, non-LRR, IR in BRI1 between LRR21 and LRR22 forms a small domain that folds back into the interior of the superhelix, where it makes extensive polar and hydrophobic interactions with LRRs 13–25 [16,17]. The LRR domain fold is characterized by an anti-parallel β-sheet, which is sandwiched between the LRR core and a 310 helix and stabilized by a disulphide bridge of the Cys cluster with Cx25C in the non-LRR, IR. Cys clusters are also present in non-LRR, IRs in the homologs of TMK, At1g74360 and TONSOKU. Thus, the non-LRR IRs may adopt similar structures with disulfide bridges. All of the non-LRR IRs would fold back into the interior or exterior of a superhelix of the LRR domains.The non-LRR IRs of BRI1 and PSKR participate in ligand/protein-protein interactions. The BRI1 non-LRR IR binds brassinosteroids [62]. The insertion of a folded domain into the LRR repeat is probably an adaptation to the challenge of sensing a small steroid ligand [16]. The PSKR non-LRR IR also binds PSK [63]. The non-LRR IRs in TLRs 7, 8, and 9 was also predicted to contribute to nucleic acid-protein interaction [66,132]. The non-LRR IRs in plant LRR@IR proteins have frequently conserved motifs that are characterized by hydrophilic residues such as Lys, Arg, Glu and Asp, as noted. Some non-LRR IRs are presumably flexible. The conservation of hydrophilic residues in the IRs is also observed in the respective families of LRRC40, LRRC9, and C. elegans LRK-1 which are LRR@IR proteins from organisms including vertebrate other than plants [47]. The IRs might contribute to ligand/protein-protein interactions [47]. Moreover, Afzals et al. [133] suggested, based on circular dichroism data, that non-LRR IRs are intrinsically unstructured, providing binding diversity to the domains.The first LRR block in tomato Cf-9, Cf-4, and Cf-2 recognize fungal avirulence proteins [134,135,136,137,138]. The recognitional specificity of Cf-2 with 37 LRRs lies between leucine-rich repeat LRR3 and LRR27, a region that differs from Cf-5 with 31 LRRs by six extra LRR and 78 amino acid substitutions [134]. Although crudely defined, this region of specificity corresponds to those in Cf-4, Cf-9, and Cf-9B responsible for recognition of their cognate ligands [135,136,137,138]. Biochemical studies show that CLV2 is essential for the stability of CLV1, in which CLV1 and CLV2 may form a disulfide-linked heterodimer of 185 kD [58]; CLV1 is an LRR-RLP having no non-LRR IR.Drosophila Toll and vertebrate TLRs 7, 8, and 9 are LRR@IR proteins [65,66,67] which contain one single transmembrane-spanning region as well as LRR-RLKs and LRK-RLPs from plant. Homo- or heterodimerization are involved in ligand-interactions of vertebrate TLRs [68,69,70,71]. A model for DrosophilaToll activation by ligand Spatzle has been proposed; the first LRR block interacts with Spatzle and the second LRR block forms strong dimer contacts that are prevented by the first block, which in the absence of ligand provides a steric constraint [67,131]. The BRI1 receptor activation involves homodimerization [139]; although Hothorn et al., [16] suggested that the superhelical BRI1 LRR domain alone has no tendency to oligomerize, indicating that BRI1 receptor activation may not be mediated by ligand-induced homodimerization of the ectodomain. Taken together, non-LRR IRs in plant LRR@IR proteins might participate in ligand/protein-interactions, dimerization or both, although an LRR-RLP, A. thaliana CLV2, remains functional without non-LRR IR, while the first and the second LRR blocks are essential for functionality [64]. N1 » N2 brings close proximity of the non-LRR IRs to interact with ligand/protein and a transmembrane region. N1 » N2might facilitate signaling in the cytoplasm through the ligand/protein- interactions. There is a possibility that Cys residues in LRRs are involved in dimerization of LRR@IR proteins. The conserved hydrophobic residues of the PS-LRR consensus sequence of LxxLxLxxNxLSGxIPxxLxxLxx at positions 1, 4, 6, 11, 15, 19, and 22 contribute to the hydrophobic cores in the LRR arcs [8,9]. The conserved hydrophobic residues at positions 1, 19 and 22, and “N” at position 9, are frequently occupied by Cys in the PS-LRRs. Moreover, Cys residues are frequently observed in noncanonical PS-LRRs which, as examples, are longer LRR motifs of 25–30 residues with the consensus of LxxLxLxxNxLSGxIPxxLCxxxxx(x/-)(x/-)(x/-)(x/-)(x/-), in which “-” indicates a possible deletion site. At the present stage it remains unknown whether the Cys residues contribute to the hydrophobic core of the LRR arcs or are exposed to solvent. However, some LRR@IR proteins contain PS-LRRs having Cys at positions 2, 3, or 5 in the HCS part (Supplementary Table S1). The Cys residues are likely to be exposed to solvent in the LRR arc and thus might induce dimerization.What is the evolutionary origin of non-LRR IRs interrupting LRRs? Previous research provided evidence that a direct duplication of the super motifs containing non-LRR regions naturally leads to the occurrence of non-LRR IRs in LRR@IR proteins, including LRR-containing 17 protein (LRRC17), LRRC32, LRR33, chondroadherin-like protein, trophoblast glycoprotein precursor, and Leishmania proteophosphoglycans, not from plants but from other eukaryotes [47]. The non-LRR IRs in plant LRR@IR proteins might originate from such similar events. The tomato Cf-2/Cf-5 homologs have PS-LRRs that include tandem repeats of the super-motif of two highly conserved LRRs, as noted [103]. The duplications of the super-motif were suggested to have occurred in the Cf-2/Cf-5 homologs [43]. Super-motifs of LRRs are observed in many LRR proteins. The SLRP subfamily (biglycan, decorin, asporin, lumican, fibromodulin, PRELP, keratocan, osteoadherin, epiphycan, osteoglycin, opticin, and podocan), the TLR7 family (TLR7, TLR8 and TLR9), the FLRT family (FLRT1, FLRT2, and FLRT3), and OMGP [65,140,141] contain tandem repeats of a super-domain of STT, where “T” is “typical” LRR and “S”is“Bacterial” LRR. Ribonuclease inhibitor also has RI- LRRs consisting of a super-motif of 57 residues that encode two LRRs [142]. The super-repeats as well as Cf-2/Cf-5 have been contributed to the duplication of their super-motifs.A large number of LRR-RLPs resembling the extracellular domains of LRR-RLKs are found in the Arabidopsis genome; although not all RLK subfamilies have corresponding RLPs [121]. Indeed, the present analysis indicates that the extracellular domain in PSYR1 is highly similar to that in RLP2. The same distributions also occur in LRR@IR proteins from other plants, such as S. bicolor and O. sativa (Supplementary Figure S2). Here four examples are described: Sb10g028170/Sb10g028210 (LRR-RLK/LRR-RLP), and Os06g0691800/Os06g0692700; all the four proteins contain 22 LRRs intersected by  a single non-LRR IR of 33 residues with N1 = 18 and N2 = 4. The others are Os07g0597200/Os03g0400850, and OsI_26735/OsI_11946; the LRR-RLKs-Os07g0597200 and OsI_26735 are homologs of Arabidopsis At1g74360. The pair-wise comparisons of the amino acid sequences exceed 50% of the identity in respective pairs. The above observations indicate that the LRR-RLKs and LRR-RLPs evolved from gene duplications and recombination [39].Two putative uncharacterized proteins from Z. mays with 717 residues [B8A2X8] and with 623 residues [B8A383_MAIZE] are paralogs of Z. mays TMK1 with 958 residues (Supplementary Figure S2). The 717-residue protein contains 6 LRRs; N1= 3 andN2 = 3. There are other examples; a hypothetical protein from Z. mays with 247 residues [C0PL86] and fasciated ear2 with 613 residues, O. sativa Os02g0782800 with 441 residues [Q6K7E5] and BRUSHY1 with 1,332 residues [Q6K7D3]. The occurrence of these proteins is attributed to gene duplication and deletions. Most plant LRR@IR proteins have LRRs intersected a single IR with N1 » N2 in which N1 is variable in their individual homologs, while N2 is highly conserved. For all known LRR-RLPs, N1 = 4. The rule of N1 » N2 plays a common, significant role in ligand-interaction, dimerization, and/or signal transduction of the LRR-RLKs and the LRR-RLPs. All of the LRR domains consisting of PS- LRRs are predicted to form a superhelix and non-LRR IRs in plant LRR@IR proteins fold back into the interior or exterior of the superhelix. The present analyses suggest that some LRR-RLKs and LRR-RLPs evolved from gene duplications and recombination. The present review will stimulate various experimental studies to understand the structure and evolution of the LRR domains with non-LRR IRs and their proteins.We thank Robert H. Kretsinger of the University of Virginia for his valuable suggestion and comments. This study was supported by a grant from a Grant-in-Aid for Scientific Research (C) No. 23500368 from the Japan Society for the Promotion of Science (to N. M).

Med-MDPI/biomolecules/biomolecules-02-03-00312.txt

@@ -0,0 +1 @@
+licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Post-translational modifications are able to regulate protein function and cellular processes in a rapid and reversible way. SUMOylation, the post-translational modification of proteins by the addition of SUMO, is a highly conserved process that seems to be present in modern cells. However, the mechanism of protein SUMOylation in earlier divergent eukaryotes, such as Giardia lamblia, is only starting to become apparent. In this work, we report the presence of a single SUMO gene encoding to SUMO protein in Giardia. Monoclonal antibodies against recombinant Giardia SUMO protein revealed the cytoplasmic localization of native SUMO in wild-type trophozoites. Moreover, the over-expression of SUMO protein showed a mainly cytoplasmic localization, though also neighboring the plasma membrane, flagella, and around and even inside the nuclei. Western blot assays revealed a number of SUMOylated proteins in a range between 20 and 120 kDa. The genes corresponding to putative enzymes involved in the SUMOylation pathway were also explored. Our results as a whole suggest that SUMOylation is a process conserved in the eukaryotic lineage, and that its study is significant for understanding the biology of this interesting parasite and the role of post-translational modification in its evolution. Giardia lamblia is one of the most prevalent parasitic protozoan in developing countries, causing an intestinal pathology known as giardiasis, which in many cases produces diarrhea and nutrient malabsorption in humans [1,2]. It has a simple life cycle with two major stages: infectious cysts and trophozoites [2], which have specific mechanisms enabling them to adapt to their environment [3]. These mechanisms involve the preferential expression of genes and proteins to allow parasite survival and the transmission of the pathology to susceptible hosts. Although its phylogenetic position in the eukaryotic lineage is controversial at the moment, Giardia is considered an early divergent eukaryote in evolution and possesses unusual features, such as the presence of two transcriptionally active diploid nuclei and the absence of mitochondria and peroxisome [4], which make this an attractive model to study the evolution of regulatory systems. Post-translational modifications are one of the most effective ways by which evolution has increased versatility in protein function, providing the cell with the flexibility to respond to a broad range of stimuli [5,6]. These modifications are essential and reversible mechanisms by which the functions, activities, and stabilities of preexisting proteins can be rapidly and specifically modulated, thereby controlling dynamic cellular processes [7]. Interaction with Small Ubiquitin-like Modifier (SUMO) is, in particular, one of the most complex, conserved, and interesting characteristic mechanisms of protein regulation in eukaryotes, with diverse targets and functions such as nuclear transportation, transcriptional regulation, maintenance of genome integrity, and signal transduction [6,8,9].SUMO belongs to the ubiquitin-like protein family (Ubl), displaying a three-dimensional structure similar to ubiquitin, although it shares only 18% identical amino acids and differs in the distribution of charged residues on the surface [5,8]. Like ubiquitin, SUMO is expressed as a precursor protein and requires a maturation process, by specific SUMO proteases (SENPs) (Figure 1), to expose the carboxy-terminal double-glycine motif (GG) required for conjugation to substrate proteins [10]. SUMO is covalently attached to target proteins, via an isopeptide bond between a C-terminal glycine of SUMO and a lysine residue within the consensus sequence defined by ψKXE (where ψ is a large hydrophobic amino acid, K is the lysine to which SUMO is conjugated, X is any amino acid, and E is glutamic acid residue) [8,11]. As an ubiquitination process, conjugation to SUMO involves an enzymatic cascade, which includes an E1-activating enzyme, an E2-conjugating enzyme, and sometimes the assistance of a ligase that increases the efficiency of transferring to substrate [12,13]. Unlike the ubiquitin E1 enzyme, which functions as a single subunit enzyme, the SUMO E1 enzyme consists of a heterodimer of two polypeptides known as SUMO Activation Enzyme 1 and 2 (SAE1 and SAE2) [5]. SAE1 contains a single domain that adenylates SUMO and is homologous to the N-terminal portion of the ubiquitin E1 enzyme [5,14]. SAE2 is homologous to the C-terminal portion of the ubiquitin E1 enzyme and mediates exclusively the E1–SUMO interaction [5,15,16].The SUMO conjugation pathway. SUMO is expressed as an inactive propeptide and is processed by a SUMO-specific protease (SENP) to expose the C-terminal GG, required by the SUMO conjugation to targets (maturation). Mature SUMO is activated by the SUMO activating enzyme (E1) and is transferred through a transesterification process to Ubc9 (E2). SUMO is next conjugated to the target lysine of a substrate, defined by the consensus motif ΨKXE. E3 ligase enzyme can facilitate this process. Specific proteases can remove SUMO from modified substrates maintaining the reserve of free SUMO. Gene ID corresponding to homologous Giardia enzymes involve in the SUMOylation process is depicted in green. Modified from [10].Through a transesterification reaction, activated SUMO is subsequently transferred to the catalytic cysteine of the unique SUMO conjugating (E2) enzyme, Ubc9 [17] which, in contrast to ubiquitin conjugating enzymes, has the ability to recognize target proteins directly and catalyze the formation of an isopeptide bond between the C-terminal glycine of SUMO and the ε-amino group of a target lysine [18]. Consistent with structural studies showing direct recognition of this consensus motif by the Ubc9 active site, recombinant E1, E2, and SUMO are sufficient for ATP-dependent SUMO modification of substrates in vitro [8,18]. SUMOylation is a dynamic and reversible process, and requires SENPs to remove SUMO conjugates from substrates, maintaining the reserve of the free SUMO form [8]. These proteins in general are able to play both roles in SUMO regulation: cleaving the isopeptide bond between SUMO and its substrate, and processing SUMO precursors to mature forms [5,18]. Despite the extensive proteomic analyses of SUMOylated proteins that have been conducted in several eukaryotic cells including parasites such as Toxoplasma gondii [19], Plasmodium falciparum [20]and Trypanosoma [21,22], the SUMOylation system of Giardialamblia has not been investigated until now. Previously, we demonstrated that the enzyme arginine deiminase is a SUMOylated protein [23] and this is so far the only evidence of SUMOylation in Giardia. Therefore, the identity, function, and regulation of SUMO proteins and the cellular processes they regulate are largely unknown. In the present work, we disclosed the presence of a single gene in Giardia (gsumo) that codifies to SUMO protein (gSUMO) and identifies gene encoding to putative enzymes of the SUMOylation pathway. To understand the role of SUMOylation in this parasite, we over-expressed gSUMO in transgenic trophozoites and produced monoclonal antibodies against recombinant SUMO protein (GST-gSUMO). These tools allowed us to reveal the presence of several putative SUMOylated proteins, suggesting SUMO conjugation as a functional system in Giardia and formerly in evolution.By searching in the GDB, we found that the G. lamblia genome encodes only one putative SUMO gene (ID GL7760). This matches what was reported for Saccharomyces cerevisiae [16,24,25], invertebrates [5,18] and other parasites like Plasmodium falciparum [20], Toxoplasma gondii [19], Trypanosoma brucei [22,26] and Trypanosoma cruzi [21] but is dissimilar to what was found in Arabidopsis thaliana [27] and mammals [10,16], where multiple members of the SUMO family are present. ID GL7760 codes for a putative gSUMO protein of 102 amino acid residues, with a predicted molecular mass of 12 KDa. Comparative analysis of SUMO proteins. (A) Multiple sequence alignment was constructed between the conserved domain of gSUMO and other 13 SUMO sequences, plus Human Ubiquitin. Dark shading show identical residues, while light shading shows similar residues between the respective sequences. Red asterisks indicate conserved GG motifs. (B) Structural alignment was constructed with alfa-helix (α) and beta-sheet (β) portions of the predicted 3D structure of gSUMO and the Chain A of the Protein Data Bank file 2K8H from T. brucei (TbSUMO).Multiple Sequence Alignment (MSA) of conserved regions of known SUMO family members, showed that gSUMO possesses the characteristic C-terminal region, which includes the conserved GG motif required for isopeptide bond formation [Figure 2(A)]. Sequence comparison also demonstrated that gSUMO is somehow homologous to the SUMO-1 member of the human SUMO family. The three-dimensional model of gSUMO was obtained using Phyre2, based on crystal structures of human SUMO-1 (PDB code 1A5R). Among the available PDB structures, human SUMO-1 was the one that shared higher sequence identity with gSUMO, revealing also the three-dimensional structure alignment of more conserved regions, high overlap between predicted gSUMO and the crystal structure of SUMO from Trypanosoma brucei (PDB code 2K8H) [Figure 2(B)]. All this suggests that SUMO is a highly conserved protein among all organisms. In order to characterize the SUMOylation process in Giardia lamblia, we produced WB-strain transgenic trophozoites that stably expressed SUMO, containing three hemagglutinin (HA) epitope sequences in the N-terminus (HA-gSUMO). Over-expression of HA-gSUMO in trophozoites under the constitutive tubulin promoter revealed, by immunofluorescence assays (IFA) in epifluorescence microscopy, mainly cytoplasmic localization but nuclear localization in some cells (data not shown).However, HA-gSUMO showed a more variable localization pattern (from underneath the plasma membrane to the nuclei) considering different focal planes of confocal microscopy (Figure 3A), which might suggest the presence of different putative substrates by SUMO. SUMOylation is known to target mainly nuclear proteins, although several studies point to many roles of SUMO in the soluble phase of the cytoplasm, the plasma membrane, mitochondria and the endoplasmic reticulum, documenting the presence of putatively SUMOylated proteins in both nucleus and cytoplasm, depending on the process being regulated [8,15,28].Western blot analysis of HA-gSUMO transgenic trophozoites enabled the free SUMO form (~20 kDa), as well as many bands, to be observed (Figure 3B). Although SUMO has a molecular mass of approximately 11 kDa, it appears larger on SDS-PAGE and adds ~20 kDa to the apparent molecular weight of most substrates [29]. The other bands that range from ~50 kDa to 85 kDa might correspond to SUMOylated proteins. After cell fractionation and comparison with what we found in the IFA assays, these proteins were observed in both cytoplasmic and nuclear fractions (data not shown).It is known that identification of SUMOylated proteins is not simple, for several reasons: (i) many SUMOylated proteins are present at a level below the normal detection limits [25,30], (ii) for most SUMO target proteins, only a small fraction of the substrate is SUMOylated at any given time, and (iii) there are strong SUMO protease activities in native cell lysates [25]. The list of newly discovered SUMO substrates is expanding only recently in other model systems, whereas only the arginine deiminase enzyme has been identified as a SUMOylated substrate in G. lamblia [23]. To analyze stage-specific dynamics of the protein SUMOylation in more detail, we generated a novel set of monoclonal antibodies against a recombinant GST–gSUMO protein. Mouse antisera against fusion protein were analyzed for specificity by dot-blot using GST–gSUMO, and IFA and dot-blot using total protein from wild-type trophozoites (data not shown). Mice showing the strongest positive reaction were later sacrificed and utilized to produce monoclonal antibodies (mAbs). Although several mAbs recognized the SUMO protein, we chose the clone 13C5, which showed the strongest reactivity in Western blot and in IFA assays.SUMO Over-expression. (A) HA-gSUMO localizes in both cytoplasm and nuclei in transgenic cells. IFA using anti-HA mAb and confocal microscopy shows that HA-gSUMO (red) has a variable pattern of localization from underneath plasma membrane and cytoplasm to nuclei. DIC (Differential Interference Contrast microscopy). Nuclei are stained with DAPI (blue). Scale bar: 10 µm. (B) Several putative SUMOylated substrates can be detected by Western blotting. Western blotting using anti-HA mAb shows a ~20 kDa band, likely corresponding to free SUMO form, plus bands that range from 50 kDa to 85 kDa that might correspond to SUMOylated substrates in HA-gSUMO transgenic trophozoites. Lane 1: Standars of the indicated molecular weights. Western blot assays using homogenate of wild-type or gSUMO-transgenic trophozoites showed that this mAbs was able to recognize endogenous gSUMO in a similar pattern of bands to the one observed for transfected cells (Figure 4A). Similarly, IFA using the mAbs showed that the localization of gSUMO was cytoplasmic in wild-type, and concentrated around nuclei, plasma membrane and flagella in HA-gSUMO transgenic trophozoites. The difference of pattern localization between wild type and transgenic trophozoites is probably due to the increased expression of the protein in HA-gSUMO cells. Given the homology between gSUMO and SUMO-1, we analyzed in a eukaryotic CHO cell line the cross-reaction of the mAbs generated against the SUMO protein. We found a clear cytoplasmic subcellular localization, similar to the localization observed in Giardia wild-type trophozoites (Figure 4B). This result reinforces the feature of SUMO as a conserved protein in evolution. gSUMO antibody reactivity. (A) gSUMO mAb is able to recognize endogenous gSUMO. Western blotting of wild-type (WT) and HA-SUMO transgenic Giardia trophozoites (TT) show that the 13C5 mAb recognizes endogenous gSUMO (arrow) and several bands (from 50 kDa to 120 kDa) likely corresponding to putative SUMOylated protein. Lane 1: Standards of the indicated molecular weights. (B) IFA and confocal microscopy using the 13C5 mAb in permeabilized wild-type trophozoites showed a main pattern of gSUMO localization (green) in the cytoplasm while in and HA-gSUMO transgenic trophozoites the label is in the cytoplasm (botton panel) and concentrated close to the plasma membrane (upper panel). In CHO cells, a cytoplasmic localization is observed using the 13C5 mAb. Scale bar: 10 µm. The SUMOylation pathway is a reversible process that creates an on and off state, which is essential for biological regulation. Besides the variation in sumo gene complexity in different species, the SUMOylation and de-SUMOylation components are conserved in most of the eukaryotes where this process was described [5]. Although the SUMOylation system appears to be a functional pathway for protein modification in Giardia, only the gene that encodes to putative gSUMO Protease (gSP: ID GL16438) protein, with a predicted molecular mass of 60 kDa was found in itsgenome. In silico analysis revealed the presence of a nuclear localization signal (NSL) RPKR at 333 position, and disclosed that the putative gSP might be a cysteine protease included in the C48 cysteine protease family, similar to other SUMO proteases described [5,31]. Although gSP presents low identity with other SUMO proteases characterized, it possesses two preserved C-terminal domains and contains a putative catalytic triad (histidine, aspartate, and cysteine), with a conserved glutamine residue essential for the formation of the oxyanion hole in the active site [31,32] (Figure 5A), suggesting that the essential catalytic residues are conserved. Over-expression of the C-terminus HA-tagged gSP (gSP-HA) was used to characterize gSP in growing trophozoites by IFA and Western blot (Figure 5B). Like gSUMO, gSP-HA localized in the cytoplasm, concentrated close to the plasma membrane including the flagella (Figure 5C) and showed nuclear localization in some trophozoites. Giardia lamblia has one predicted gSP. (A) Schematic representation of the putative gSP containing two catalytic C-terminal domains (green) with the essential catalytic residues (red). (B) Western blotting showing a band of ~ 60 kDa (arrow) corresponding to free gSP in gSP-HA transgenic trophozoites (TT). (C) IFA using anti-HA mAb and confocal microscopy show gSP-HA (red) in the cytoplasm and nuclei (DAPI) of transfected Giardia trophozoites. WT: wild-type. Scale bar: 10 µm.All SUMO proteases possess a large N-terminal domain with minimal or no homology to each other’s domain. It has been suggested that the diversified N-terminal domains of these proteases determine their substrate specificity by controlling their subcellular localization [33,34,35]. Studies in both yeast and mammalian systems suggest that subcellular localization contributes to substrate selection by the SUMO proteases in vivo [36]. Distinct subcellular localization has been found by the mammalian SUMO proteases. Thereby, SENP1 has been localized to the nucleoplasm and nuclear bodies, SENP2 has been found at the nuclear pore, SENP3 localized to the nucleolus, and SENP6 was cytoplasmic [5,8,37]. Nevertheless, very little is known about the substrate specificity of each of the SUMO enzymes, and how a specific subcellular localization is linked to the function of each SUMO protease remains poorly understood. An interesting evidence was provided by Itahana et al. who observed a nucleocytoplasmic shuttling of human SENP2. Like SENP2 it is possible that gSP may have substrates in the nucleus and the cytoplasm showing similar pattern subcellular localization to gSUMO protein, and may the deSUMOylation of substrates be regulated by cellular processes, such as cell cycle progression. By controlling the function of the nucleocytoplasmic shuttling of gSP, cells could be able to selectively deSUMOylate specific substrates in the nucleus and/or in the cytoplasm, thereby achieving a greater flexibility in regulating gSP activity. Undoubtedly, future studies will reveal more about the functional specificity of the SUMO protease in Giardia and the relationship between the enzyme and the SUMO protein localization.By searching the G. lamblia genome using the SAE1/2 sequence homologies, we identified one gene encoding putative gSAE1 (ID GL10661) and one gene that codified to putative gSAE2 (ID GL6288). Although putative gSAE1 is included by in silico analysis in the Ubiquitin E1 enzymes family protein it does not present the typical structure observed in other SAE1 isoforms characterized [38]. gSAE2 is a homologous protein to human SAE2 or yeast Ulp2p subunits and, similar to those, possesses a putative Zn2+ motif formed by Cys residues 152, 155, 416, and 419 [14] involved in direct binding of SUMO for adenylation, and three conserved domains that include the adenylation domain (ThiF family) (3–134), which binds both SUMO and ATP, the putative catalytic Cys domain (151–183) with the catalytic cysteine (putative C167 in Giardia) responsible for E1-SUMO-thioester bond formation [39], and the UbL or ubiquitin-like domain (315–378), due to its structural similarity to Ub and other Ubl modifiers [15] (Figure 6A). In silico analyses predicted a cytoplasmic localization for the putative gSAE2. IFA revealed a cytoplasmic localization but also nuclear and perinuclear (Figure 6B). With bioinformatics tools, we also found one gene encoding by putative gUbc9 (ID GL24068) that, similar to other E2 family members, shares a conserved UbL domain of approximately 14 kDa and contains a conserved cysteine residue (putative C98 in Giardia) required for the thio-ester formation between SUMO and the E2 member (Figure 7A) [40]. In silico analyzing predicted a nuclear subcellular localization of putative gUbc9, similar to other SUMO-conjugating enzymes described [5]. However, IFA assays enabled us to observe that gUbc9 is present in the cytoplasm and surrounding the nuclei but not inside them (Figure 7B).In silico analysis and localization of the putative gSAE2. (A) Jalview image shows amino acid sequence alignment of the SAE2 from G.mlamblia (glam), S. cerevisiae (scer), and H. sapiens (hsap). Bar graphs showing the amino acid conservation at each residue are shown. The putative active-site cysteine residue (C178 in Giardia) is indicated by empty black box while the cysteines corresponding to putative Zn+ motive are denoted by asterisks. (B) IFA using anti-HA mAb (red) and confocal microscopy show the cytoplasmic and nuclear (DAPI) localization of gSAE2-HA (arrow). Scale bar: 10 µm.Putative gUbc9. (A) Alignment of putative gUbc9 with yeast and human Ubc9 orthologous shows high number of conserved amino acids residues and conserved putative cysteine (C98 in Giardia) required for thio-ester formation between SUMO and the Ubc9 enzyme (boxed and highlighted by a black bar). Bar graph shows the amino acid conservation at each residue. (B) IFA using anti-HA mAb (red) and confocal microscopy show that the enzyme possesses a cytoplasmic localization with the main distribution surrounding the nuclei (arrow) in gUbc9-HA transgenic trophozoites. Scale bar: 10 µm.It is known that SUMO is conjugated to target proteins by an analogous but distinct pathway from ubiquitin conjugation [41], and that ubiquitin enzymes (specifically E1-activating and E2-conjugating enzymes) are highly related to the SUMOylating E1 and E2 enzymes [8]. In contrast to ubiquitin system where E3 ligase enzymes are generally a requirement for ubiquitination, in the SUMOylation pathway the requirement of E3 ligases is a source of debate because SUMO conjugation can be reconstituted under select conditions in vitro using only E1, Ubc9, SUMO, and ATP [16]. Although no putative E3-like ligase was found in the Giardia genome previous reports suggest that Ubc9 is sufficient in SUMOylation as long as the consensus sequence is present [18]. Nearly all SUMO-modified proteins identified to date contain a conserved motif that surrounds the modified lysine, and structural and mutational analyses of this motif indicate that it is recognized directly by Ubc9. Thus, the direct interaction between Ubc9 and SUMO substrates seems to preclude an absolute requirement for E3-like factors [42]. Also, Giardia’s genome encodes a simplified form of many cellular processes: fewer and more basic subunits, incorporation of single-domain bacterial and archaea-like enzymes, and a limited metabolic repertoire that makes some proteins functionally redundant with other proteins in the same or another pathway. Therefore, it is possible that Giardia has acquired a basal protein SUMOylation system based on unusual features compared to other organisms [4]. G. lamblia trophozoites of the isolate WB, clone 1267 (WB/1267) were axenically cultivated in screw-cap borosilicate glass tubes in modified TYI-S-33 medium enriched with 10% heat-inactivated fetal bovine serum at pH 7.5 and supplemented with 0.1% bovine bile [43]. Trophozoites of the WB/1267 clone were transfected by electroporation and selected with puromycin [44,45]. The transfection of stable WB/1267 cells was 100%, as determined by IFA and flow cytometry. Cultures were harvested by chilling on ice followed by agitation to dislodge attached cells. Trophozoites were collected by centrifugation at 500 × g for 10 min at 4°C and washed three times with PBS. The mouse myeloma cell line NSO (ECACC85110503) was grown in RPMI 1640 (GIBCO) supplemented with 10% fetal bovine serum.Chinese hamster ovary (CHO) cells were grown to confluence in D-MEM (GIBCO) medium supplemented with 10% fetal bovine serum. The cells were maintained in a humidified incubator at 37 °C with 5% CO2.Purebred female BALB/c mice (aged 10–12 weeks) were purchased from the Facultad de Ciencias Veterinarias, Universidad de La Plata, and housed at the vivarium of the Instituto Mercedes & Martín Ferreyra (INIMEC-CONICET). They were maintained in our animal facilities, which meet the conditions of the Guide to the Care and Use of Experimental Animals, published by the Canadian Council on Animal Care (with the assurance number A5802-01 being assigned by the Office of Laboratory Animal Welfare (NIH)). Our Institutional Experimentation Animal Committee also approved the animal handling and experimental procedures.Searching in the Giardia genome data base (GDB), only one SUMO homologue (gSUMO, ID GL7760) was identified using the human SUMO-1 sequence as a query. In order to test the conservation of the amino acid composition, a multiple sequence alignment was performed using T-Coffee [46] with default settings and SUMO sequences from H. sapiens, P. troglodytes, M. musculus, D. rerio, D. melanogaster, C. elegans, S. cerevisiae, S. pombe, T. brucei, P. falciparum and A. thaliana. Following the alignment, Block Mapping and Gathering with Entropy software [47] was used to select regions of the alignment with a higher percentage of identity. The resulting shorter alignment (79 residues on average, between positions 34 and 111) was then manually curated with GeneDoc [48]. 3D structure of gSUMO was predicted using Phyre2 [49]. Alfa-helix and Beta-sheet regions (between positions 23 and 96) corresponding to gSUMO and Chain A of the Protein Data Bank file 2K8H from T. brucei [22] were aligned with the program Friend 2.0 [50].Genes coding for putative Ubiquitin-like proteins in G. lamblia were obtained from GiardiaDB. To identify the putative components of the SUMOylation pathway, previously characterized yeast and human orthologues were obtained from NCBI (http://www.ncbi.nlm.nih.gov/BLAST/) and used as query on gBLASTp. Multiple sequence alignments were constructed using Muscle and Jalview algorithm. PSORT (http://psort.hgc.jp/) and phobius (http://phobius.sbc.su.se/) were used to in silico analysis of proteins. Genomic DNA was prepared from G. lamblia as described earlier and used as a template to amplify the gsumo gene with the primers gsumoF and gsumoR (forward primer: 5'- CATTGGATCGGATGACGAAGGAGACGTCCCCCCAATT-3'; reverse primer: 5'-CATTGCGG CCGCCTAGTGGCCGCCAATCTGATTTCGCATCA-3') containing a BamH1 and Not1 restriction enzyme site, respectively. To amplify the gsp gene we used the primers gSPF and gSPR (forward primer: 5’- CATTCCATGGCTGCTGAACTGTTGCAGCTCAAA-3’; reverse primer: 5’-CATTGAT ATCGCAGAGCTCGTCCAGATCTTGCGG-3’) containing a NcoI and EcoRV restriction enzyme site. To amplify the gsae2 gene we used the primers: gsae2F 5’-CATTCCATGGATCTGTGCATCGT CGGGTGCGGC-3’ and gsa2R: 5’-CATTGATATCGCAGAGCTCGTCCAGATCTTGCGG-3’ containing a NcoI and EcoRV restriction enzyme site. To amplify the gubc9 gene we used the primers gubc9F and gubc9R (forward primer: 5’-CATTCCATGGAATTGGCTTTTAAAACACGAAATTCA GTTAAAATGG -3’; reverse primer: 5’- CATTGATATCCTTTTTGCTGGCGTAGTCAAGCGT-3’) containing a NcoI and EcoRV restriction enzyme site. PCR reactions were carried out in a thermal cycler (Eppendorff, Germany) with denaturation at 94 °C for 2 min before 30 cycles of 94 °C for 1 min, 57 °C for 30 s, 72 °C for 1 min 30 s and 72 °C for 7 min for final extension. PCR products were purified (QIAquick PCR purification kit, Qiagen) and digested with BamH1 and Not1 enzymes or NcoI and EcoRV enzymes respectively. Standard recombinant DNA techniques were used to construct the plasmid gSUMO-ptubHA and ptubApaHA-gSP, ptubApaHA-gSAE2, and ptubApaHA-gUbc9. Fragments of 309 bp corresponding to gsumo, and 1620 bp, 1614 bp and 538 bp corresponding to the putative gsp gene, putative gsae2 gene and gubc9 gene respectively, were generated by PCR amplification of genomic DNA. PCR products were purified (QIAquick PCR purification kit, Qiagen), then digested with BamHI-NotI, and NcoI-EcoRV ligated in frame with HA tag sequence at the N-terminus of ptubHA expression vector or with HA tag sequence at the C-terminus of ptubApaHA expression vector with T4 DNA ligase (MBI fermentas). Plasmids were introduced into E. coli X-10 Gold by CaCl2 transformation and the insert was confirmed by colony PCR and DNA sequencing (Macrogen, South Korea).Production of the recombinant protein GST-gSUMO has been described in detail previously [51]. Briefly, cDNA encoding Giardia SUMO was amplified and cloned into the gluthatione-S-transferase (GST) fusion expression vector, pGEX-4T 3 (Amersham Pharmacia Biotech, Little Chalfont, United Kingdom) via BamHI and NotI restriction sites. The GST-tagged protein was expressed in Escherichia coli strain BL21-Codon-Plus (Stratagene, Valencia, CA) and purified using Gluthatione-Sepharose 4B beads (Amersham Pharmacia Biotech) yielding amounts sufficient for mouse immunization.The eluted recombinant gSUMO protein fractions were separated on 12% SDS-PAGE and visualized by Coomassie blue staining.The recombinant protein GST-gSUMO was used as antigen for mouse immunization and monoclonal antibody production. Three female BALB/c mice were subcutaneously injected with 100 μg of antigen emulsified with TiterMax Gold Adjuvant (Sigma, St. Louis, MO) (1:1) on days 1 and 15. On day 30, mice were boosted intravenously with 100 μg of the antigen in PBS. The mouse myeloma cell line NSO was used for fusion with spleen cells obtained from immunized mice. Antibody secreting hybridomas were screened by indirect immunofluorescence and dot-blotting, using non-encysting WB trophozoites, and were then grown, screened and finally cloned.For Western blot assays, parasite lysates were incubated with sample buffer with b-mercaptoethanol, boiled for 10 min, and separated in 10% Bis-Tris gels using a Mini Protean II electrophoresis unit (Bio-Rad). We used 50 µg and 500 µg of proteins for wild type or transgenic trophozoites, respectively. Samples were transferred to nitrocellulose membranes, blocked with 5% skimmed milk and 0.1% Tween 20 in TBS, and then incubated with hybridoma supernatants (1:200) or anti-HA monoclonal antibody (Sigma) for an hour. After washing 3 times with 0.1% Tween 20 in TBS, the strips were incubated for 1 h with horseradish peroxidase-conjugated polyclonal goat anti-mouse Igs (Dako) and then visualized with autoradiography. Controls included the omission of the primary antibody and the use of an unrelated antibody.Trophozoites cultured in growth medium were harvested and processed as described. Briefly, cells were washed with PBSm (1% growth medium in PBS, pH 7.4), allowed to attach to multi-well slides in a humidified chamber at 37 °C for an hour, and the wells were fixed for 40 min with fresh 4% formaldehyde. The cells were incubated sequentially with blocking solution (10% goat serum and 0.1% triton X-100 in PBS) at 37 °C for 30 minutes followed by incubation with anti-HA mAb (1/300) or undiluted hybridoma supernatant at 37 °C for an hour. After washing three times with PBS, the cells were incubated for 1 h in the dark with FITC-conjugated goat anti-mouse secondary antibody (Cappel, Laboratories) or Texas red anti-mouse secondary antibody. Finally, preparations were washed and mounted in Vectashield mounting media. Fluorescence staining was visualized by using a conventional (Zeiss Pascal) inverted confocal microscope, using 100× oil immersion objectives (NA 1.32, zoom X). Differential interference contrast images were collected simultaneously with fluorescence images by the use of a transmitted light detector. Images were processed using FV10-ASW 1.4 Viewer and Adobe Photoshop 8.0 (Adobe Systems) software.In this work we present evidence of SUMOylation in G. lamblia, a protozoan parasiteconsidered a basal organism in eukaryotic evolutionary history. By searching the GiardiaDB, we identified a SUMO gene with products that are highly homologous to SUMO-1 isoforms and SMT3C of yeast, a gene that encodes for a putative gSP, and two proteins, putative gSAE2 and putative Ubc9, with high identity and homology to the SUMOylation enzymes, presumed to function in both the SUMOylation and the ubiquitination pathways [5,8,52]. The over-expression of SUMO in Giardia and in the mAbs produced, enabled us to describe the localization of SUMO in wild-type and transgenic trophozoites and the presence of potential SUMO conjugates. However, research about how SUMO affects biological processes is only in its early stages. Knowledge of the proteins targeted by this modification is of the utmost importance in deciphering the impact of SUMOylation on the biology of the organism. Experiments on SUMOylated candidates are currently underway and will help us to disclose how SUMOylation is regulated, how the SUMOylation process functions to integrate signal pathway networks, and the role of this post-translational modification in the G.lamblia life cycle and in the evolution of the eukaryotes.We thank all the members of our laboratory for helpful discussions. This research was supported in part by the Argentine National Agency for the promotion of Science and Technology (FONCyT), the Secretary of Science and Technology of the National University of Córdoba (SECYT) and the National Council for Sciences and Technology (CONICET). Cecilia V. Vranych produced the specific anti-gSUMO mAb, over-expressed gSUMO protein and SUMOylating enzymes, performed confocal immunofluorescence assays and performed in silico analysis of proteins. María C. Merino performed immunofluorescence assays and Western blotting. Nahuel Zamponi performed bioinformatics analysis of gSUMO protein and three dimensional analysis. María C. Touz performed the figures of paper and three dimensional analysis. Andrea S. Rópolo assisted in the production of the anti-gSUMO mAb, and conceived and co-ordinated the project. Cecilia V. Vranych, María C. Touz and Andrea S. Rópolo wrote the paper, and all authors discussed the results and commented on the paper.The authors declare no conflict of interest.

Med-MDPI/biomolecules/biomolecules-02-03-00331.txt

@@ -0,0 +1 @@
+These authors contributed equally to this work.licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).Small ubiquitin-related modifier (SUMO), an ~90 amino acid ubiquitin-like protein, is highly conserved throughout the eukaryotic domain. Like ubiquitin, SUMO is covalently attached to lysine side chains in a large number of target proteins. In contrast to ubiquitin, SUMO does not have a direct role in targeting proteins for proteasomal degradation. However, like ubiquitin, SUMO does modulate protein function in a variety of other ways. This includes effects on protein conformation, subcellular localization, and protein–protein interactions. Significant insight into the in vivo role of SUMOylation has been provided by studies in Drosophila that combine genetic manipulation, proteomic, and biochemical analysis. Such studies have revealed that the SUMO conjugation pathway regulates a wide variety of critical cellular and developmental processes, including chromatin/chromosome function, eggshell patterning, embryonic pattern formation, metamorphosis, larval and pupal development, neurogenesis, development of the innate immune system, and apoptosis. This review discusses our current understanding of the diverse roles for SUMO in Drosophila development.As its name implies, small ubiquitin-related modifier (SUMO) is a Ubiquitin-Like protein (UbL) with sequence and structural similarity to ubiquitin [1]. While the human genome encodes at least four SUMO family proteins [2,3], the Drosophila genome encodes one such protein [4] which is more similar to human SUMO-2/3/4 than it is to human SUMO-1. Although the Drosophila gene encoding SUMO was originally termed smt3 by analogy with its yeast counterpart, we will use the more familiar “SUMO” (italicized for the gene name, but not for the protein) throughout this review. Like ubiquitin, SUMO becomes covalently attached to lysine side chains in a variety of target proteins in a process termed SUMOylation. It is synthesized as an inactive precursor that must undergo maturation before it can proceed through the enzymatic steps required for conjugation to target proteins. Maturation of SUMO is accomplished by the activity of a ubiquitin-like protease (Ulp), which removes a C-terminal extension from the immature protein exposing a Gly-Gly motif at the C-terminus. The Ulp family is similar to the sentrin/SUMO-specific protease (SENP) family in humans. In Drosophila SUMO, this C-terminal extension is just two residues in length [5].Once processed, the mature SUMO is ready to enter the three-step conjugation pathway. In the initial step, which is coupled to the hydrolysis of ATP to AMP and pyrophosphate, the C-terminal carboxyl group of SUMO forms a thioester bond with an active site cysteine of the heterodimeric SUMO activating enzyme SAE1/SAE2. This thioester linkage is then transferred to an active site cysteine within the conjugating enzyme Ubc9 (which is encoded by the Drosophila gene lesswright (lwr)) [6]. Ubc9 is capable of recognizing the substrate and catalyzes the formation of an isopeptide bond between the C-terminus of SUMO and an ε-amino group of a lysine residue within the substrate. The target lysine residues usually fall within a SUMO conjugation consensus motif Ψ KxE (where Ψ is a large hydrophobic residue and x is any amino acid) [7]. Isopeptide bond formation can be facilitated by the activity of SUMO E3 ligases that bind both Ubc9 and the target, increasing the rate of SUMO conjugation to a specific target or, in some cases, promote poly-SUMO chain formation [8]. In Drosophila, it is likely that the protein inhibitor of activated STAT (PIAS) family protein suppressor of variegation 2-10 (Su(var)2–10) functions as a SUMO E3 ligase [9,10].SUMO conjugation is readily reversed by the action of the Ulps, which function as deconjugases in addition to maturases. Thus, for many targets, the conjugate appears to be transient and it is possible that the transient nature of the modification is an important feature in SUMO function [5]. The Ulps are often found in specific subcellular locales and their ability to mediate deconjugation in a cell-compartment specific manner may play an important role in controlling SUMO pathway function. The Drosophila genome encodes at least two such deconjugases, termed Ulp1 and Ulp2. The first of these, Ulp1, has been extensively characterized [5]. It is primarily associated with the nucleoplasmic face of the nuclear pore complex where it may deconjugate proteins as they exit the nucleus thus serving as a molecular switch to control the biochemical properties of a protein as a function of its subcellular location.SUMO may have roles in both meiotic and mitotic chromosome function. The earliest evidence for a role in meiosis was provided by the discovery that mutations in the gene encoding the SUMO conjugating enzyme Ubc9 can suppress the no distributivedisjunction (nod) meiotic nondisjunction phenotype. It appears that Ubc9 promotes the dissociation of homologous heterochromatic regions at the end of meiotic prophase I [11]. Evidence for a role of the SUMO pathway in coordinating the mitotic chromosome cycle comes from the discovery of chromosomal hypercondensation and aberrant segregation defects observed in SUMO loss-of-function embryos. SUMO is required for progression through the cell cycle and, consistent with this observation, numerous cell cycle regulators have been identified as SUMO conjugation targets [12].Further evidence for a role of SUMO in chromatin function is provided by the finding that SUMO contributes to the ability of the gypsy transposon to insulate chromosomal domains. Two protein components of the gypsy chromatin insulator complex, Mod(mdg4)2.2 and CP190, can be modified by SUMO invitro and invivo. SUMOylation of these insulator components may antagonize insulator activity as disruption of the SUMO pathway can increase the enhancer blocking activity of the insulator. SUMOylation does not affect the ability of CP190 and Mod(mdg4) to bind chromatin, but rather appears to regulate the stability of gypsy insulator complexes. For example, Ubc9 overexpression leads to disintegration of such complexes, whereas reduced levels of SUMOylation partially restore insulator body formation lost due to the absence of Mod(mdg4) [13]. SUMOylation of these insulator body components is apparently regulated by dTopors, a dual ubiquitin/SUMO ligase, which could stimulate insulator body function by interfering with SUMOylation of CP190 and Mod(mdg4). However, further study is required as published data regarding the role of SUMO and Topors in gypsy insulator function are inconsistent [13,14,15,16].Links between SUMO and chromatin function are also suggested by the multiple connections between the SUMO pathway and position effect variegation (PEV). PEV arises from the establishment of mitotically stable transcriptionally silent heterochromatic domains [17]. Suppressors of PEV encode proteins that favor heterochromatic silencing, and one such gene, Su(var)2-10, encodes the Drosophila orthologue of the human SUMO E3 ligase, PIAS, [18]. Furthermore, the conjugation of SUMO to lysine 839 in another such gene, Su(var)3-7 is required to target it to heterochromatic regions and promote PEV [19].SUMO may also be required for the function of the Polycomb group (PcG). This set of gene products functions to establish and maintain an epigenetically stable silent transcriptional state that has some similarity to the heterochromatic state [20]. In Drosophila development, the PcG serves to maintain the silent transcriptional state of the homeotic complex (Hox) genes, such as Ultrabithorax (Ubx), throughout the latter stages of embryogenesis and throughout larval and pupal development. The PcG includes two protein complexes termed Polycomb Repressive Complexes 1 and 2 (PRC1 and PRC2) that interact with cis-regulatory elements termed Polycomb Response Elements (PREs) to maintain a transcriptionally silent state.Evidence that SUMO is required for PcG function came from the discovery that Sex comb on midleg (Scm), a substochiometric but nonetheless essential component of PRC1, is efficiently sumoylated in vivo. In S2 cells, Scm SUMOylation interferes with its association with the major Ubx PRE. Thus, depletion of SUMO by RNAi or mutation of the three SUMO-acceptor sites in Scm lead to increased association of Scm with the PRE and increased repression of Ubx, whereas fusion of SUMO to the N terminal region of Scm interfered with the recruitment of Scm to the PRE [21].Evidence that SUMO function in the PcG pathway is biologically relevant comes from the observation that SUMO has a role in determining the identity of the third thoracic segment in the adult fly. Classic studies on the function of Ubx as a selector gene showed that its function in the third thoracic segment from which the halteres arise is required for the identity of this segment [22]. When Ubx is removed from the third thoracic segment, the resulting homeotic transformation converts this segment into a second thoracic segment with the accompanying transformation of the halteres into wings. As mentioned above, cell culture studies show that SUMO functions via Scm to prevent PcG hyperactivity and thus over-repression of Ubx. Consistent with this observation, when SUMO is depleted from developing halteres via clonal RNAi in the haltere disc, the result is a partial transformation of the haltere to a wing presumably due to inappropriate repression of Ubx by the PcG [21]. Thus, SUMO may negatively regulate Scm function by inhibiting its recruitment to the Ubx major PRE (Figure 1).Role of SUMOylation in Sex comb on midleg (Scm) protein function.SUMOylation (S) of Scm blocks its recruitment to the Ultrabithorax (Ubx) Polycomb Response Element (PRE) and thus inhibits Polycomb Repressive Complexes 1 (PRC1)-mediated repression of Ubx. SUMOylation of Scm could also interfere with the binding of PRC1 to Scm, but this idea has not been directly tested. SUMO may also modulate the function of other PcG proteins. For example, Pleiohomeotic, a DNA binding transcription factor with a primary role in PRE recognition, is a SUMO conjugation target [12]. Furthermore, the human PcG protein Pc2 acts as a SUMO E3 ligase, and recruits Ubc9 and the transcriptional corepressor CtBP into subnuclear compartments termed Polycomb bodies, where SUMOylation of CtBP may occur [23].A role for SUMOylation in regulating eggshell patterning was first suggested by the observation that SUMO hypomorphic mutations act as enhancers of the hypomorphic Ras1 ventralized eggshell phenotype [24]. During oogenesis, the Ras signaling cascade is activated in the follicle cells that surround the developing oocyte by the transmembrane receptor tyrosine kinase (RTK) epidermal growth factor receptor (EGFR) to pattern the dorsoventral axis of the eggshell, which is synthesized by the follicle cells [25]. Since Ras signaling is required for specification of the dorsal pattern elements in the eggshell including two processes termed the dorsal appendages, mutations in components of the Ras signaling cascade result in ventralized eggshells as manifested by dorsal appendage defects. SUMO mutations enhanced this phenotype. Furthermore, reduced SUMO levels suppress the eggshell dorsalization that results from expression of an activated form of EGFR suggesting that SUMO acts downstream of EGFR in the follicle cells [24].Further support for the conclusion that SUMO acts downstream of EGFR was provided by experiments examining activation of the Ras signaling pathway components mitogen activated protein kinase (MAPK) and MAPK/Erk kinase (MEK) in EGFR-expressing S2R+ cells. Activation of the Ras pathway with the EGFR ligand Spitz or the insulin-like receptor ligand insulin induces phosphorylation of both MAPK and MEK, and this phosphorylation was significantly attenuated in cells depleted of SUMO by RNAi. In contrast, depletion of SUMO had no effect on MAPK activation by a constitutively active form of Ras1 suggesting that SUMO is required for Ras activation [12].The possibility that SUMO modulates Ras1 function directly is supported by two observations. First, Ras1 can be efficiently sumoylated at multiple lysine residues present within its C-terminal hypervariable region. Second, Ras1 was identified as a SUMO-associated protein in a proteomic screen to identify novel SUMOylation targets in the early embryo. This screen also identified many other proteins involved in Ras signaling as either SUMOylation substrates or SUMO-interacting proteins suggesting that eggshell patterning is regulated at multiple levels by SUMO-modulation of Ras pathway activity [12].Maternally deposited SUMO and SUMO conjugation pathway components (Ubc9, SAE1, SAE2, and Ulp1) are present at high concentrations in the early Drosophila embryo [6,26,27]. The first evidence for a role of SUMOylation in embryonic pattern formation came from an analysis of the embryonic phenotype resulting from loss-of-function mutations in the gene encoding Ubc9. Such mutations lead to a loss of anterior segments, a phenotype that resembles those that result from mutations in hunchback (hb). Reduction of Ubc9 levels in the early embryo produce defects that range from a missing first thoracic segment to the absence of all segments from the first thoracic to the fifth abdominal segment [28]. Consistent with this phenotype, a proteomic screen showed that both Hb and its positive regulator Bicoid are SUMOylation targets in the early embryo and that mutations in the gene encoding SUMO produce a range of defects that include anteroposterior defects [12]. Interestingly, SUMO appears to have opposing, direct and indirect effects on Bicoid function. Experiments looking at embryos with reduced Ubc9 levels suggest that SUMOylation is required for Bcd function [28], while direct modification of Bcd by SUMO appears to negatively regulate Bcd-dependent transcriptional activation [29].Mutations in the gene encoding SUMO also lead to dorsoventral patterning defects [12], an observation that is consistent with the finding that the Rel family morphogen Dorsal is sumoylated in the early embryo [12]. Studies of Dorsal SUMOylation in cultured cells suggest that SUMOylation stimulates Dorsal-dependent transcription [30,31] although the mechanism behind this stimulation is not clear.An additional and better understood role for the SUMO pathway in directing dorsoventral pattern formation is in Decapentaplegic (Dpp) signaling [32]. Dpp is a TGF-β family ligand that is distributed in a concentration gradient in the extraembyonic space on the dorsal side of the early embryo. Dpp signals through transmembrane receptors to trigger the activation of target genes by SMAD family transcription factors, including the Drosophila Mad and Medea (Med) proteins. Specifically, the Dpp signal leads to the phosphorylation of Mad, which then enters the nucleus and activates Dpp target genes as a heterotrimer with Med [33,34].The SUMO pathway may regulate patterning by restricting the range of action of the Dpp morphogen as suggested by experiments in which reduced SUMO pathway activity was found to result in expanded expression of Dpp targets. In accord with this finding, a triple mutant form of Med lacking all three potential SUMO acceptor sites increased Med transcriptional activity, leading to expanded expression of Dpp targets. Overexpression of a Med-SUMO fusion protein led to retraction of the gene expression pattern of Dpp target genes similar to what is observed in dpp heterozygous embryos. This suggests that Med SUMOylation negatively regulates Dpp target genes [32] and is in agreement with findings that demonstrate reduced Smad4 transcriptional activity when this mammalian Med homologue is sumoylated [35].While it is not clear how Med SUMOylation reduces its activity, an effect of SUMOylation on Med mobility may play a role. This is suggested by experiments in which nuclei containing a GFP-Med fusion protein were photobleached followed by determination of the rate of recovery of nuclear fluorescence due to import of unbleached GFP-Med from the cytoplasm into the nucleus. Mutation of the three SUMO acceptor lysines in Med greatly reduced the rate of fluorescence recovery suggesting that SUMOylation of Med may increase its mobility [32].One possible model to explain the link between SUMOylation, Med mobility, and Dpp signaling is the following: First, it is known that Dpp signaling leads to Mad phosphorylation and that the resulting phospho-Mad then enters the nucleus bringing Med with it [33]. Once in the nucleus, Med would encounter the SUMOylation machinery (which is almost exclusively nuclear [5,26,27]) and become sumoylated. The increased mobility of sumoylated Med would then facilitate its export from the nucleus (Figure 2), explaining the broader nuclear distribution of Med that results from disruption of the SUMO pathway. Since the SUMO deconjugating enzyme Ulp1 resides in the nuclear pore complex, it is possible that Med would be de-sumoylated on its way out of the nucleus decreasing its mobility and thus decreasing the rate of Med re-import. This cyclic conjugation and deconjugation of Med during nucleocytoplasmic shuttling could thus render Dpp signaling self-limiting by ensuring that Med only enters the nucleus transiently after a Dpp signal is received, thereby preventing the inappropriate activation of Dpp target genes after signaling stops.Med SUMOylation limits Dpp signaling by stimulating Med nuclear export. Decapentaplegic (Dpp) signaling enhances nuclear import of Med through an interaction with phospho-Mad (pMad). Once in the nucleus, Med is sumoylated. This increases Med mobility and may thus permit Med nuclear export, thereby limiting the time of Med residence in the nucleus. Since less sumoylated Med is detected in the presence of the Dpp signal, it is possible that phosphorylated Mad decreases the rate of Med SUMOylation perhaps by interfering with the Ubc9-Med interaction.SUMO appears to have multiple roles in wing morphogenesis. For example, SUMO modulates the function of the wing selector gene vestigial (vg). This gene, which encodes a transcriptional coactivator, is essential for wing formation and, furthermore, its misexpression can lead to ectopic wing formation. Mutations in SUMO, Ubc9, or Su(var)2-10 all enhance the wing notching phenotype that results from reduced function of vestigial (vg). In addition, SUMO overexpression enhances the ectopic wing outgrowth phenotype that results from misexpression of vg in the eye imaginal disc [36].A further role for SUMO in wing development is demonstrated by a study examining interactions between the SUMO pathway and the Drosophila Spalt-like (Sall) family proteins Spalt (Sal) and Spalt-related (Salr). These zinc finger transcription factors control growth and wing venation in the central part of the wing during larval and pupal development [37]. A vertebrate member of this family is a known SUMOylation target although the functional significance of its SUMOylation is not known [38].Both loss and gain of function analyses were employed to determine if the SUMO pathway is required for the function of the Drosophila Sall proteins in the wing. Flies heterozygous for a deficiency that removes both sal and salr exhibited a modest decrease in wing size, an effect that was significantly increased upon removal of one copy of the gene encoding SUMO or the gene encoding Ubc9, implying a synergistic interaction between Sall proteins and the SUMO pathway. In addition, flies doubly heterozygous for the deficiency and the gene encoding SUMO exhibited ectopic vein material, a phenotype not observed in either single heterozygote [39].Both Sal and Salr were found to contain two evolutionarily conserved potential SUMOylation sites and mutagenesis of critical aspartate or glutamate residues downstream of each SUMO acceptor lysine was found to significantly reduce SUMOylation [39]. Phenotypes due to overexpression of wild-type and mutant forms of Sall proteins were then assessed by examining wing size, wing venation, and the expression of knirps, a known target of Sall family proteins in the wing. Induction of ectopic wing vein formation and broadening of the knirps expression domain by Sal overexpression is partially suppressed by mutation of the SUMO acceptor sites in Sal. In contrast induction of ectopic wing vein formation and broadening of the knirps expression domain is enhanced by mutation of the SUMO acceptor sites in Salr. Thus, SUMO appears to affect Sal and Salr in contrasting ways.To determine how SUMOylation might alter Sall protein functions, the effects of the SUMO on Sall protein subnuclear localization in S2 cells was assessed [39]. Overexpressed Sal and Salr both localized to punctate nuclear bodies. Co-overexpression of SUMO, presumably leading to increased SUMOylation, blocked punctate body formation in the case of Sal, but led to the formation of large nuclear aggregates in the case of Salr. The effects of SUMO co-overexpression were not observed when SUMOylation deficient mutants of Sal and Salr were employed in place of the wild-type Sall proteins. It thus appears that SUMO may differentially regulate the two Drosophila Sall proteins by differential affects on their subnuclear localization. Consistent with this view, SUMOylation differentially regulates Sal and Salr function in cell culture transfection assays [40]. SUMO-mediated regulation of wing development via the modulation of protein subcellular localization may be a common theme in wing morphogenesis. In addition to the effect of SUMO on the Sall protein localization discussed above, SUMO also appears to alter the subcellular localization of the homeodomain interacting protein kinase (Hipk), thereby influencing c-Jun N-terminal kinase (JNK) signaling during wing development. JNK is a member of the MAP kinase family with multiple roles in development. Studies of vertebrate JNK signaling suggest that it might be regulated by Hipk family kinases [41,42], although the mechanistic basis for this regulation is not understood.SUMO depletion in the wing disc by RNAi leads to a marked increase in the number of apoptotic cells, a phenotype that was significantly reduced when JNK activity was reduced either by RNAi against JNK or through the use of a dominant negative JNK allele. In support of the hypothesis that SUMO depletion leads to apoptosis via increased JNK signaling, depletion of SUMO in the wing disc caused ectopic expression of downstream targets of JNK signaling. The upregulation of JNK signaling in SUMO-depleted wing discs required the action of Hipk. Sumoylated Hipk is found in the nucleus and reducing SUMO levels results in translocation of Hipk from the nucleus to the cytoplasm. These findings suggest that SUMOylation of Hipk serves to sequester it in the nucleus, and loss of Hipk SUMOylation allows it to be exported to the cytoplasm where it can then activate the JNK signaling pathway [43] (Figure 3). The discovery that JNK itself is a direct SUMO conjugation target suggests that this signaling pathway may be regulated at multiple levels by SUMOylation [12].Homeodomain interacting protein kinase (Hipk) SUMOylation may down-regulate c-Jun N-terminal kinase (JNK) signaling. Phosphorylation of JNK in the cytoplasm by Hipk may stimulate JNK pathway activity. SUMOylation of Hipk may interfere with JNK signaling by sequestering Hipk in the nucleus. A role for SUMOylation in triggering metamorphosis was first suggested by the observation that Ubc9 mutant larvae exhibit prolonged larval life, but die before pupariation [44]. Consistent with this observation, RNAi-mediated depletion of SUMO from the ecdysone (E)-producing prothoracic gland (PG) resulted in developmental arrest at the third instar larval stage (L3). These larvae failed to pupariate but continued to grow in size (achieving twice their normal weight) and survived for an additional three weeks [45]. E is a precursor to 20-hydroxyecdysone (20E), a hormone that is a crucial regulator of metamorphosis. SUMO depleted larvae had reduced total ecdysteroid levels, suggesting that the L3 arrest phenotype was due to an inability to produce 20E. Consistent with this idea, feeding the SUMO knock-down larvae media containing 20E prevented the L3 developmental arrest and allowed pupariation, although the pupae failed to develop into adults [45].The reduced levels of 20E caused by loss of SUMO in the PG can be explained by changes in the localization and expression levels of steroidogenic factors. The PG-expressed genes phantom (phm), disembodied (dib) and shadow (sad) encode for cytochrome p450 enzymes that are responsible for converting cholesterol to 20E [46]. SUMO depletion in the PG had no effect on Phm. However, Dib protein was present at a lower level in the mitochondria while Sad (normally nuclear and cytoplasmic) accumulated in the cytoplasm and was excluded from the nucleus. Multiple transcription factors that regulate expression of the ecdysteroid biosynthesis enzymes also displayed aberrant localization and expression levels in PG cells with reduced SUMO or Ubc9.SUMO RNAi produced morphological changes in the plasma membrane and nuclei of PG cells but did not result in obvious structural defects in the ER or mitochondria (the primary sites of E biosynthesis from cholesterol). A thickening of the nuclear lamina beneath the inner nuclear membrane was observed, however this did not appear to affect the distribution or function of the nuclear pores as a global effect on nucleo-cytoplasmic transport was not apparent. In wild-type PG cells, the plasma membrane forms many invaginations that extend deep within the cell and the formation of these intracellular channels is believed to facilitate the process of lipid uptake and ecdysteroid secretion. In SUMO depleted PG cells the number and size of these intracellular channels was significantly reduced and the accumulation of sterol-containing lipid droplets in the cytoplasm was inhibited. These findings suggest that the impaired metamorphosis phenotype observed in SUMO knock-down larvae results from a reduction in cholesterol uptake in the PG cells that contributes to an inability to produce the ecdysteroid levels required for the developmental transition from the larval to the pupal stage [45]. A number of SUMO conjugation targets have essential roles in the development of both the central and peripheral nervous system. Tramtrack69 (Ttk69) is a transcriptional repressor that antagonizes neuronal fate determination in the peripheral nervous system [47]. Sumoylated Ttk69 binds to its target DNA sequence with an affinity equal to that of the unmodified protein and Ttk69 colocalizes with SUMO on polytene chromosomes. While the functional effects of SUMOylation on Ttk69-mediated repression remain unknown, the high levels of SUMO present in sensory bristle cells suggest a possible role for SUMOylation during sense organ differentiation [26].The activity of SoxNeuro, a member of the SRY high mobility group (HMG) box (Sox) family with roles in central nervous system (CNS) development also appears to be regulated by SUMOylation. Preventing SUMOylation of SoxNeuro (by mutating the single SUMO acceptor lysine to an arginine or through expression of a dominant-negative form of Ubc9) resulted in increased SoxNeuro-mediated transcriptional activity. While embryonic overexpression of wild-type SoxNeuro had no detectable effect on CNS development, overexpression of the SoxNeuro SUMOylation deficient lysine to arginine mutant resulted in significant CNS defects including missing or reduced longitudinal axon tracts, absent or fused neural commissures, and an aberrant axonal fasciculation pattern. These results suggest that repression of SoxNeuro transcriptional activity by SUMOylation is necessary for the proper development of the embryonic CNS [48]. The zinc finger transcription factor Senseless (Sens), which has a role in peripheral nervous system development, also appears to be directly modulated by SUMOylation [49]. Sens is a member of the GPS (Gri1/Pag-3/Senseless) family of proteins and plays an essential role in maturation of sense organ precursor (SOPs) by synergizing with proneural basic-helix-loop-helix (bHLH) transcription factors as they autoregulate the genes that encode them. Sens is a direct target for SUMOylation and mutagenesis of its SUMO-acceptor lysine reduces synergy between Sens and the proneural transcription factors both in cell culture and in vivo. This suggests that SUMOylation of Sens promotes its synergistic interaction with proneural proteins, to regulate a critical step in the maturation of the SOPs [49]. The innate immune response in Drosophila consists of a humoral component (the expression of anti-microbial peptides in the fat body) and a cellular component (the production of hemocytes capable of recognizing and neutralizing foreign material through phagocytosis and/or encapsulation). Multiple studies have shown that the SUMO conjugation pathway acts as a critical negative effector of Toll/Rel family-mediated regulation of anti-microbial peptide expression and hematopoiesis. Similar to phenotypes observed with Toll gain-of-function alleles, Ubc9 mutants display “melanotic tumors”, an overproliferation of hemocytes, and they exhibit constitutive expression of the anti-microbial peptides Drosomycin and Cecropin in the absence of immune challenge. These defects in hematopoietic proliferation and anti-microbial peptide expression are suppressed by mutations in the Rel family genes dorsal and dif [44]. In Ubc9 mutant hemocytes and fat body cells, Dorsal and Dif accumulate in the nucleus in a manner similar to that observed during activation of the Toll signaling pathway [50,51]. In the absence of Toll signaling, Dorsal is retained in the cytoplasm through an interaction with the IκB family protein Cactus. Cactus protein levels are significantly reduced in Ubc9 mutant fat body cells and a gain-of-function allele of Cactus (resistant to Toll signal-dependent degradation) suppresses “tumorigenesis”, overproduction of hemocytes, and constitutive expression of Drosomycin caused by loss of Ubc9 activity [51]. These findings are consistent with the mammalian model of inhibition of the Rel family factor NFκB by SUMO, in which SUMOylation of IκB protects it from ubiquitin-mediated degradation [52]. While SUMOylation of Cactus has not been demonstrated, Ubc9 does physically associate with both Cactus and Dorsal [31].While the above findings suggest that SUMO downregulates the innate immune response under some circumstances, SUMO may also stimulate Rel family protein activity and therefore the immune response in other contexts. For example, mutations in Ubc9 or SUMO interfere with antimicrobial peptide production in first instar larvae in response to treatment with the bacterial cell wall component lipopolysaccharide [30]. This is consistent with the observation that SUMOylation of the Rel family protein Dorsal makes it a more potent transcriptional activator in S2 cells [30,31].An additional SUMOylation target that has not been discussed above and that may link SUMO’s roles in embryonic patterning, wing morphogenesis, and neurogenesis is Groucho (Gro), a factor that has broad roles in all these processes. This factor is broadly distributed in the embryo and imaginal discs. It is a corepressor, meaning that, while it is required for transcriptional repression, it does not bind to DNA directly, but rather is recruited to DNA by DNA-binding repressor proteins. It is required for the function of many of the repressors that control development [53].Gro contains highly conversed regions at its N and C-termini (the Q and WD-repeat domains, respectively) that are essential for function and a poorly conserved central region (consisting of GP, CcN, and SP domains) [54]. While these central domains appear to be disordered, they nonetheless play critical positive and negative roles in repression–the GP and CcN domains are required for repression, while the SP domain dampens Gro function, and therefore its deletion results in Gro hyperactivity [55]. The disorder in this region could allow it to adapt to and bind multiple targets. The resulting low affinity, but high specificity interactions could be readily modulated by posttranslational modification of this region making these central domains into prime candidates for regulation by such modifications [53,55]. In accord with this idea, this region is known to be targeted for phosphorylation by a variety of protein kinases and is also the target of SUMOylation [53] (Figure 4).The role of SUMOylation in Groucho (Gro)-mediated repression. (A) Gro consists of five domains (the Q, GP, CcN, SP, and WD-repeat domains). Known SUMO acceptor residues are shown in red, while known phospho-acceptor residues are shown in green; (B) Proposed mechanism by which SUMOylation of Gro enhances its ability to repress by increasing its affinity for a SUMO interaction motif (SIM) in HDAC1; (C) Proposed mechanism by which SUMOylation of Gro inhibits Gro-mediated repression through an inhibitory interaction between sumoylated Gro and a SIM in the SUMO-targeted ubiquitin ligase Dgrn. This interaction may lead to relocalization of Gro to a cellular compartment where it is inactive. Dgrn may also direct ubiquitylation of the Gro-dependent repressor Hairy, inhibiting Hairy’s ability to interact with Gro, thus further relieving the repression of a subset of Gro targets. Four potential SUMOylation sites have been identified in Gro, two in the positively acting GP domain and two in the negatively acting SP domain, suggesting that SUMOylation could have both positive and negative roles in controlling Gro function [56]. In support of the possibility that SUMO upregulates Gro, mutagenesis of the SUMO acceptor sites in Gro results in decreased Gro-mediated repression in mammalian cultured cells. Furthermore, Ubc9 knockdown appears to compromise repression by wild-type Gro, but not by the quadruple SUMO acceptor site mutant [56].The ability of SUMO to enhance Gro function may result from the ability of SUMO to stabilize an interaction between Gro and the histone deacetylase HDAC1, which a number of studies have shown to be required for repression by Gro family proteins. HDAC1 contains a conserved SUMO-interaction motif (SIM). These motifs, which consist of a hydrophobic stretch and a nearby charged region, mediate non-covalent interactions with SUMO and are found in a number of SUMO-interacting proteins. In co-immunoprecipitation experiments, a SUMO-Gro fusion protein was found to bind wild-type HDAC1 more efficiently than did the Gro protein not fused to SUMO, while an HDAC1 mutant lacking the SIM did not [56]. These analyses suggest that Gro SUMOylation enhances its corepressor activity due to efficient recruitment of HDAC1 through its SIM domain (Figure 4B). These findings are consistent with the observation that two of the SUMO acceptor sites in Gro are in the GP domain, which has a major role in HDAC1 recruitment. However, it should be noted that it has not been determined if the SUMO-acceptor sites in the GP domain are specifically required for the enhancement of HDAC1 binding.While the above findings suggest that SUMO regulates Gro function positively, another study suggests that SUMO may antagonize Gro function, which could be in accord with the presence of SUMOylation sites in the negatively acting SP domain. In this study, Gro was found to be a target of regulation by the SUMO-targeted ubiquitin ligase (STUbL) Degringolade (Dgrn) [57]. STUbLs are RING domain ubiquitin ligases that contain SIMs through which they recognize sumoylated proteins [58]. Tissue culture experiments suggest that Dgrn and Gro work antagonistically and that this antagonism requires a functional SUMO pathway [57]. For example, repression by Gro in S2 cells of a reporter under the control of a pro-neural promoter is blocked by wild-type Dgrn, but not by an inactive form of Dgrn containing a mutated RING domain [57]. Furthermore, the effect of Dgrn on Gro-mediated repression is reduced when SUMO levels are reduced. Evidence that the antagonistic interaction between Gro and Dgrn is functional in vivo is provided by the ability of Dgrn, but not the Dgrn SIM mutant, to suppress the small eye phenotype that results from overexpression of Gro in the eye disc. Similarly, Dgrn prevents the loss of wing bristles that results from Gro overexpression in the developing wing [57].As a major role of ubiquitin is to target proteins for destruction by the proteasome, we might expect the antagonistic relationship between the ubiquitin ligase Dgrn and Gro to be the result of ubiquitin-dependent proteoloysis of Gro. Surprisingly, however, it does not appear that Gro levels are reduced in the presence of Dgrn, nor is there any evidence that Gro is ubiquitylated by Dgrn. Rather, it appears that, in the presence of Dgrn, Gro may become sequestered to some cellular compartment where it is inactive (Figure 4C), as suggested by the observation that extraction of sumoylated Gro from Dgrn-containing cells requires harsher conditions that does extraction of sumoylated Gro from Dgrn-deficient cells. Programmed cell death, apoptosis, plays widespread roles in development by eliminating excess cells and thus helping to sculpt structures during morphogenesis. As already mentioned above, SUMO may regulate apoptosis in response to JNK signaling by modulating Hipk subcellular localization. In addition, SUMO may regulate apoptosis via effects on the activity of the tumor suppressor protein p53. This DNA binding transcription factor is critical for triggering the expression of apoptotic genes, and the Drosophila p53 orthologue (Dmp53) is known to activate the expression of pro-apoptotic genes such as Reaper, Grim, and Hid. In the Drosophila eye, apoptosis is a normal part of development, serving to remove excess cells during pupal eye development, and can also be induced by events such as DNA damage that interfere with normal cell cycle progression.In mammals, SUMOylation has been implicated as a regulator of p53 subcellular localization, nucleocytoplasmic translocation, and intermolecular interactions including DNA-binding [59,60]. However, the biological relevance of these functions is controversial [61,62]. Like mammalian p53, Dmp53 appears to be a target of SUMOylation [63,64,65,66]. However, unlike mammalian p53, which contains a single SUMO acceptor lysine, Dmp53 contains two such residues located at opposite ends of the protein, neither of which seems to correspond to the SUMO acceptor lysine in the mammalian protein. Reporter assays in Drosophila S2 cells suggest that these two SUMO acceptor residues are required for full Dmp53 transcriptional activity [63]. The fact that either SUMO acceptor site alone is sufficient for activity, suggests that SUMO does not modify p53 activity through a direct effect on one of its biochemical functions (e.g., DNA binding, oligomerization), but rather that SUMO just needs to be present at the site of Dmp53 action, perhaps to recruit coregulatory factors [63,66].In support of the notion that Dmp53 SUMOylation modulates the apoptotic response in development, it was found that mutagenesis of the two SUMO acceptor lysines attenuated the apoptotic response that results from ionizing radiation-induced DNA damage in larval eye discs. [63]. This is consistent with studies conducted in mammalian cultured cells and in mice in which mutations in p53 SUMO acceptor sites were found to cause defects in the DNA damage response [67,68]. In sum, these results suggest that Dmp53 SUMOylation is required for induction of apoptosis following DNA damage. However, genetic evidence that SUMO pathway components contribute to the apoptotic response is lacking. SUMO is a reversible regulatory switch that can be used to control many (perhaps most) developmental processes [69]. This regulation is often complex and often targets multiple components of a pathway, sometimes in contrasting ways. In a sense, attempting to determine the role of SUMOylation in development is like trying to formulate a coherent picture regarding the role of any common protein modification (e.g., phosphorylation, acetylation, glycosylation, etc.) in development. The pleiotropic roles of SUMO together with the redundancy associated with the existence of multiple SUMOylation targets in any given pathway and multiple SUMO acceptor sites in any given target have made this analysis a challenge. To overcome the challenge of interpreting data produced from experiments in which SUMOylation has been globally perturbed, SUMO acceptor site mutations within the substrate have been utilized. However, this approach can be complicated by the possibility that these sites could serve as acceptors for other posttranslational modifications (e.g., acetylation and methylation). An experimental approach that combines SUMO acceptor site mutants, SUMO-substrate fusions, and overexpression/knockdown of pathway components is perhaps the best way to gain a clear picture of the effects of substrate-specific SUMO modification.An additional complicating factor in studies of the biological roles of SUMOylation is the so-called “SUMO enigma” [60]. In particular, it is often found that only a small proportion of the population of any given substrate needs to be sumoylated to see robust effects on substrate targeting or activity. This suggests that cyclic conjugation and deconjugation may leave behind an unmodified protein that nonetheless retains a memory of having been sumoylated. For example, perhaps SUMOylation leads to the formation of macromolecular complexes, which then, because of some kind of “hysteresis”, hold together even after deconjugation. Alternatively, perhaps cyclic conjugation and deconjugation serves to take target proteins through multiple functional states that are needed in succession for pathway activity.Given the above complexities, a facile genetic model, such as Drosophila, has proved invaluable in efforts to decipher the roles of SUMO in development. Using such a model, it has been possible to combine biochemical and cell culture studies with genetic studies targeting the genes encoding the proteins in the SUMOylation pathway, and with genetic studies targeting the SUMO acceptor sites in the substrate proteins. In this way, it has been possible to build up compelling evidence for functional links between numerous SUMOylation substrates and a variety of developmental pathways.

Med-MDPI/biomolecules/biomolecules-02-03-00350.txt

@@ -0,0 +1 @@
+These authors contributed equally to this work.licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).DNA double-strand breaks (DSBs) comprise one of the most toxic DNA lesions, as the failure to repair a single DSB has detrimental consequences on the cell. Homologous recombination (HR) constitutes an error-free repair pathway for the repair of DSBs. On the other hand, when uncontrolled, HR can lead to genome rearrangements and needs to be tightly regulated. In recent years, several proteins involved in different steps of HR have been shown to undergo modification by small ubiquitin-like modifier (SUMO) peptide and it has been suggested that deficient sumoylation impairs the progression of HR. This review addresses specific effects of sumoylation on the properties of various HR proteins and describes its importance for the homeostasis of DNA repetitive sequences. The article further illustrates the role of sumoylation in meiotic recombination and the interplay between SUMO and other post-translational modifications. Maintenance of genetic information is essential for genomic integrity, but it is constantly challenged by enormous amounts of DNA damage. DNA double-strand breaks (DSBs) comprise one of the most serious kinds of DNA lesions. If left unrepaired, these could lead to aneuploidy, genetic aberrations or cell death. Defects in DSB repair are linked to many human syndromes, such as neurodegenerative diseases, immunodeficiency and cancer. Two major pathways have evolved for repair of DSBs: non-homologous end joining (NHEJ) and homologous recombination (HR). The error-prone NHEJ is used throughout the cell cycle and is most prominent in G1 phase of the cell cycle. On the other hand, the error-free HR generally uses sister chromatid for repair and is therefore dominant in the S and G2 phases [1,2,3]. The pathway choice between the DSB repair mechanisms also varies among species. This could reflect different expression of various recombination proteins, presence of factors that suppress or promote individual pathway as well as regulation on the level of post-translational modifications [1,2,3,4,5,6,7]. Increasing evidence indicates that not only phosphorylation, but also sumoylation is possibly a key regulatory component. Here, we review the role of sumoylation during DSB repair while focusing especially on the HR pathway. For clarity’s sake, we predominantly summarize findings from yeast, as this model system provides most of the pioneering studies, but we will also integrate these findings with data from other organisms. The two pathways for DSB repair differ in their mechanisms and enzymatic requirements. The repair of DSBs via NHEJ is characterized by binding of the Ku70/80 heterodimer to the broken ends followed by recruitment of the Mre11-Rad50-Xrs2 (MRX) complex (Figure 1). The main function of the Ku70/80 complex is to protect DNA ends against nucleolytic degradation and to recruit additional NHEJ proteins. Meanwhile, MRX complex bridges the ends and prevents their separation. In some cases, the ends require removal of damaged nucleotides to allow conjugation by NHEJ-specific DNA ligase IV and its associated factor, Lif1 (reviewed in [8]; Figure 1). In the second DSB repair pathway, homologous recombination (Figure 1), the broken ends need to be nucleolytically processed to produce 3’ single-stranded DNA (ssDNA) overhangs. During HR, MRX complex binding to the broken ends catalyzes removal of short oligonucleotides from the 5’ ends in collaboration with Sae2 nuclease. Two alternative pathways then extensively process the short 3’ overhangs. One is characterized by an action of Exo1 (5’–3’ exonuclease), while the other is dependent on the activities of Dna2 endonuclease and the Sgs1-Top3-Rmi1 complex [9]. During resection, the resulting ssDNA strand is rapidly bound by replication protein A (RPA), which not only promotes end resection but also prevents formation of secondary structures. In addition, RPA creates a barrier for the binding of Rad51 recombinase, thus prohibiting the formation of Rad51 presynaptic filament. To overcome this inhibitory effect, the action of recombination mediators, in particular Rad52 and the Rad55–57 complex, is required [10]. These assemble Rad51 on the RPA-coated ssDNA and promote the formation of Rad51 filament. Upon Rad51 filament formation, HR continues by searching for a homologous sequence, followed by DNA-strand invasion, and results in a D-loop formation. These reactions are catalyzed by Rad54 protein [11]. The double-strand break repair pathways in S. cerevisiae. After DNA damage, DSBs can be either resected to generate 3’ ssDNA tails and directly ligated by non-homologous end joining (NHEJ) (I) or processed by homologous recombination (HR) (II). In HR, resection of a DSB is followed by formation of a Rad51 presynaptic filament invading into the homologous strand to form a D-loop structure. The invading strand is then extended by DNA synthesis. The resulting extended D-loop could then be processed by one of three alternative mechanisms: synthesis-dependent strand annealing (SDSA) (A); double-strand break repair (DSBR) (B); or break-induced replication (BIR) (C); Proteins involved in DSB repair that undergo sumoylation are depicted. An alternative pathway–single strand annealing (SSA)–can be used for DSBs occurring between repeated DNA sequences (D).The invading strand of the D-loop is at this point extended by a polymerase, which is followed by resolution in one of the three possible sub-pathways. The first pathway, designated double-strand break repair (DSBR, [12]), proceeds by capturing the second DSB end to the extended D-loop and formation of double Holliday junctions (dHJ) (Figure 1B). dHJs are then resolved into either crossover or non-crossover products characteristic of meiotic recombination (recombination in meiosis will be discussed later). The second pathway, synthesis-dependent strand annealing (SDSA, [13]), is characterized by displacement of the extended strand from the D-loop and its consequent annealing with the complementary strand of the other resected DSB end (Figure 1A). This pathway eliminates formation of crossovers and is therefore typical for the repair of DSBs during mitosis. In the third alternative pathway, the D-loop assembles into a full-fledged replication fork in a process called break-induced replication (BIR) (Figure 1C).This mechanism can lead to a loss of heterozygosity and is also often used to repair broken or shortened telomeres [14]. For further details about HR, see additional review articles [15,16,17,18].DSBs can alternatively be repaired by a mechanism known as single-strand annealing (SSA). SSA is used for DSB repair if directly repeated DNA sequences are present (Figure 1D). After the resection of DSBs, the generated single-stranded DNA overhangs can anneal to the complementary DNA strand with the help of Rad52 and Rad59 proteins. The 3’ non-homologous tails, produced during annealing as an intermediate, are removed by the Rad1-Rad10 endonuclease followed by gap filling and ligation (reviewed in [17]).Sumoylation is a post-translational modification characterized by an attachment of SUMO (small ubiquitin-like modifier) peptide to target proteins. SUMO, known as Smt3 in S. cerevisiae, is an 11 kDa protein that modifies many proteins participating in diverse cellular processes, including DNA repair, replication, basic metabolism, gene transcription, ion and protein transport, and others [19,20,21,22] The conjugation of SUMO to a target protein involves a 3-step mechanism analogous to ubiquitylation (reviewed in [19,20,21,22]). It is initiated by an ATP-dependent activation of the SUMO protein by SUMO-activating enzyme (E1), a heterodimer consisting of Aos1 and Uba2. SUMO is then transferred to SUMO-conjugating enzyme (E2), Ubc9, which catalyzes the conjugation of SUMO to a target protein. To accomplish efficient and specific sumoylation of the substrate, however, the presence of SUMO ligases (E3) is usually necessary [20,23]. To date, four SUMO E3 ligases have been identified in budding yeast: Siz1, Siz2, Mms21, and meiosis specific Zip3 [20,23,24,25,26]. Attachment of SUMO can block interactions occurring at or near the attachment site, or, more often, it provides a binding surface for new protein interactions or stimulates the existing ones. In the latter case, the binding protein contains a SUMO-specific binding site, known as SIM (SUMO-interacting motif). The SIM has been shown to play essential role in multi-step enzymatic processes, and affect the assembly and disassembly of dimeric and multimeric protein complexes (reviewed in [27]). The crucial component of SIM is a hydrophobic amino acid core, which provides an interface for non-covalent interaction with SUMO. However, also position of acidic residues juxtaposed to SIM can further stimulate SIM-SUMO interaction [27,28,29,30,31]. Alternatively, negative charge in the SIM can be introduced by phosphorylation of serine or threonine residues and can additionally lead to increased selectivity of the protein interaction [30,32,33]. At functional level, SUMO has been shown to change protein interaction, localization, stability or activity [19,20,21]. Importantly, sumoylation is a reversible process and SUMO can be rapidly deconjugated from the target protein by the action of SUMO-specific proteases (Ulp1 and Ulp2 in yeast), which makes sumoylation ideal for regulatory purposes [19].As mentioned above, SUMO modification has been implicated as a possible key player in the regulation of DSB repair. Mutations or deletions of the components of SUMO machinery lead to severe defects, including recombination abnormalities, thus implicating sumoylation as a potential regulator of recombinational repair [25,34,35,36,37]. Ubc9 mutant cells as well as cells carrying a SUMO ligase-deficient allele of MMS21 exhibit increased sensitivity to DNA-damaging agents [25,35]. Moreover, during replication of damaged template both mutants accumulate cruciform structures in a Rad51-dependent manner [34]. In addition, a strong hyper-recombination phenotype has been observed in a mutant of SUMO protease Ulp1. This mutant has also been shown to be synthetically lethal with mutations in genes involved in HR (such as srs2, rad51, rad52, rad54, rad55, rad50, and mre11) [37]. Recently, a breakthrough study from the Zhao laboratory has revealed a comprehensive role of sumoylation in maintaining genome stability [38]. Using a biochemical screen in yeast, they identified a large group of proteins participating in DNA repair and undergoing sumoylation, mainly in response to DNA damage. This revelation greatly broadens the potential roles of sumoylation in genome maintenance. Additional discussion on this work will be described in another review article by Zhao et al. in the next special issue focusing on “DNA damage response”.In HR, the spectrum of SUMO-modified proteins includes all steps indicating SUMO’s substantial role in HR regulation (see Table 1). However, the exact role of sumoylation in HR regulation is often elusive, due to the problems in identification of modified sites and their possible redundancy. Identification of the downstream SUMO-interacting partners, analysis of sumo-deficient alleles as well as permanent sumo-fusion of target proteins that can only partially mimic the effect of sumoylation further hinder the task. Sumoylated proteins involved in DSB repair.1 also involved in NHEJ and SSA; 2 also involved in SSA.Nevertheless, the available data of selected examples described below show the diversity of the effects of SUMO modification, including its ability to regulate the intracellular localization, stability, and conformation of target protein as well as their interactions or biochemical activities. Future studies will be needed to uncover the molecular mechanism and biological function of sumoylation in HR.The initiation of the end resection turns the DSB repair into HR pathway. The MRX and Sae2 represent the key components of end processing machinery. The role of sumoylation in this step is clearly indicated by DNA damage induced SUMO targeting of these proteins [38]. Furthermore, defective sumoylation results in an impaired DNA end resection, suggesting that recombination is facilitated by sumoylation [38]. This further supported by the fact that the deletion of Mre11 (leading to disruption of the MRX complex) causes decreased sumoylation of several downstream proteins participating in presynaptic filament formation, such as Rad52, Rad59, Rfa1 and Rfa2 [38]. Perhaps the extent of ssDNA generated through the end processing is being monitored and correlated by DNA damage-induced sumoylation machinery. This is an interesting reminiscence of DNA-damage checkpoint signalling in human that is sensed by ATR-ATRIP complex via extent of ssDNA bound by RPA [39,40]. Further discussion on the relationship between the two will be covered in the above-mentioned review in next issue.Sumoylation strikes at the heart of HR by modifying the crucial recombination mediator Rad52. SUMO-modified Rad52 has been found in S. cerevisiae, S. pombe and human cells indicating a conservation of this process [41,42]. However, SUMO conjugation sites (K10, 11, and 220) identified in S. cerevisiae Rad52 are located outside the highly conserved domain and sumoylation patterns are obviously different in yeast and human Rad52 leading to a hypothesis that sumoylation may have various regulatory roles in yeast and mammalian cells [42,43]. Studies from budding yeast have shown that sumoylated Rad52 is DNA damage-induced and occurs in mitotic as well as meiotic cells [42]. Nonsumoylatable Rad52 exhibits no significant hypersensitivity to MMS and is also not defective in spore viability and sporulation, indicating that SUMO modification maintains Rad52 function [42]. Correspondingly, impaired sumoylation of Rad52 does not significantly affect major mitotic and meiotic recombination frequencies but rather influences the choice and efficiency of the recombination pathway with slight shift towards SSA in sumoylation defective rad52 mutant [44]. This might reflect the defects in biochemical properties of sumoylated Rad52 such as decreased DNA binding, annealing activity and corresponding shorter duration of rad52 sumo-deficient foci [44]. In addition, ssDNA stimulates Rad52 sumoylation and this is not blocked when coated by RPA [44]. This is in good correlation with reduced Rad52 sumoylation in mutants of MRX complex that fail to generate ssDNA due to block of DSB end processing [38,42]. On the other hand ssDNA coated by Rad51 protein is not anymore capable of stimulating Rad52 sumoylation indicating that Rad52 sumoylation proceeds prior to Rad51 filament formation [44]. This hypothesis is further supported by the fact that deletion of RAD51 leads to accumulation of sumoylated Rad52, while deletion of factors participating in subsequent steps of HR (such as Sgs1, Srs2, Rad55, Rad54, Rad59) or replacing Rad51 with an ATPase defective Rad51-K191R mutant suppresses this effect [45]. This evokes an attractive possibility that Rad51-dependent reactions require sumoylated Rad52. Further, accumulation of Rad51-intermediates results in desumoylation of Rad52 leading to its increased proteasomal degradation, a phenotype observed for sumoylation-defective Rad52. This behaviour was even more pronounced in double mutants of Srs2, Sgs1, or Rrm3 helicases, which are known to accumulate recombination intermediates and loss of Rad52 function rescues the cell growth [42]. Moreover, sumoylated Rad52 has been found as an in vitro substrate for Slx5–Slx8 complex, which is a member of SUMO-Targeted Ubiquitin Ligase (STUbL) family of proteins [46]. Slx5 and Slx8 are both RING finger proteins containing multiple SUMO-interacting motifs for binding to conjugated SUMO on a target protein. Such interaction can serve as a signal for Slx5-Slx8-mediated ubiquitylation that could potentially lead to ubiquitin-dependent degradation [47]. However, neither slx5 nor slx8 cells display slower degradation of SUMO-fused Rad52 indicating a possible different function of Slx5–Slx8-mediated ubiquitylation for sumoylated Rad52 [46]. As mentioned above, the single-stranded DNA-binding protein RPA is also a target for sumoylation and SUMO has also been observed to modify its mammalian homolog [48]. While the role of RPA sumoylation in yeast has not yet been addressed, studies of human RPA support the pro-recombination role of sumoylation as SUMO-modified RPA70 initiates Rad51-dependent HR. After treatment with the replication stress inducer camptothecin, RPA70 dissociates from SUMO-specific protease SENP6 and is modified by SUMO-2/3 thus increasing its association with RAD51. This enhancement could be due to interaction of RAD51 with SUMO. Sumoylated RPA70 then facilitates formation of RAD51 foci and promotes HR [48]. Interestingly, the observation that RAD51 interacts with SUMO and also with UBC9 provokes an idea about potential regulation of RAD51 filament formation by sumoylation in humans [49,50]. Nevertheless, sumoylation of Rad51 has not yet been observed either in yeast or in mammals.The effect of recombination mediators can be counteracted by the helicase Srs2, which potently dismantles Rad51 filaments to prevent inappropriate recombination [51,52] and Srs2 is also sumoylated in response to DNA damage [53,54]. Though the function of Srs2 modification is unclear, unscheduled sumoylation has been found to impair SDSA in non‑phosphorylatable Srs2 mutant [54]. Moreover, sumoylation of Srs2 modifies its affinity towards SUMO-PCNA and might be involved in regulation of the diverse roles of Srs2 (the interplay between Srs2 and SUMO-PCNA will be discussed later) [53]. During synapsis, Rad51 presynaptic filament is stabilized by Rad54 protein (a member of the Snf2/Swi2 family of DNA-dependent ATPases), which further stimulates DNA-strand invasion and D-loop formation [55,56]. Even though sumoylation of Rad54 has not been observed, another member of the Snf2/Swi2 family, Uls1, has been proposed to be a SUMO-targeted ubiquitin ligase (STUbL) [57]. Uls1 is reported to bind both SUMO and the ubiquitin-conjugating enzyme Ubc4 and is required to ubiquitylate SUMO conjugates [57]. Nevertheless, the biochemical evidence about its SUMO-dependent ubiquitylation activity is still missing. Importantly, together with other translocases-Rad54 and Rdh54, Uls1 is important for the removal of Rad51 recombinase from chromatin [58]. Recently, Uls1 has been also implicated in replication stress response, and especially in cells lacking Rad52 mediator proteins or Mus81/Mms4 nuclease [59]. However, additional studies will be required to clarify the function of Uls1 as a STUbL ligase potentially targeting HR proteins. SUMO also modifies proteins participating in the post-synaptic phase of HR, including members of the RecQ helicase family, which are involved in resolution of recombination intermediates. Sumoylation of Sgs1 helicase is stimulated by DSB formation induced by ionizing radiation or chemicals [60]. The observation that mutations in Ubc9, Mms21 and Sgs1 results in similar phenotypic outcome suggests that Sgs1 sumoylation is important for resolution of the X-shaped structures formed during DNA replication [34]. It is noteworthy that sumoylation of Sgs1 at K621 was found to be dispensable for homologous recombination but functionally important for telomere–telomere recombination [60]. Studies from other organisms have shown that SUMO modification is a conserved mechanism among RecQ helicases. Sumoylation of Rqh1, a Sgs1 orthologue in Schizosaccharomyces pombe, controls its activity at telomeres [61]. Further, SUMO also regulates the pro- and anti-recombinogenic roles of the human Sgs1 ortholog BLM (deficient in Bloom syndrome). The cells expressing SUMO-deficient BLM mutant are defective in HR and display a defect in Rad51 localization to stalled replication forks [62].The role of SUMO in DSB repair is not restricted to HR, as proteins involved in NHEJ undergo sumoylation as well. Interestingly, a SIM-containing peptide has been found to inhibit NHEJ in humans cell lines, though, the underlying mechanism remains unknown [63]. Ku70 protein, which forms a heterodimer with Ku80, has been shown to be sumoylated in both yeast and humans and can also interact with the SIM-containing peptide after radiation [25,63]. Since the main role of Ku70 is to recognize and protect DSBs as well as load other NHEJ factors, it is possible that SUMO-SIM mediated interaction can either affect the dynamics of the recruitment of additional NHEJ proteins or regulate the removal of the Ku70/80 heterodimer from DNA ends. Also sumoylation of Lif1 and MRX complex was recently observed, but the biological function is not known [38]. Interestingly, the sumoylation of NHEJ factors (Ku70, Ku80 and Lif1) was not influenced by deletion of Mre11 in contrast to proteins involved in recombinational repair, suggesting that resection contributes to the HR proteins sumoylation induction [38]. Therefore it might be intriguing to speculate if sumoylation might play important role in pathway choice between NHEJ–HR, or consequent amplification of the decision signal. Though HR plays a major role in the repair of the DNA containing repetitive sequences, it has to be tightly regulated, as the presence of multiple homologous sequences can lead to unequal sister chromatid exchange and subsequent loss of genetic information. SUMOylation may represent one of the control mechanisms at the repetitive sequences and we will illustrate this on examples of the ribosomal (rDNA) and telomeric DNA.The rDNA is localized in the nucleolus and in S. cerevisiae is formed by 150 tandem repeats encoding the 35S and 5S ribosomal RNAs. Sumoylation plays an essential role in regulation of rDNA recombination as the loss of sumoylation severely impairs rDNA stability [71,72]. The important role of SUMO is further suggested by the striking localization of sumoylated proteins in the nucleolus, when SUMO deconjugation is blocked [73]. One main target of SUMO regulation of HR in nucleolus is the recombination mediator Rad52. In wild type cells, sumoylated Rad52 is excluded from the nucleolus and the DSB is repaired in the nucleoplasm. In contrast, SUMO-deficient mutant of Rad52 forms foci within the nucleolus resulting in hyperrecombination at the rDNA locus and rDNA marker loss [72]. Rad52 sumoylation therefore seems to specifically inhibit formation of Rad52 foci in the nucleolus in order to preserve rDNA integrity. Similar phenotype have been also observed in the Smc5–Smc6 mutant cells [72]. Smc5-Smc6 (structural maintenance of chromosomes) heterodimer forms a core of a multimeric complex containing also six non-Smc elements (Nse1–Nse6) [74]. The complex regulates sister chromatid cohesion, HR, chromatin structure and its dynamics, however, many aspects of its function still remain unclear [74,75]. Smc5–Smc6 is highly enriched at the rDNA and other repetitive sequences and is required for their proper segregation. The complex associates with E3 SUMO ligase Mms21 (Nse2), which activity is required for the integrity of repetitive sequences [25,76]. However, Mms21 in the nucleolus probably targets other factors than Rad52, as Rad52 sumoylation is independent of the Smc5–Smc6 complex and they rather act synergistically in limiting recombination at rDNA [72]. Another protein complex that is enriched in the nucleolus and is involved in the rDNA recombination regulation is the Slx5–Slx8 SUMO-targeted ubiquitin ligase. The slx8Δ cells exhibit increased rDNA recombination and nucleolar Rad52 foci formation [67]. The slx8Δ cells also accumulate Smt3 foci in the nucleolus, suggesting the role of Slx5–Slx8 in proteasomal degradation of sumoylated nucleolar proteins [71]. Altogether these findings demonstrate SUMO’s central role in controlling HR at the rDNA locus. However, further studies are required to determine specificity of the DNA-damage response in nucleolus, other nucleolus-specific SUMO targets, dynamics and molecular mechanism of protein re-localization.Telomeres-the structures located at the ends of chromosomes, are vital for the stability and complete replication of the genome. In S. cerevisiae they consist of 300 base pairs of telomeric repeats terminated by short 3’ single-stranded overhangs. Their resemblance to DSBs necessitates specific set of proteins that protects telomeres from recognition as DSBs and against exonucleolytic cleavage [77]. SUMO is an important regulator of the telomeres and many telomeric proteins undergo sumoylation [78]. Sumoylation limits telomere length both in budding and fission yeast [25,78,79,80,81]. A main target of sumoylation is the Cdc13 protein, an important telomerase regulator. Siz-dependent sumoylation of Cdc13 strengthens its interaction with the telomerase inhibitor Stn1 and thus suppresses telomerase function [78]. Siz2 was also shown to sumoylate Ku70/80 and Sir4 proteins and thus stimulate anchoring of telomeres to the nuclear envelope [82]. The observations that Siz2Δ mutant displays telomerase–dependent telomere extension and elongating telomeres shift away from the nuclear envelope, led to the hypothesis that sumoylation represses telomerase by tethering telomeres to the nuclear periphery, whereas their release is connected to telomere elongation [82]. The importance of sumoylation in directing telomeres to nuclear envelope is also supported by impaired telomere clustering in sumoylation deficient Mms21 cells [25]. Though the mechanism of SUMO-dependent telomere anchoring to the nuclear envelope remains elusive, it possibly involves multiple SUMO–SIM interactions, as both structures are associated with profound sumoylation. The Smc5–Smc6 complex is enriched at the telomeres in the budding and fission yeast [83,84,85]. The whole complex as well as Mms21 activity are particularly important for telomere maintenance in telomerase deficient cells, as their absence cause accumulation of HR intermediates at telomeres, aberrant recombination between sister telomeres and growth termination [86,87]. Contradictory to SUMO’s role in telomere length restriction, SUMO was also found to play an important role in telomere length increase in the telomerase-deficient cells. Though in most cells lacking telomerase the telomere length gradually decreases and finally leads to cell cycle arrest [88], some cells are able to maintain the telomere length by employing recombination-mediated pathways, which are in humans referred to as alternative lengthening of telomeres (ALT) [89]. Sumoylation of Sgs1 was shown to specifically promote telomeric recombination in telomerase-deficient cells [60]. Similarly, sumoylation of Rqh1, the RecQ homologue in S. pombe, stimulates ALT-like recombination events in the taz1Δ and taz1Δ trt1Δ cells [61]. Though the mechanism by which sumoylation of Sgs1 and Rqh1 increases the telomeric recombination is unclear, it was suggested that sumoylation mediates localization of the RecQ helicases to telomeres where they facilitate restart of collapsed replication forks that subsequently lead to telomere-telomere recombination [60,61]. Whether sumoylation of human RecQ homologues has a similar importance in the telomeric recombination remains to be determined. However, conserved sumoylation and the role in telomere maintenance among RecQ helicases indicate their roles may be maintained [68,69,90].The situation observed in the telomerase-deficient yeasts resembles the one occurring in human ALT cancer cell, and importantly in both sumoylation plays a central role in the promoting of telomeric recombination. Similarly to the absence of telomerase in the tlc1Δ yeasts, telomerase is downregulated in most human cells. However, the cancer cells are able to elongate their telomeres and achieve unlimited replicative potential either by upregulating telomerase transcription (85% of cancers), or by the ALT mechanism using recombination between telomeres (15% of cancers) [89,91,92]. The telomeres of the ALT cells usually associate with PML (promyelocytic leukemia) nuclear bodies, which are in this context referred to as ALT-associated PML bodies (APBs) [93]. SUMO plays a central role in the PML bodies as sumoylation and subsequent noncovalent SUMO–SIM interaction between PML subunits is absolutely necessary for PML body formation [94,95]. The association of telomeres and APBs is thought to facilitate recombination between telomeric repetitive sequences, as APBs are known to contain various HR proteins and artificially created APBs cause telomere elongation by a DNA repair mechanism [96,97]. The Smc5–Smc6 complex is also localized in APBs and is required for the telomere–PML colocalization [98]. The necessity of Mms21-dependent sumoylation of telomere-binding proteins for APB formation [98] suggests that the telomere–PML interaction may be stimulated by multiple noncovalent SUMO–SIM interactions, as is the case other PML-interacting partners [68,99,100]. The observation that Smc5–Smc6 depletion inhibits HR at telomeres and their elongation further supports its major role in ALT cells [98]. Moreover, the Smc5–Smc6 complex and Mms21 activity is also important for de-novo formation of PML bodies on telomeric DNA [96,101]. The above-mentioned roles of sumoylation in the telomere length restriction and elongation nicely illustrate how sumoylation of the same substrates occurring in different cellular conditions can lead to completely different outcomes. Recombination during meiosis is a key event that mediates the pairing of homologous maternal and paternal DNA chromosomes, thus ensuring proper exchange of genetic information. Meiotic recombination is characterized by the DSBR pathway that results in generation of crossovers providing a connection between homologues (chiasmata) and facilitating their accurate segregation. In contrast to mitotic recombination, DSBs in meiosis are programmed and endogenously generated by Spo11. After resection, the 3’ ssDNA tails assemble together with Rad51 or meiosis-specific recombinase Dmc1 into nucleoprotein filaments that catalyze a strand-exchange reaction between homologous sequences. While no direct evidence indicates sumoylation of scDmc1, the homologue of Dmc1 in basidiomycete Coprinus cinereus (CcLim15) has been shown to interact with Ubc9 and to be sumoylated both in vitro and in vivo [102]. Another link between SUMO and meiotic recombination is represented by Ecm11 protein. Ecm11 is required for normal DNA synthesis and meiotic recombination, interacts with SUMO and Siz2 ligase, and can also be sumoylated in vivo [103,104]. Correspondingly, nonsumoylatable ecm11 mutant exhibits severe sporulation defects corresponding to the phenotype of the ecm11Δ mutant [104]. However the effect on molecular or biochemical activities remains to be determined. The SUMO modification of HR factors during meiosis is not well studied and awaits further characterization. Nevertheless, evidence already exists to suggest its important role for protein stability and function during meiotic DSB repair. Meiotic recombination proceeds in coordination with the assembly of proteinaceous structures between homologous chromosomes known as synaptonemal complexes (SCs). These structures are formed by two lateral elements (or axes) and a central region which “zips” the axes together. The formation of SCs, which is essential for crossing over, is dependent on initiation of recombination, thus suggesting close connection between recombination and SCs. How the assembly and disassembly of SC is regulated remains unclear (reviewed in [105,106]). Recent observations indicate SUMO as a central player in formation of the SC complex (Figure 2). For example, mutation in Ubc9 leads to delay in synapsis and a major component of the SC central region, the Zip1 protein, has been found to co-localize with SUMO along the synapsed chromosomes [26,107]. Zip1 can also interact with SUMO chains or SUMO-conjugated proteins through the SIM located at its C-terminus [26], suggesting that these interactions may mediate SC formation. This hypothesis is further supported by interaction between Zip1 and the axial element protein Red1. Red1 has been found to interact not only with Zip1 but also with SUMO chains, Ubc9, and SUMO protease Ulp2. These interactions are mediated via the C-terminus of Red1, harboring two SIM motifs [108,109]. The interaction between SUMO chains and Red1 is important for initiating the SC assembly, as it facilitates the Zip1 and Zip3 recruitment [109]. Moreover, the interaction is also essential to promote Tel1- and Mec1-dependent Hop1 phosphorylation, an important step in the cascade promoting interhomologue recombination and ensuring normal meiotic progression [109]. Indeed Red1 can also be covalently modified by SUMO during meiosis [26,108] and its sumoylation seems to be critical for efficient Red1-Zip1 interaction, as interaction Zip1 with SUMO-defective red1-KR mutant is significantly decreased. Furthermore red1-KR exhibits a substantial delay in SC formation resulting in reduced spore viability. This suggests that SUMO-promoted Red1-Zip1 interaction is important for timely SC formation [108]. Eichinger et al. have further shown that the level of Red1 sumoylation is impaired in Δzip3 mutant, thus indicating that Zip3 might be directly linked to Red1’s sumoylation process [108]. It has been shown that Zip3 can function as a SUMO [26] or ubiquitin ligase [110]. As zip3 mutant accumulate high molecular weight SUMO conjugates similar to slx5 and slx8 mutant, it raises an intriguing possibility that Zip3 may also serve as a STUbL [26,46,47,57]. This leads to an interesting hypothesis that coordinated desumoylation can drive disassembly of SC to ensure proper segregation of chromosomes.Several studies indicate that the role of SUMO in meiosis is conserved among eukaryotes. Similarly to budding yeast Smt3, the SUMO homologue in S. pombe (Pmt3) has been found to co-localize along linear elements (LinEs), structures resembling the axial elements of SC [111]. Mutation of the SUMO ligase Pli1 was shown to cause reduced genetic recombination and abnormal LinE formation, thus implicating sumoylation in the regulation of meiotic recombination in S. pombe [111].The formation of synaptonemal complex (SC). During SC assembly Zip3 recruits Ubc9 and SUMO to the synapsis sites thus facilitating formation of SUMO chains and conjugation of SUMO to other proteins (such as Red1). Zip1 dimers polymerize along the lateral elements (LE) where they can bind to Red1 and SUMO chains leading to generation of central region (CR). SC disassembly could proceed by dissociation of SUMO conjugates by the action of SUMO proteases or other counteracting mechanism. For clarity, Red1 and Hop1 along LEs are illustrated, even though the exact distribution of Hop1, Red1 and sumoylated Red1 is unknown.Further, the putative human functional Red1 homologue SYCP3 has been shown to be modified by SUMO2 [108]. A recent study from mammalian cells also identified SCP1 and SCP2 proteins (SC components) to be conjugated to SUMO1 in human testis [112]. Interestingly, SUMO1 and SUMO-2/3 were found to localize to meiotic chromosomes, but with the distinct patterns of their localization, indicating their separate functions in the cell [112]. Taken together, the aforementioned data implies sumoylation to be a potential key player not only in mitotic and meiotic recombination but also in successful progression and completion of meiosis. SUMO is just one component in the intricate network of post-translational protein modifications (PTMs) (Figure 3). Through the years, many examples of various types of interplay between SUMO and other PTMs have been suggested, and these can occur at different levels. Individual pathways can either interact by modifying the same substrate or the modification can target an enzyme belonging to another PTM pathway and thus regulate that PTM pathway’s activity. Here, however, we will only outline the PTM interplay at substrate level. Where this cannot be currently directly illustrated on proteins belonging to the HR pathway, examples from other pathways will be also used. The interplay between sumoylation and other post-translational modifications. Sumoylation does not exist alone but is often influenced by and itself affects other PTMs. Phosphorylation can influence sumoylation both in positive as well as negative manner and also regulate interaction between SUMO and SIM motif in various proteins. Furthermore, acetylation can compete with sumoylation for the same lysine residue, similarly to ubiquitylation that was also reported to cooperate or lead to subsequent reaction with sumoylation. See text for more details.First, phosphorylation of multiple proteins can stimulate their sumoylation (Figure 3) [113]. In this case, a kinase usually targets a serine residue situated at a specific phosphorylation-dependent sumoylation motif (PDSM) and leads to stimulation of the Ubc9 binding by creating negative charge [114]. In other cases, phosphorylation inhibits protein sumoylation. For example, it may counteract unscheduled sumoylation, as in the case of Srs2 [54,113]. In addition, phosphorylation has been reported also to regulate SUMO–SIM interactions, a feature seen in the SUMO ligase PIAS1 or the Daxx protein [32,33]. While phosphorylation is known to regulate various metabolic processes it remains to be determined whether it plays wider role in the regulation of sumoylation during recombination. Second, PTM by acetylation often targets the same lysine residues as does sumoylation. This creates competition for the substrate, as illustrated on the histone modifications (Figure 3) [113,115]. Finally, ubiquitin, SUMO’s most famous cousin, can also modify the same lysine residues (Figure 3). It has been suggested, for example, that SUMO could protect the substrate from ubiquitylation and proteasomal degradation, as reported for Rad52 protein [42]. It seems, however, that in most cases SUMO and ubiquitin do not compete for the substrate; they rather act together to take advantage of individual regulations depending on the actual needs of the cell [116]. Sumoylation and ubiquitylation can even act sequentially as seen in such STUbL ligases as Slx5–Slx8 complex and possibly Uls1, which selectively ubiquitylate sumoylated proteins and have been implicated in HR [57].The complex regulation of DNA repair by PTMs can be ideally illustrated on the proliferating cell nuclear antigen (PCNA). PCNA is a ring-shaped homotrimeric protein that is loaded on DNA to constitute a sliding clamp of DNA polymerases. It provides a platform for recruiting multiple proteins to the DNA and thus coordinates various processes associated with replication and repair [117]. Subsets of these interactions are regulated through the modification of PCNA by ubiquitin and SUMO. PCNA is modified by ubiquitin in response to DNA-damaging agents, including MMS, 4-Nitroquinoline 1-oxide and UV light [118,119]. Such DNA damage leads to monoubiquitylation of PCNA on K164, which strengthens the interaction between PCNA and specialized ubiquitin-binding motif containing DNA polymerases, and enables synthesis through the damaged site [120]. This pathway, known as translesion synthesis, constitutes an error-prone branch of the post-replication repair (PRR) pathway (reviewed in [121]). Moreover, ubiquitin attached to K164 of PCNA can be further modified by K63 linkages to form a polyubiquitin chain [118,122]. Polyubiquitylated PCNA is a prerequisite for proceeding along the error-free branch of PRR, possibly using the template switch/gap repair mechanism [123,124].PCNA can, however, be ubiquitylated on K164 in an alternative pathway dependent on Asf1, which is involved in the deposition of histones H3 and H4 onto newly synthesized DNA and was implicated in the processing of stalled replication forks [125]. Furthermore, persistent nicks in DNA ligase I deficient cells result in PCNA’s ubiquitylation at lysine K107 and consequent S-phase checkpoint response [126]. These observations suggest that different types of DNA damage may lead to different PCNA ubiquitylation patterns and different cellular outcomes. Prior to the S phase, PCNA is also modified by SUMO on lysine K164 and, to a lesser extent, on K127 [118]. While PCNA is sumoylated even in the absence of exogenous DNA damage, massive DNA damage leads to its heightened sumoylation [118]. Sumoylated PCNA (SUMO-PCNA) then recruits Srs2, which has been shown to inhibit unwanted HR and to channel DNA lesions into the PRR pathway [127,128]. SUMO-PCNA interacts with Srs2 through two interaction sites. One site is represented by SUMO interaction motif (SIM) composed of last 5 C-terminal amino acids (IIVID). The second site includes PCNA-specific interaction motif (PIM). Importantly, both sites are necessary for efficient function in vivo [53,129,130,131]. How exactly Srs2 inhibits HR at the replication fork is not quite clear, but two possible mechanisms have been suggested. First, the inhibitory effect may be mediated by Srs2’s ability to disrupt Rad51 presynaptic filaments [51,52]. Second, Srs2 may block the extension of recombination intermediates by outcompeting SUMO-PCNA from its complex with Polδ during the repair synthesis [132]. Though sumoylation and ubiquitylation target the same lysine residue of PCNA (K164) they do not compete for the substrate and rather act in concert to favor PRR [123,127]. The presumed SUMO–ubiquitin cooperation is further evidenced by the observation that both modifications of K164 are important for the break-induced repair [133]. The K164 modifications seem not to compete with PCNA-interacting proteins, since K164 is located at the back of the PCNA ring while most proteins interact at the front of that ring [134,135]. On the other hand, K127 lies at the site responsible for these interactions and it has been proposed that its sumoylation blocks interactions with most PCNA-interacting proteins [136]. The ubiquitylation and sumoylation of PCNA on K164 are conserved among eukaryotic species, and their effects share certain similarities [118,120,137,138,139,140,141]. Though PCNA sumoylation was originally thought not to be present in human cells, a recent study opposes this supposition and suggests that SUMO-PCNA recruits the PARI protein by a similar mechanism and with an effect similar to that of Srs2 [140]. In addition to sumoylation and ubiquitylation, mammalian PCNA has been shown to undergo phosphorylation and acetylation. Phosphorylation can stabilize PCNA and stimulate cell proliferation [142]. The different acetylation statuses lead to the appearance of three PCNA isoforms that differ in their localization and affinities towards DNA polymerases β and δ or the MTH2 protein [143,144].Homologous recombination is a complex multistep pathway that allows various outcomes, depending on the specific situation and subcellular localization. Moreover, HR is interlinked with various other pathways, which is underlined by the fact that they share some of the same protein factors. In such cases, regulation is paramount, and evidence suggests that sumoylation may be an important process to regulate and coordinate the interplay between HR and other pathways (Figure 4).Role of sumoylation on the relationship of homologous recombination and other DNA metabolic processes in S. cerevisiae. Sumoylation influences not only repair of DSBs but also homeostasis of rDNA and telomeres, DNA replication, and meiosis. Examples of sumoylation’s involvement in this interplay are illustrated.In some studies, sumoylation has been found to repress recombination [34,67,127,128]; in others, sumoylation has been seen to promote it [38]. It is therefore possible that sumoylation influences the utilization of positive and the suppression of negative recombination outcomes. On the molecular level, sumoylation often stimulates protein–protein interactions. Therefore, it may bring together proteins to facilitate a certain DNA repair pathway, and by doing so to block alternative repair pathways. Although the evidence supporting this idea can be found in PML bodies or PCNA-Srs2 complex formation [95,127,128], the identification of SUMO-dependent DNA repair complexes remains a challenge for future years.Alternatively, sumoylation can also serve to disassemble or dissociate proteins after fulfilling corresponding task to allow subsequent steps and completion of HR. Future mechanistic work to understand how SUMO uses these different ways to regulate each step of HR will bring clarity and generate a more comprehensive view of the role of sumoylation. In addition, it will be interesting to understand how these regulations occur in response to DNA damage. Last but not least, understanding the role of sumoylation and its regulatory function could potentially be used in development of novel chemotherapeutic treatments.This study was supported by a Wellcome International Senior Research Fellowship (WT076476); the Ministry of Education Youth and Sport of the Czech Republic (ME 10048, Mendel Centre for Education in Biology, Biomedicine and Bioinformatics–CZ.1.07/2.3.00/09.0186); Czech Science Foundation (GACR 301/09/1917, GACR 203/09/H046, and GACR P207/12/2323); and European Regional Development Fund–Project FNUSA-ICRC (CZ.1.05/1.1.00/02.0123). We would like to thank Victoria Marini, Mario Spirek and Katerina Krejci for reading and commenting upon the manuscript.